Google Cloud introduces a specialized agentic AI solution for dental treatment planning, while researchers demonstrate significant advances in autonomous agent self-evolution and multi-agent safety evaluation. These developments highlight the growing sophistication of AI applications in healthcare and complex systems.
Google Cloud has announced a new specialized agentic AI solution designed to address gaps in dental skills, particularly in orthodontic treatment planning. This system leverages AI to assist dental professionals in creating more accurate treatment plans and aligner designs, integrating with existing workflows to reduce human error and improve patient outcomes. Concurrently, researchers have made significant progress in autonomous agent systems, with MOSS demonstrating source-level code rewriting that enables agents to self-evolve through deterministic, Turing-complete evolution. Additionally, new benchmarks like Boiling the Frog reveal critical safety vulnerabilities in tool-using AI models, emphasizing the need for robust evaluation frameworks in multi-turn agent interactions. These developments represent important steps toward more reliable, specialized, and safe AI applications across healthcare and complex autonomous systems.
AI Model Releases
Filling the gaps in dental skills with specialized agentic AI | Google Cloud Blog
Google Cloud has announced a new specialized agentic AI solution designed to address gaps in dental skills, particularly in orthodontic treatment planning. The system leverages AI to assist dental professionals in creating more accurate treatment plans and aligner designs. It integrates with existing dental workflows and is intended to reduce human error and improve patient outcomes. The solution is part of Google's broader effort to bring AI-powered tools to healthcare and specialized professional domains. This represents a significant step toward AI-assisted decision-making in precision healthcare.
Why it matters: This development signals a growing trend of AI being tailored for niche professional applications, particularly in healthcare. It demonstrates how agentic AI can be deployed to enhance expertise in specialized fields, potentially reshaping how professionals approach complex tasks.
Elon Musk has given up on solar power (on Earth)
Research Papers
MOSS: Self-Evolution through Source-Level Rewriting in Autonomous Agent Systems
arXiv: 2605.22794
MOSS represents a significant leap in autonomous agent evolution by addressing a fundamental limitation of existing self-evolving systems: they typically operate only on text-mutable artifacts like skill files or prompt configurations, leaving the core agent architecture untouched. This approach is problematic because critical elements such as routing, hook ordering, state invariants, and dispatch logic reside in code rather than text, making structural failures inaccessible through text-layer modifications alone.
The system's innovation lies in its ability to perform source-level rewriting on production agentic substrates, anchoring each evolution to automatically curated batches of production failures. This deterministic pipeline ensures that changes are grounded in actual system behavior rather than theoretical assumptions, with code modification delegated to external coding agents while MOSS maintains control over stage ordering and verdicts.
The practical impact is substantial: MOSS achieved a 47% improvement in mean grader score on OpenClaw (from 0.25 to 0.61) in a single cycle without human intervention, demonstrating that source-level adaptation can dramatically enhance agent performance. This approach also provides robust safety mechanisms through ephemeral trial workers, user-consent gated promotions, and health-probe gated rollbacks, making it suitable for production environments where reliability is paramount.
Key insight: MOSS demonstrates that source-level code rewriting enables autonomous agents to self-evolve by modifying core architectural elements, not just text artifacts, achieving deterministic, Turing-complete evolution that outperforms traditional approaches.
Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention
arXiv: 2605.22791
Gated DeltaNet-2 addresses a critical limitation in existing linear attention mechanisms where scalar gates control both erasure and writing operations, creating an inherent coupling that limits performance. By introducing separate channel-wise erase gates (b_t) and write gates (w_t), the model achieves a more nuanced approach to memory management that generalizes both Gated DeltaNet and Kimi Delta Attention while maintaining their strengths.
The theoretical foundation of Gated DeltaNet-2 includes a fast-weight update view, chunkwise WY algorithm with channel-wise decay, and gate-aware backward pass that preserves efficient parallel training. These innovations allow the model to maintain computational efficiency while achieving superior performance across language modeling, commonsense reasoning, and retrieval tasks, particularly excelling in long-context RULER needle-in-a-haystack benchmarks.
At 1.3B parameters trained on 100B FineWeb-Edu tokens, Gated DeltaNet-2 outperforms variants of Mamba-2, Gated DeltaNet, KDA, and Mamba-3 across multiple evaluation metrics. The advantage is most pronounced on long-context tasks, where it improves multi-key retrieval settings and remains strong in both recurrent and hybrid settings, demonstrating that the decoupled approach to memory management provides substantial gains in handling extended sequences.
Key insight: Gated DeltaNet-2 advances linear attention by decoupling erase and write operations through channel-wise gates, enabling more precise memory management and superior performance on long-context tasks compared to previous approaches.
LCGuard: Latent Communication Guard for Safe KV Sharing in Multi-Agent Systems
arXiv: 2605.22786
LCGuard tackles a critical security challenge in multi-agent LLM systems where KV caches encode sensitive information that can propagate across agents without explicit disclosure. The framework treats shared KV caches as latent working memory and learns representation-level transformations before cache artifacts are transmitted, addressing the opacity of KV-based communication channels.
The approach formalizes sensitive information leakage operationally through reconstruction - if an adversarial decoder can recover agent-specific inputs from shared cache artifacts, the communication is considered unsafe. This leads to an adversarial training formulation where LCGuard learns transformations that preserve task-relevant semantics while reducing reconstructable information, creating a robust defense against information leakage.
Empirical evaluations across multiple model families and multi-agent benchmarks demonstrate that LCGuard consistently reduces reconstruction-based leakage and attack success rates while maintaining competitive task performance compared to standard KV-sharing baselines. This represents a significant advancement in enabling efficient, secure communication in multi-agent systems without sacrificing performance.
Key insight: LCGuard introduces a representation-level transformation framework that safely enables KV cache sharing in multi-agent systems by reducing sensitive information leakage through adversarial training.
Deep Reinforcement Learning for Flexible Job Shop Scheduling with Random Job Arrivals
arXiv: 2605.22773
This work addresses the fundamental challenges in Flexible Job Shop Scheduling (FJSP) where unpredictable job arrivals and combinatorial complexity make conventional optimization approaches intractable. The proposed event-based DRL approach employs Proximal Policy Optimization with lightweight MLPs to train agents that minimize total completion time, using state representations directly accessible from the environment.
The key innovation lies in limiting the learning agent to select from well-established dispatching rules rather than exploring an unbounded action space, which provides a practical balance between exploration and exploitation. This approach allows the agent to learn when and how to apply different dispatching strategies based on environmental conditions, rather than attempting to learn complex scheduling policies from scratch.
Simulation results show that the DRL approach outperforms individual dispatching rules across datasets with varying heterogeneity and job arrival rates, particularly excelling in heterogeneous environments where traditional methods struggle. Benchmarking against arrival-triggered mixed-integer linear programming solutions demonstrates that the DRL method achieves good performance especially when dealing with complex, variable datasets, making it a promising approach for real-world scheduling applications.
Key insight: DRL-based scheduling with event-based learning and dispatching rule selection outperforms traditional methods in handling unpredictable job arrivals and combinatorial complexity in flexible job shop scheduling.
Advancing Mathematics Research with AI-Driven Formal Proof Search
arXiv: 2605.22763
This research demonstrates the transformative potential of AI in mathematics research by showing that LLMs can be effectively used to generate formal proofs in languages like Lean, achieving results that were previously unattainable with traditional methods. The most capable agent autonomously resolved 9 of 353 open Erdős problems at a cost of a few hundred dollars per problem, proving 44 of 492 OEIS conjectures, and being deployed in diverse mathematical research areas.
The approach combines LLM-based generation with Lean-based verification, creating a hybrid system that leverages the creativity of LLMs with the rigor of formal verification. This method represents a significant advancement over basic agents that alternate between generation and verification, which proved costlier on the hardest problems, demonstrating that sophisticated agent designs are crucial for achieving breakthrough results.
The findings shed light on the agent designs that enable AI-aided formal proof search, showing that the combination of LLMs with formal verification systems can systematically advance mathematical research. This approach not only demonstrates the power of AI in solving complex mathematical problems but also provides a scalable framework for tackling open problems in various mathematical domains.
Key insight: AI-driven formal proof search using LLMs can autonomously resolve open mathematical problems, with the most capable agents achieving significant breakthroughs in Erdős problems and OEIS conjectures at a fraction of traditional costs.
Towards a General Intelligence and Interface for Wearable Health Data
arXiv: 2605.22759
This work presents a foundation model for wearable health that leverages over one trillion minutes of unlabeled sensor signals from five million participants, demonstrating that joint scaling of model capacity and pretraining data volume leads to systematic improvements in performance across diverse health prediction tasks. The approach addresses key challenges in wearable health data analysis, including high phenotypic diversity, individual baseline variations, and the scarcity of high-quality labeled data.
The model's ability to unlock label-efficient few-shot learning and generative capabilities for robust daily metric estimation represents a significant advancement in health data analysis. The deployment of a classroom of LLM agents to autonomously search downstream predictive heads built on model embeddings shows how the learned representation can be effectively leveraged for various health applications, with performance improvements increasing with LLM model capacity.
The validation through 1,860 ratings from clinicians demonstrates the practical utility of integrating downstream predictors into a Personal Health Agent, supporting more relevant, contextually aware, and safe responses. This approach not only advances the technical capabilities of wearable health monitoring but also provides a framework for developing more personalized and effective health interventions.
Key insight: Foundation models pretrained on massive wearable sensor data enable label-efficient few-shot learning and generative capabilities for robust health prediction, with LLM agents autonomously searching downstream predictive heads for improved performance.
HarnessAPI: A Skill-First Framework for Unified Streaming APIs and MCP Tools
arXiv: 2605.22733
HarnessAPI addresses a fundamental inefficiency in LLM tool deployment where Python functions must exist in two forms: HTTP endpoints for human-facing clients and CI pipelines, and MCP tool registrations for agent runtimes. This duplication leads to divergent code that drifts apart as underlying code evolves, creating maintenance overhead and potential inconsistencies.
The framework's innovation lies in treating a typed skill folder as the single source of truth, automatically deriving multiple representations from a unified codebase. This includes streaming HTTP endpoints with Server-Sent Events, interactive OpenAPI/Swagger UI, and zero-configuration MCP tools, all served from a single process with dual-mode content negotiation that handles both SSE-streaming and JSON-returning clients seamlessly.
Measured across six representative skills, HarnessAPI reduces framework-facing boilerplate by 74% compared to manually maintained dual-stack implementations, demonstrating significant efficiency gains. The framework's inheritance of FastAPI's full middleware, dependency-injection, and deployment ecosystem ensures compatibility with existing toolchains while providing the benefits of unified development and deployment.
Key insight: HarnessAPI eliminates duplication in LLM tool deployment by treating typed skill folders as single source of truth, automatically deriving streaming HTTP endpoints, interactive UIs, and MCP tools from a unified framework.
Can AI Make Conflicts Worse? An Alignment Failure in LLM Deployment Across Conflict Contexts
arXiv: 2605.22720
This research reveals a critical gap in AI safety evaluation: existing practices do not adequately check whether LLM outputs can make conflicts worse in sensitive contexts. The study tested nine model configurations across 90 multi-turn scenarios designed to surface misaligned behavior in conflict contexts, including false equivalence between documented atrocities, denial of genocide, and failure to recognize ethnic slurs.
The findings show that failure rates vary dramatically between models, with the worst performing models failing 47% of the time, making model choice a safety question in its own right. When users pushed for 'balance' in cases where international courts have already assigned responsibility, five of nine configurations failed 80 to 100 percent of the time, demonstrating the potential for AI systems to exacerbate rather than resolve conflicts.
The research introduces the first evaluation framework for this domain and proposes adding it to alignment evaluation portfolios, highlighting the urgent need for specialized safety measures in conflict-sensitive AI deployment. This work underscores that AI systems must be evaluated not just for accuracy but for their potential to cause harm in fragile societies, particularly when their outputs feed into journalism, humanitarian reporting, or public debate.
Key insight: LLM deployment in conflict contexts reveals systematic alignment failures where model outputs can make conflicts worse, with failure rates ranging from 6% to 47% across different configurations, highlighting the urgent need for specialized evaluation frameworks.
AMEL: Accumulated Message Effects on LLM Judgments
arXiv: 2605.22714
The study reveals a significant bias phenomenon in LLM judgments where the polarity of prior conversation history systematically influences subsequent evaluations, termed the accumulated message effect on LLM judgments (AMEL). Across 75,898 API calls to 11 models, identical test items showed shifts toward the conversation's prevailing polarity, with the effect being particularly pronounced for items where models are genuinely uncertain at baseline.
Key findings include that bias does not grow with context length - 5 prior turns and 50 produce the same shift - and there is a strong negativity asymmetry where negative histories induce 1.62x more bias than positive ones. This asymmetry has both token-level and semantic components, though attribution remains exploratory at current sample sizes. The research also shows that position within the conversation does not matter, as five biased turns anywhere in a 50-turn history produce the same shift.
The implications are significant for evaluation pipelines, with the simplest fix being a fresh context per item. When batching is unavoidable, balancing the history helps mitigate the effect. The study demonstrates that scaling helps but does not solve the problem, with even the most advanced models showing persistent bias. This research provides crucial insights for designing more reliable automated evaluation systems and highlights the need for careful consideration of conversation history in LLM applications.
Key insight: LLMs exhibit accumulated message effects where prior conversation history biases subsequent judgments, with negative histories inducing 1.62x more bias than positive ones, and this bias persists regardless of context length or position within the conversation.
Beyond the Org Chart: AI and the Transformation of Invisible Work
arXiv: 2605.22707
This research explores how AI adoption is reshaping work processes in an AI-forward company, revealing that AI's impact extends beyond formal role responsibilities to influence informal cultural practices such as professional mentoring. The study interviewed 24 product-focused individuals at a large technology firm, uncovering how AI is changing both formal and informal aspects of professional development.
The findings show that while AI can facilitate smoother collaboration between peers, it also creates nuanced challenges that put traditional career growth opportunities at risk. The research identifies that AI's influence on mentoring and feedback mechanisms may be altering the typical pathways for professional development, potentially affecting how individuals receive support and guidance in their careers.
The authors propose concrete steps that AI companies can take to make invisible work more visible and suggest efforts that individuals and leaders can implement to support colleagues through AI transformation while preserving healthy company cultures that support diverse thinking, collaboration, and informal interactions. This work provides valuable insights for organizations navigating the cultural implications of AI adoption.
Key insight: AI adoption is transforming informal cultural practices like professional mentoring, creating both opportunities for smoother collaboration and risks to traditional career growth opportunities and mentorship dynamics.
Tokenisation via Convex Relaxations
arXiv: 2605.22821
ConvexTok represents a paradigm shift in tokenization by formulating the problem as a linear program and solving it using convex optimization tools, rather than relying on greedy algorithms like BPE and Unigram that make locally optimal decisions without considering the resulting vocabulary as a whole. This approach yields a new algorithm that consistently improves intrinsic tokenization metrics and bits-per-byte (BpB) achieved by language models.
The method allows users to certify how far their tokeniser is from optimal with respect to a certain objective via a lower bound, finding it to be within 1% of optimal at common vocabulary sizes. This certification capability is particularly valuable for understanding the trade-offs between tokenization quality and computational efficiency, providing users with confidence in their tokenization choices.
Empirical results show that ConvexTok improves downstream task performance, though less consistently than intrinsic metrics. The ability to provide near-optimal solutions while maintaining practical computational scaling makes ConvexTok a promising approach for tokenization that could improve the efficiency and effectiveness of language models across various applications.
Key insight: ConvexTok tokenization via convex optimization outperforms greedy algorithms like BPE and Unigram by consistently improving intrinsic metrics and downstream task performance while providing certification of optimality.
Evaluating Commercial AI Chatbots as News Intermediaries
arXiv: 2605.22785
This evaluation of six AI chatbots across 2,100 factual questions derived from BBC News reporting reveals that while these systems achieve over 90% accuracy on well-formed questions about recent events, they lose significant accuracy under free-response evaluation and when dealing with subtle false premises. The study highlights three key failure patterns that expose fundamental limitations in current AI news intermediation.
The research identifies a clear regional inequity where models achieve 79% accuracy on Hindi questions compared to 89-91% elsewhere, with citations indicating an Anglophone retrieval bias. This demonstrates that AI systems are not equally effective across languages and regions, potentially exacerbating information access disparities. The retrieval failures drive over 70% of all errors, showing that the problem lies in source selection rather than reasoning.
The findings reveal a detection-accuracy paradox where the best false-premise detector ranks second in adversarial accuracy, indicating that premise detection and answer recovery are partially independent capabilities. Overall, these results suggest that high accuracy can mask systematic regional inequity, near-total dependence on retrieval infrastructure, and vulnerability to imperfect queries that real users pose, highlighting the need for more comprehensive evaluation approaches.
Key insight: Commercial AI chatbots show high accuracy on well-formed questions but struggle with regional inequity, retrieval dependence, and false premise detection, revealing systematic biases and limitations in their news handling capabilities.
Reducing Political Manipulation with Consistency Training
arXiv: 2605.22771
The research identifies a phenomenon called covert political bias where LLMs handle counterpart topics from opposing political sides asymmetrically, manifesting in 7 categories of techniques that operate systematically across sensitive contexts. This asymmetric handling creates a significant concern for AI systems deployed in politically sensitive environments.
Political Consistency Training (PCT) addresses this issue through two complementary paradigms: Sentiment Consistency Training, which measures symmetry in rhetoric and framing across paired political prompts, and Helpfulness Consistency Training, which measures symmetric depth and engagement. The approach demonstrates that PCT preserves overall helpfulness while substantially reducing covert political bias and generalizing to held-out benchmarks.
The method's effectiveness is demonstrated through comprehensive evaluation showing that PCT achieves substantial bias reduction while maintaining performance on standard benchmarks. This represents a significant advancement in addressing political bias in LLMs, providing a practical framework for developing more balanced AI systems that can be deployed in politically sensitive contexts without sacrificing their utility.
Key insight: Political Consistency Training (PCT) reduces covert political bias in LLMs through two complementary paradigms - Sentiment Consistency Training and Helpfulness Consistency Training - achieving substantial bias reduction while preserving helpfulness.
Understanding Data Temporality Impact on Large Language Models Pre-training
arXiv: 2605.22769
This work introduces a comprehensive benchmark of over 7,000 temporally grounded questions to study how pre-training dynamics affect the acquisition of time-sensitive factual knowledge, focusing specifically on data ordering. The research demonstrates that sequentially trained models match shuffled baselines on general language understanding and common knowledge while consistently exhibiting more up-to-date and temporally precise knowledge.
The key finding is that temporally ordered pre-training yields improved factual freshness, with shuffled pre-training peaking on older data possibly due to increased factual repetition. This suggests that the temporal ordering of training data has a significant impact on how LLMs acquire and retain factual knowledge, particularly time-sensitive information that changes over time.
The implications are substantial for continual learning in LLMs, as the study provides a foundation for future research on temporal grounding and knowledge evolution. The release of code, checkpoints, and datasets enables further investigation into how temporal ordering affects model behavior and opens new avenues for developing LLMs that can better handle time-sensitive information and adapt to changing factual landscapes.
Key insight: Temporally ordered pre-training improves factual freshness and temporal precision while maintaining general language understanding, demonstrating that data ordering significantly impacts LLM knowledge acquisition and temporal grounding.
ChronoMedKG: A Temporally-Grounded Biomedical Knowledge Graph and Benchmark for Clinical Reasoning
arXiv: 2605.22734
ChronoMedKG addresses a critical limitation in existing biomedical knowledge graphs where disease associations are treated as static facts, ignoring the crucial temporal information needed for clinical reasoning. The system constructs a temporal biomedical knowledge graph containing 460,497 evidence-linked triples covering 13,431 diseases, with each association tied to temporal components like onset window or progression stage.
The approach uses a disease-autonomous multi-agent pipeline where multiple frontier LLMs independently extract knowledge from PubMed and PMC literature, with only relations supported by multi-model consensus, credibility filtering, and ontology alignment being retained. This ensures high-quality, temporally grounded knowledge that can be used for clinical reasoning tasks.
The benchmark results demonstrate that ChronoMedKG provides a crucial temporal axis for retrieval-augmented clinical systems that was previously absent, rescuing 47-65% of long-tail failures in frontier LLMs compared to 17-29% for HPOA-RAG. This represents a significant advancement in clinical reasoning systems, enabling more accurate diagnosis and treatment recommendations by incorporating the temporal dynamics of diseases.
Key insight: ChronoMedKG introduces temporal grounding to biomedical knowledge graphs, enabling more accurate clinical reasoning by incorporating onset windows, progression stages, and temporal components that were previously absent from static KGs.
Tokenization with Split Trees
arXiv: 2605.22705
Tokenization with Split Trees (ToaST) introduces a novel subword tokenization method that directly optimizes compression under a new recursive inference procedure, representing a significant departure from traditional greedy algorithms like BPE and Unigram. The method greedily splits each pretoken into a full binary tree using precomputed byte n-gram counts, independent of any vocabulary.
The approach formulates vocabulary selection as an Integer Program (IP) that minimizes the total token count over all split trees under the inference procedure, with the Linear Programming (LP) relaxation being near-integral in practice. This allows for provably near-optimal vocabularies with training time empirically scaling quadratically in the number of split trees.
On English text, ToaST reduces token counts by more than 11% compared to BPE, WordPiece, and UnigramLM at vocabulary sizes of 40,960 and above, leading to substantial improvements in Renyi efficiency. In experiments training 1.5B parameter language models, ToaST achieves the highest CORE score, outperforming baselines by 2.6%-7.6%, with significance for two of three, and scoring best on 13 of 22 individual tasks, demonstrating its practical effectiveness.
Key insight: ToaST tokenization method directly optimizes compression through recursive inference procedures, achieving 11% reduction in token counts compared to BPE, WordPiece, and UnigramLM at vocabulary sizes of 40,960 and above.
Self-Policy Distillation via Capability-Selective Subspace Projection
arXiv: 2605.22675
Self-Policy Distillation (SPD) addresses a fundamental weakness in existing self-distillation methods where self-generated outputs entangle task-relevant capability with stylistic patterns, formatting artifacts, and model-specific errors, diluting the signal for specific capability improvement. SPD extracts a low-rank capability subspace from the model's own gradients on correctness-defining tokens.
The method projects key-value (KV) activations into this subspace during self-generation and fine-tunes on the resulting raw outputs with standard next-token prediction loss. This approach achieves generalizable, capability-selective improvement without any external signal, outperforming state-of-the-art self-distillation methods by up to 13% without external signals and by up to 16% over pre-trained baselines.
Notably, SPD demonstrates superior generalizability, achieving 15% better performance under out-of-domain generalization settings, making it particularly valuable for applications where models need to adapt to new domains or tasks. The capability-selective nature of the approach ensures that improvements are focused and meaningful, rather than being diluted by irrelevant variations in the generated outputs.
Key insight: SPD achieves generalizable, capability-selective self-distillation without external signals by extracting low-rank capability subspaces from model gradients and projecting KV activations into these subspaces during self-generation.
Moral Semantics Survive Machine Translation: Cross-Lingual Evidence from Moral Foundations Corpora
arXiv: 2605.22660
This research investigates whether LLM-based translation can bridge the gap in moral values research, where language-specific annotated corpora exist almost exclusively in English. Using ~50k morally-annotated social media posts from Polish, the study applies a principled four-method validation pipeline including LaBSE cross-lingual embedding similarity, Centered Kernel Alignment (CKA), LLM-as-judge evaluation, and deep learning classifier parity tests.
Despite shortcomings in handling slang, vulgarity, and culturally-loaded expressions, direct translation preserves subtle moral cues well enough to be harvested by cross-lingual machine learning. The results show that translation maintains mean cosine similarity of 0.86 and AUC gaps of 0.01-0.02 across all moral foundations, closing further under fine-tuning of language models. This demonstrates that machine translation is a practical and cost-effective path to moral values research in languages currently under-resourced in this domain.
The findings are particularly significant for Polish as a representative Slavic language, with expected generalization to related languages, providing a foundation for cross-lingual moral values research. This approach not only addresses the resource gap in moral values research but also opens new possibilities for understanding how moral language varies across cultures and languages, potentially leading to more inclusive and globally representative AI systems.
Key insight: Machine translation preserves subtle moral cues well enough for cross-lingual machine learning applications, with mean cosine similarity of 0.86 and AUC gaps of 0.01-0.02 across all foundations, demonstrating that translation can bridge language gaps in moral values research.
Seeing the Poem: Image-Semantic Detection of AI-Generated Modern Chinese Poetry with MLLMs
arXiv: 2605.22654
This paper evaluates and enhances LLM performance as detectors for modern Chinese poetry by innovatively incorporating images that reflect the content of the poetry, addressing a gap in previous detection studies that focused primarily on English poetry. The approach uses example-driven methods to integrate information such as meaning, imagery, and feeling from images, then forms a complementary judgment with the poem text.
Experimental results demonstrate that LLM detectors based on this method outperform baseline detectors based on plain text and even surpass the best-performing traditional detector, RoBERTa. The Gemini detector using this method achieves a Macro-F1 score of 85.65%, reaching state-of-the-art level, proving the effectiveness of the image-semantic guidance approach for poetry detection.
The performance improvements across multiple LLM detectors on LLM-generated data prove the method's effectiveness, showing that combining visual and textual information significantly enhances detection accuracy for modern Chinese poetry. This approach represents a novel direction for poetry detection that could be extended to other forms of culturally specific text generation.
Key insight: Image-semantic guided poetry detection using LLMs achieves state-of-the-art performance (85.65% Macro-F1) by integrating meaning, imagery, and feeling from images with poem text, outperforming baseline detectors based on plain text.
Boiling the Frog: A Multi-Turn Benchmark for Agentic Safety
arXiv: 2605.22643
Boiling the Frog introduces a benchmark that evaluates whether tool-using AI models deployed in corporate and office settings are susceptible to incremental attacks, where scenarios begin with benign workspace edits and later introduce risk-bearing requests. The benchmark focuses on stateful multi-turn evaluation with persistent workspaces, placing risk-bearing payloads at controlled positions in the turn sequence.
The results show that across a nine-model panel, aggregate strict attack success rate (ASR) is 44.4%, with model-level ASR ranging from 20.5% for Claude Haiku 4.5 to 92.9% for Gemini 3.1 Flash Lite. Average chain category-level ASR reaches 93.3% for Code of Practice loss-of-control scenarios, demonstrating that even sophisticated models are vulnerable to incremental attacks that exploit the persistent state of agent interactions.
The benchmark's three-level operational risk taxonomy grounded in the AI Act Annex I and Annex III high-risk contexts, and EU AI Act's Code of Practice on General-Purpose AI provides a comprehensive framework for evaluating agent safety. This work highlights the critical need for safety evaluation that goes beyond simple prompt-based testing to examine how agents behave in complex, multi-turn environments where state persistence creates new attack vectors.
Key insight: Boiling the Frog benchmark reveals that tool-using AI models are susceptible to incremental attacks, with aggregate strict attack success rate of 44.4% across nine models, highlighting the need for robust safety evaluation in multi-turn agent interactions.
More Context, Larger Models, or Moral Knowledge? A Systematic Study of Schwartz Value Detection in Political Texts
arXiv: 2605.22641
This systematic study examines when context and explicit moral knowledge help with sentence-level value detection, comparing sentence, window, and full-document inputs; no-RAG and retrieval-augmented settings with a curated moral knowledge base; supervised DeBERTa-v3-base/large encoders; and zero-shot LLMs from 12B to 123B parameters. The research reveals that while more context is not uniformly better, retrieved moral knowledge is more consistently useful in matched comparisons.
The findings show that full-document context improves supervised DeBERTa encoders by 3.8-4.8 macro-F1 points over sentence-only input, but does not consistently help zero-shot LLMs. Retrieved moral knowledge is more consistently useful in matched comparisons, improving each tested model family and context condition under early fusion. However, scaling from DeBERTa-v3-base to large and from 12B to larger LLMs does not guarantee gains.
Per-value analyses show that context and retrieval help most for socially situated or conceptually confusable values, suggesting that value-sensitive NLP should evaluate context, knowledge, and model family jointly rather than treating longer inputs or larger models as universal improvements. This research provides crucial insights for designing effective value detection systems that appropriately leverage context and knowledge resources.
Key insight: Context and explicit moral knowledge both help with Schwartz value detection, but more context is not uniformly better - full-document context improves supervised DeBERTa encoders by 3.8-4.8 macro-F1 points over sentence-only input, but does not consistently help zero-shot LLMs.
Integrable Elasticity via Neural Demand Potentials
arXiv: 2605.22820
ICDN represents a novel approach to multiproduct retail demand modeling by learning log-demand as a smooth, context-conditioned function of log-prices, allowing elasticities to be derived exactly from the learned demand surface. This approach addresses limitations of traditional demand models that often require parametric assumptions or cannot handle complex, context-dependent relationships.
The model's ability to provide exact elasticity estimates, especially for weakly identified cross-price effects, represents a significant advancement in demand modeling. On the Dominick's beer dataset, ICDN improves out-of-sample generalization over a directed log-log benchmark and yields more stable, economically plausible elasticity estimates, demonstrating the practical value of the approach.
This work contributes to the broader field of demand modeling by providing a framework that combines neural network flexibility with exact mathematical derivations, enabling more accurate and reliable demand predictions. The approach's ability to handle complex, context-dependent relationships while maintaining mathematical rigor makes it particularly valuable for applications in retail, pricing, and market analysis.
Key insight: Integrable Context-Dependent Demand Network (ICDN) learns log-demand as a smooth, context-conditioned function of log-prices, enabling exact derivation of elasticities from the learned demand surface for multiproduct retail demand.
Vector Policy Optimization: Training for Diversity Improves Test-Time Search
arXiv: 2605.22817
Vector Policy Optimization (VPO) addresses a critical limitation in standard LLM post-training where optimizing for a pre-specified scalar reward leads to low-entropy response distributions and struggles with diverse downstream tasks. VPO explicitly trains policies to anticipate diverse downstream reward functions and produce diverse solutions by leveraging that rewards are often vector-valued in practice.
The approach is essentially a drop-in replacement for the GRPO advantage estimator but trains the LLM to output a set of solutions where individual solutions specialize to different trade-offs in the vector reward space. This enables effective test-time search procedures like AlphaEvolve that select rollouts with a variety of task-specific reward functions, which is crucial for generalizing to novel environments.
Across four tasks, VPO matches or beats the strongest scalar RL baselines on test-time search (e.g. pass@k and best@k), with the gap widening as the search budget grows. For evolutionary search, VPO models unlock problems that GRPO models cannot solve at all. This demonstrates that optimizing for diversity may need to become the default post-training objective as test-time search becomes more standardized, representing a fundamental shift in how LLMs are trained for generalization.
Key insight: Vector Policy Optimization (VPO) trains policies to anticipate diverse downstream reward functions and produce diverse solutions, outperforming scalar RL baselines on test-time search with the gap widening as search budget grows.
Remember to be Curious: Episodic Context and Persistent Worlds for 3D Exploration
arXiv: 2605.22814
This work addresses the challenge of curiosity-driven reinforcement learning in complex, photorealistic 3D environments where agents can become trapped in local loops and receive fresh rewards for revisiting forgotten states. The key insight is that effective curiosity requires a model of the world that is persistent and continuously updated, paired with an agent that maintains an episodic trajectory history to navigate toward novel regions.
The approach uses an online 3D reconstruction as a persistent model of the world, while the agent policy is parameterized as a sequence model over RGB observations to maintain episodic context. This design enables effective exploration during training while allowing the agent to navigate using solely RGB frames at deployment, demonstrating the importance of both persistent world models and episodic context for successful exploration.
Trained purely via curiosity on HM3D, the agent outperforms RL-based active mapping baselines and generalizes zero-shot to Gibson and AI-generated worlds. The end-to-end policy enables efficient adaptation to downstream tasks such as apple picking and image-goal navigation, outperforming from-scratch baselines. This work demonstrates that combining persistent world models with episodic context is crucial for effective exploration in complex 3D environments.
Key insight: Effective curiosity in 3D exploration requires persistent world models and episodic trajectory history, with online 3D reconstruction and sequence modeling enabling efficient exploration and zero-shot generalization to new environments.
The Matching Principle: A Geometric Theory of Loss Functions for Nuisance-Robust Representation Learning
arXiv: 2605.22800
The Matching Principle offers a novel geometric theory that unifies various approaches to robustness, domain adaptation, and compositional generalization under a single framework. It argues that much of their shared structure is one statistical problem: estimating the covariance of label-preserving deployment nuisance and regularizing the encoder Jacobian along a matrix whose range covers that covariance.
The theory demonstrates that CORAL, adversarial training, IRM, augmentation, metric learning, Jacobian penalties, and alignment-style constraints are different estimators of the same object, not independent robustness tricks. The paper provides closed-form optimality theorems and consistency lemmas for estimation under standard identifiability assumptions, offering a falsifiable theory once the deployment nuisance covariance is identified.
Thirteen pre-registered blocks from classical ML through Qwen2.5-7B test the predicted matched, then isotropic, then wrong-W ordering on geometry and deployment drift, with twelve passing and the sole exception being an eigengap failure. At 7B scale, matched style-PMH improves selective honesty and preserves Style TDI where standard DPO degrades it, demonstrating the practical utility of the theoretical framework in real-world applications.
Key insight: The Matching Principle provides a unified geometric theory for loss functions that regularizes encoder Jacobian along a matrix whose range covers the covariance of label-preserving deployment nuisance, unifying various robustness approaches under a single framework.
SDPM: Survival Diffusion Probabilistic Model for Continuous-Time Survival Analysis
arXiv: 2605.22776
SDPM introduces a novel generative approach to continuous-time survival analysis that models the conditional distribution of survival outcomes using a denoising diffusion model, avoiding the structural assumptions on hazard functions and discretization of time axes that limit existing methods. This approach operates in a transformed target space using standardized log-times and continuous Gaussian-mixture representation of censoring indicators.
The model's formulation allows conditional samples generated by the model to be transformed into survival function estimates using the Kaplan-Meier estimator, providing a bridge between generative modeling and traditional survival analysis methods. This approach avoids parametric assumptions on the event-time distribution and does not require discretization of the output time space, making it more flexible than traditional methods.
Evaluation on ten real survival datasets shows that SDPM achieves competitive predictive performance across C-index, integrated time-dependent AUC, and integrated Brier score. A study on synthetic Cox-Weibull data demonstrates that SDPM can recover the shape of an underlying continuous survival distribution more accurately than a strong nonparametric baseline when sufficiently many samples are generated, showing the method's effectiveness in handling complex survival data.
Key insight: SDPM provides a generative approach to continuous-time survival analysis using denoising diffusion models, avoiding parametric assumptions on event-time distribution and discretization of output time space while achieving competitive predictive performance.
MambaGaze: Bidirectional Mamba with Explicit Missing Data Modeling for Cognitive Load Assessment from Eye-Gaze Tracking Data
arXiv: 2605.22775
MambaGaze tackles two major challenges in cognitive load assessment from eye-tracking signals: handling frequent data missingness from blinks and tracking failures, and efficiently modeling long-range temporal dependencies. The framework introduces XMD encoding that augments raw features with observation masks and time-deltas to explicitly model data uncertainty, and bidirectional Mamba-2 that captures temporal dependencies with linear computational complexity.
Experiments on CLARE and CL-Drive datasets under leave-one-subject-out evaluation show that MambaGaze achieves 76.8% and 73.1% accuracy, respectively, outperforming CNN, Transformer, ResNet, and VGG baselines by 4-12 percentage points. The approach demonstrates superior performance in handling the inherent challenges of eye-tracking data, including missing values and long temporal dependencies.
Edge deployment benchmarks on NVIDIA Jetson platforms demonstrate real-time inference at 43-68 FPS with power consumption below 7.5W, confirming feasibility for wearable cognitive load monitoring. This represents a significant advancement in making real-time cognitive load assessment practical for wearable applications, with the potential for use in safety-critical applications such as driver vigilance monitoring or automated flight deck assistance.
Key insight: MambaGaze addresses challenges in cognitive load assessment from eye-tracking signals by combining XMD encoding for explicit data uncertainty modeling with bidirectional Mamba-2 for efficient long-range temporal dependency modeling, achieving state-of-the-art accuracy.
CogAdapt: Transferring Clinical ECG Foundation Models to Wearable Cognitive Load Assessment via Lead Adaptation
arXiv: 2605.22774
CogAdapt addresses the challenge of applying clinical ECG foundation models to wearable cognitive load assessment, where sensor configuration mismatch and task differences prevent direct application of pre-trained models. The framework introduces LeadBridge, a learnable adapter that transforms 3-lead wearable signals into anatomically consistent 12-lead representations, and ProFine, a progressive fine-tuning strategy that gradually unfreezes encoder layers while preventing catastrophic forgetting.
Evaluations on two public datasets (CLARE and CL-Drive) under leave-one-subject-out cross-validation show that CogAdapt substantially outperforms baselines trained from scratch, achieving macro-F1 scores of 0.626 and 0.768. This demonstrates the effectiveness of foundation model adaptation for subject-independent cognitive load assessment from wearable sensors, representing a significant advancement in wearable health monitoring.
The approach shows that carefully designed adaptation strategies can close much of the utility gap between clinical and wearable applications, highlighting a promising direction for future research in foundation model adaptation. This work demonstrates that the rich representations learned from clinical ECG data can be effectively transferred to wearable applications, opening new possibilities for real-time cognitive load assessment in various settings.
Key insight: CogAdapt enables transfer of clinical ECG foundation models to wearable cognitive load assessment through LeadBridge adapter and ProFine progressive fine-tuning, achieving substantial performance improvements over baselines trained from scratch.
Uniform Diffusion Models Revisited: Leave-One-Out Denoiser and Absorbing State Reformulation
arXiv: 2605.22765
This work reveals a fundamental mismatch in Uniform Diffusion Models (UDM) where the standard plug-in bridge parameterization is not optimized by the denoising posterior but by a leave-one-out posterior that predicts each clean token without using its own noisy observation. This identification of the mismatch between plug-in ELBO and usual cross-entropy denoising objective provides crucial insights into UDM training.
The research characterizes the leave-one-out target and derives exact conversions between the denoiser, the leave-one-out posterior, and the score, allowing disentanglement of parameterization and training objective. These conversions lead to inference improvements without additional training through informed predictor-corrector samplers and improved temperature sampling based on the leave-one-out predictor.
The study introduces an absorbing-state reformulation of uniform diffusion that preserves the UDM joint law while decomposing it into masked-diffusion-like sampling operations with simpler denoising posteriors, carry-over unmasking, and natural remasking mechanism. On language modeling, leave-one-out parameterizations consistently improve UDM generation, while the absorbing construction matches or surpasses masked diffusion, suggesting that the empirical gap between masked and uniform diffusion is driven less by the choice of marginals themselves than by parameterization and sampling design.
Key insight: Uniform Diffusion Models reveal that standard plug-in bridge parameterization is not optimized by the denoising posterior but by a leave-one-out posterior, leading to improved inference through informed predictor-corrector samplers and temperature sampling.
Lumberjack: Better Differentially Private Random Forests through Heavy Hitter Detection in Trees
arXiv: 2605.22756
Lumberjack addresses the challenge of achieving high utility in differentially private random forests, where existing approaches typically degrade performance to the point of impracticality. The approach constructs large random decision trees and then applies aggressive, privacy-preserving pruning to retain only sufficiently populated nodes, significantly improving the privacy-utility trade-off.
A key innovation is a novel (ε,δ)-DP heavy hitter detection algorithm for hierarchical data with error scaling as O_{ε,δ}(√log h) for trees of height h, enabling the use of significantly deeper trees than in prior work. This favorable scaling allows for improved expressiveness under privacy constraints, leading to better performance than prior DP random forest methods.
Empirical evaluation on benchmark datasets shows that Lumberjack consistently outperforms prior DP random forest methods, establishing a new state of the art. The approach yields substantial improvements in the privacy-utility trade-off for practical privacy budgets, demonstrating that carefully designed DP random forests can close much of the utility gap and represent a promising direction for future research in privacy-preserving machine learning.
Key insight: Lumberjack achieves substantially higher utility in differentially private random forests by constructing large random decision trees and applying aggressive privacy-preserving pruning to retain only sufficiently populated nodes.
Cyber-Physical Anomaly Detection in IoT-Enabled Smart Grids Using Machine Learning and Metaheuristic Feature Optimization
arXiv: 2605.22749
This work investigates cyber-physical anomaly detection in IoT-enabled smart grids using machine learning combined with genetic-algorithm-based feature selection. The approach aims to classify attack and natural events accurately while determining whether a reduced set of physically informative PMU/IED measurements can support reliable detection.
The evaluation shows that tree-based ensemble models are most effective for the considered dataset, with Extra Trees providing the strongest full-feature baseline. After feature selection, the GA + Extra Trees model reduces the clean PMU feature space from 112 attributes to an average of 27.4 attributes over five runs, while increasing macro-F1 from 0.9118 to 0.9212 and ROC-AUC from 0.9791 to 0.9837.
These results indicate that many synchronized electrical measurements are redundant, and a compact subset of phasor-based features can still provide accurate and interpretable anomaly detection in smart grids. The approach demonstrates that feature optimization can significantly improve both detection accuracy and computational efficiency, making it practical for real-world smart grid applications.
Key insight: Metaheuristic feature optimization combined with machine learning improves cyber-physical anomaly detection in smart grids, reducing feature space from 112 to 27.4 attributes while increasing macro-F1 from 0.9118 to 0.9212.
Plug-in Losses for Evidential Deep Learning: A Simplified Framework for Uncertainty Estimation that Includes the Softmax Classifier
arXiv: 2605.22746
This research addresses computational challenges in Evidential Deep Learning (EDL) where Dirichlet expected objectives are more complex than standard supervised learning losses, complicating analysis and implementation. The approach approximates the objective of the first-order empirical risk minimization problem induced by EDL with a plug-in loss evaluated at the Dirichlet mean.
Under mild assumptions, the approximation error decays with growing evidence for a broad class of loss functions, including mean-squared error and cross-entropy loss. The analysis provides justification for the use of softmax in the context of uncertainty estimation, since under a particular evidence-to-Dirichlet mapping, the framework includes the standard softmax classifier as a special case.
Validation on the Google Speech Commands dataset shows that the proposed simplified objectives achieve predictive accuracy and selective prediction performance comparable to classical EDL, while being simpler to implement using standard deep learning losses and training pipelines. This represents a significant advancement in making EDL more accessible and practical for real-world applications requiring uncertainty estimation.
Key insight: Plug-in losses for evidential deep learning provide a simplified framework that approximates first-order empirical risk minimization with a plug-in loss evaluated at the Dirichlet mean, achieving comparable performance to classical EDL while being simpler to implement.
Self-Evolving Multi-Agent Systems via Decentralized Memory
arXiv: 2605.22721
DecentMem addresses limitations of centralized memory in self-evolving multi-agent systems by proposing a decentralized framework where each agent maintains its own dual-pool memory - an exploitation pool of consolidated past trajectories and an exploration pool of LLM-generated candidates for unseen contexts. The two pools are reweighted online based on stage-wise feedback from an LLM-as-a-judge.
Theoretical analysis proves that this design guarantees global reachability of the solution space and achieves O(log T) cumulative regret, matching the stochastic bandit lower bound up to constants. This provides strong theoretical foundations for the decentralized approach, demonstrating its effectiveness in balancing exploration and exploitation.
Practical evaluation across three MAS frameworks, three Qwen3 backbones, and five benchmarks spanning math, code, QA, and embodied tasks shows that DecentMem improves average accuracy by up to 23.8% over the strongest centralized memory baseline and by up to 52.5% over the no-memory baseline, while reducing token usage by up to 49%. This demonstrates the practical advantages of decentralized memory in multi-agent systems.
Key insight: DecentMem framework enables self-evolving multi-agent systems through decentralized memory where each agent maintains dual-pool memory (exploitation and exploration) with online reweighting based on LLM-as-a-judge feedback, achieving up to 52.5% improvement over no-memory baseline.
A Generalized Nash Equilibrium-Seeking Scheme for Trauma Resuscitation
arXiv: 2605.22661
This work formulates trauma resuscitation as a distributed generalized Nash equilibrium (GNE)-seeking game with coupled inequality constraints, providing a novel socio-technical approach to understanding and optimizing healthcare worker decisions in safety-critical environments. The framework addresses the challenge of designing and optimizing quantifiable metrics that accurately capture HCW decisions.
The method is optimized over a time-varying communication graph and introduces novel insights from clinical experience to model HCWs behavior, facilitating the best possible resuscitation outcome given HCWs workloads, schedules, competencies, and limited resources. This approach represents a significant advancement in applying game theory to healthcare decision-making.
The work demonstrates how mathematical optimization can be applied to complex clinical processes, potentially improving patient outcomes by better understanding how healthcare workers make decisions under resource constraints and time pressures. This framework provides a foundation for developing more effective resuscitation protocols and training systems.
Key insight: Socio-technical formulation of trauma resuscitation as a distributed generalized Nash equilibrium (GNE)-seeking game with coupled inequality constraints provides a framework for optimizing healthcare worker decisions under time-varying communication graphs and resource constraints.
Sibyl-AutoResearch: Autonomous Research Needs Self-Evolving Trial-and-Error Harnesses, Not Paper Generators
arXiv: 2605.22343
Sibyl-AutoResearch addresses the limitation that executable workflows alone do not produce research judgment, identifying where current systems lose trial experience: weak evidence becomes prose, pilot signals become broad claims, memory remains textual, and recurring process failures do not change later behavior. The framework introduces Scientific Trial-and-Error Harnesses that let agents run bounded trials, preserve outcomes, and route lessons into later planning.
The approach formalizes this through two auditable conversion units: trial-to-behavior conversion, which links trial signals to later research actions, and trial-to-harness-behavior conversion, which links recurring process failures to system updates. The framework is implemented in SIBYL, a file-backed autonomous research system that exposes state, roles, memory, gates, and artifact traces needed to inspect these conversion paths.
A retrospective audit identifies eight high-confidence conversion events with median latency of one iteration and maximum latency of three iterations, showing that the proposed conversion units are recoverable from realistic autonomous-research workspaces. The recovered-failure registry further shows how five naturally occurring failure classes were blocked, downgraded, or routed into later repair, demonstrating the practical effectiveness of the framework.
Key insight: Sibyl-AutoResearch framework built around Scientific Trial-and-Error Harnesses enables autonomous research systems to learn from trial experience, preserve positive and negative outcomes, and route lessons into later planning and validation.
ACCoRD: Actor-Critic Conflict Resolution with Deep learning for O-RAN xApps
arXiv: 2605.22306
ACCoRD addresses Conflict Mitigation (ConMit) in Open Radio Access Networks (O-RAN) by proposing a method that resolves detected control conflicts using a Conflict Resolution (CR) Agent with an Artificial Neural Network (ANN) trained with a reinforcement learning algorithm PPO-Clip. The CR Agent analyzes data about the network and conflicting control decisions to infer optimal CR actions.
The approach uses feedback from the network after each resolved conflict to assess efficiency and adjust the ANN's weights during batch training, creating an adaptive system that improves over time. The evaluation based on simulation data shows that the proposed ANN-based method improves on the efficiency of rule-based approaches by significantly reducing negative network events caused by conflicting control decisions in medium and high traffic scenarios.
This work demonstrates the practical application of deep learning for conflict resolution in complex network environments, providing a scalable solution for managing conflicts in O-RAN systems. The method's ability to learn from experience and adapt its behavior makes it particularly valuable for dynamic network environments where conflicts are common and require rapid resolution.
Key insight: ACCoRD method using reinforcement learning with PPO-Clip for conflict resolution in O-RAN xApps significantly reduces negative network events caused by conflicting control decisions in medium and high traffic scenarios.
Emergence of agriculture in an artificial society of reinforcement learning agents
arXiv: 2605.22256
This work introduces an artificial society of reinforcement learning agents embedded in a dynamic ecological environment to identify general principles underlying the emergence of agriculture as a major evolutionary transition. The system demonstrates that agricultural practices emerge spontaneously without explicit instruction through the coupled dynamics of learning and environmental modification.
The research identifies four key ingredients governing this transition: individual planning through the valuation of delayed rewards, social vulnerability to cheaters, stabilization via social learning, and an emergent lock-in effect that renders agriculture effectively irreversible once established. The study shows that social learning acts as a 'firewall' that suppresses cheater invasion and enables propagation of successful strategies, leading to sustained population growth and nonlinear amplification of domesticated resources.
These results reveal universal mechanisms linking individual decision-making, social interactions, and ecological feedbacks, highlighting the potential of artificial societies as experimental platforms to study emergence of cultural innovations and major evolutionary transitions. The work demonstrates that complex collective behaviors can emerge from simple interactions, providing insights into how human societies might have developed complex practices like agriculture.
Key insight: Artificial society of reinforcement learning agents spontaneously develops agricultural practices through coupled dynamics of learning and environmental modification, with four key ingredients enabling the transition: individual planning, social vulnerability to cheaters, stabilization via social learning, and emergent lock-in effect.
Planning, Scheduling, and Behavior in EV Charging Systems: A Critical Survey and Trilemma Framework
arXiv: 2605.21665
This survey develops a three-layer Planning-Scheduling-Behavior (PSB) framework to organize EV charging research according to decision horizon, actor objective, and coupling structure, identifying a fidelity-tractability tradeoff termed the PSB trilemma. The trilemma shows that each layer is computationally difficult in isolation, and realistic integration across layers generally requires reducing the fidelity of at least one layer.
Reviewing the three pairwise-coupling literatures - Planning-Scheduling, Scheduling-Behavior, and Planning-Behavior - the study shows that the omitted third layer is typically fixed exogenously or represented by a static aggregate surrogate. These simplifications enable tractability but impose distinct costs: they can obscure long-term investment feedback, temporal grid and emissions dynamics, or heterogeneous user response and equity outcomes.
The survey identifies open challenges in emerging charging technologies, behavioral incentives, equity metrics, and city-scale learning-based methods that balance fidelity, interpretability, and policy relevance. This work provides crucial insights for researchers and practitioners working on EV charging systems, emphasizing the need for careful consideration of trade-offs in cross-layer integration approaches.
Key insight: PSB trilemma reveals that each layer of EV charging research is computationally difficult in isolation, and realistic integration across layers generally requires reducing the fidelity of at least one layer, highlighting the need for balanced approaches to cross-layer integration.
Argo: Efficient Importance Labeling for Enterprise Email Systems
arXiv: 2605.21604
Argo addresses the challenge of email importance labeling at enterprise scale, where traditional approaches like keyword matching and user-defined rules demand extensive manual feature engineering and fail to scale effectively. The framework explores the trade-off space of using alternative labeling schemes versus GPT-4.1 scale LLMs to achieve near GPT level labeling quality with significantly lower cost.
The approach constructs a profiler to efficiently search the cost-quality trade-off space of labeling and identifies cost-efficient alternatives to labeling emails. Additionally, it designs an on-demand provisioning scheme to intelligently scale Argo with real-time load, minimizing cost increases during peak load inference. This enables large-scale, context-aware email labeling practical for enterprises.
Over three open-source email datasets, Argo achieves 148-167X inference cost reduction with negligible quality degradation and 20-640000X lower profiling costs, demonstrating the practical viability of cost-effective LLM-based email labeling. The framework represents a significant advancement in making sophisticated AI capabilities accessible to enterprise users without prohibitive computational costs.
Key insight: Argo framework achieves 148-167X inference cost reduction with negligible quality degradation by constructing a profiler to search the cost-quality trade-off space and implementing on-demand provisioning for real-time load scaling.
Superhuman Safe and Agile Racing through Multi-Agent Reinforcement Learning
arXiv: 2605.22748
This work demonstrates that multi-agent reinforcement learning provides the essential safety scaffolding required for real-world interaction, contrasting with the dominant single-agent paradigm where other actors are ignored or treated as environmental noise. The approach uses high-speed quadrotor racing as a high-stakes testbed to train agents that navigate complex aerodynamic interactions and strategic maneuvering with variable numbers of racers.
Through league-based self-play, agents evolve sophisticated anticipatory behaviors including proactive collision avoidance, overtaking, and handling multi-agent physical interactions, including aerodynamic downwash. The agents outperform a champion-level human pilot in multi-player races at speeds exceeding 22 m/s, while simultaneously reducing collision rates by 50% compared to state-of-the-art single-agent baselines.
Crucially, training with diverse artificial agents enables zero-shot generalization to safer human interaction, suggesting that the path to robust robotic co-existence lies not in isolated safety constraints, but in the rigorous demands of multi-agent interaction. The work demonstrates that multi-agent systems can achieve superhuman performance while maintaining safety, representing a significant advancement in autonomous systems for real-world applications.
Key insight: Multi-agent reinforcement learning provides essential safety scaffolding for real-world interaction, enabling agents to evolve sophisticated anticipatory behaviors including proactive collision avoidance, overtaking, and handling multi-agent physical interactions.
Cross-domain benchmarks reveal when coordinated AI agents improve scientific inference from partial evidence
arXiv: 2605.22300
This research evaluates when coordinated AI agents add value over simpler scientific workflows using a cross-domain benchmark spanning four scientific tasks: mapping molecular structure into musical representations, detecting historical paradigm shifts in science, identifying vector-borne disease emergence, and vetting transiting-exoplanet candidates. Each case uses a frozen evaluation panel, predefined scoring protocols, explicit baselines, ablations or null controls, and stated limitations.
The results define three operating regimes where coordination improves performance. When different disciplines each capture only part of the phenomenon, cross-channel composites improve over single-channel baselines: climate-vector emergence reaches AUROC 0.944 and exoplanet vetting reaches AUROC 0.955. However, the exoplanet workflow is effectively tied with a strong combined-summary baseline, showing that decomposition does not always improve top-line performance.
The study demonstrates that coordination mainly improves interpretation and traceability when one signal dominates, as in paradigm-shift detection, and provides representational rather than predictive gains for molecular sonification. ScienceClaw x Infinite provides the auditable artifact and provenance layer for this evaluation, assigning value to coordination only when the corresponding performance, provenance, or representation claim is supported by explicit comparators.
Key insight: Cross-domain benchmarks reveal that coordinated AI agents improve scientific inference when different disciplines each capture only part of the phenomenon, with cross-channel composites improving over single-channel baselines in climate-vector emergence and exoplanet vetting.
The Log is the Agent: Event-Sourced Reactive Graphs for Auditable, Forkable Agentic Systems
arXiv: 2605.21997
ActiveGraph represents a fundamental shift from traditional agent frameworks where the language model comes first, then tools, then rules, and finally a logging layer bolted on for observability. Instead, the append-only event log is the source of truth; the working graph is a deterministic projection of that log; and behaviors react to changes in the graph and emit new events.
This single design decision yields three properties that retrieval-and-summarization memory systems do not provide: deterministic replay of any run from its log, cheap forking that branches a run at any event without re-executing the shared prefix, and end-to-end lineage from a high-level goal down to the individual model call that produced each artifact. The framework presents a determinism contract that makes replay sound and a worked diligence example whose full causal structure is reconstructable from the log alone.
The approach is unusually well suited to self-improving agents and extends the BabyAGI lineage and prior graph-memory research. By making the event log the foundation of agent behavior, ActiveGraph provides a robust framework for building auditable, forkable agentic systems that can be reliably reproduced and modified, representing a significant advancement in agent architecture design.
Key insight: ActiveGraph framework inverts traditional agent design by making the append-only event log the source of truth, enabling deterministic replay, cheap forking, and end-to-end lineage from high-level goals to individual model calls.
AI-Enabled Serious Games: Integrating Intelligence and Adaptivity in Training Systems
arXiv: 2605.21962
This chapter examines how contemporary AI approaches may support real-time instructional adaptation in serious games, distinguishing between instructional intelligence (capacity to infer learner knowledge and reason about pedagogically appropriate responses) and adaptivity (ability to modify instructional actions during interaction). The work traces developments from early computer-assisted instruction through intelligent tutoring systems (ITS), dynamic difficulty adjustment (DDA), authoring platforms, learning analytics, and recent AI-enabled architectures.
The research highlights how large language models (LLMs), reinforcement learning (RL), and agent-based architectures may contribute to more integrated forms of intelligence and adaptivity in serious games. It discusses practical and research challenges associated with AI-enabled systems, including explainability, validation, computational cost, and the limited empirical evidence regarding long-term learning outcomes in AI-enabled serious games.
The chapter provides a comprehensive synthesis of adaptive learning systems, demonstrating how AI can address persistent challenges in serious games such as static scenario design, authoring bottlenecks, limited learner modeling, and difficulty implementing meaningful real-time instructional adaptation. This work represents a significant step toward more sophisticated and effective training systems using AI technologies.
Key insight: AI integration in serious games enables dynamic scenario variation, contextual feedback, adaptive pacing, and learner-state modeling, addressing persistent challenges in static scenario design and authoring bottlenecks.