This week's research highlights include MOSS's breakthrough in autonomous agent self-evolution through source-level rewriting, Google Cloud's AI-powered video editing platform, and significant advances in multi-agent systems and mathematical proof search. These developments demonstrate AI's growing capability to automate complex creative and analytical tasks while addressing critical safety and alignment concerns.
Google Cloud's Glance is leveraging AI to automatically transform hours of raw video content into mobile-ready clips, significantly reducing manual editing time and enabling scalable content production. Meanwhile, MOSS introduces a novel approach to autonomous agent evolution by operating at the code level rather than just text artifacts, achieving deterministic, Turing-complete evolution that can fix structural failures inaccessible through text-only methods. The research also covers advances in multi-agent systems with LCGuard's representation-level transformation framework for safe KV cache sharing, and Gated DeltaNet-2's decoupling of erase and write operations in linear attention mechanisms. Additionally, AI-driven formal proof search is showing remarkable success in resolving open mathematical problems, while new frameworks like HarnessAPI streamline LLM tool deployment and CogAdapt enables transfer of clinical ECG foundation models to wearable cognitive load assessment.
AI Model Releases
How Glance turns hours of video into mobile-ready clips with AI | Google Cloud Blog
Google Cloud's AI platform is being used by Glance to automatically transform hours of raw video content into mobile-ready clips. The system leverages machine learning models to identify key moments, cut scenes, and optimize content for mobile consumption. This process reduces manual editing time significantly, enabling content creators to produce multiple versions of videos quickly. The technology is designed to work with various video formats and can adapt to different content styles and audience preferences. Glance is using Google Cloud's AI infrastructure to scale this process across its content library.
Why it matters: This represents a significant step toward automated content creation workflows that could transform how media companies produce and distribute video content. The integration of AI into video editing pipelines suggests broader adoption of AI tools for creative industries, potentially reducing production costs and increasing output capacity.
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 introducing source-level rewriting as a mechanism for self-improvement. Unlike previous approaches that could only modify text-based artifacts like skill files or prompt configurations, MOSS operates at the code level, enabling changes to routing, hook ordering, state invariants, and dispatch logic that are fundamental to agent behavior. This approach addresses a critical gap in existing self-evolving systems where structural failures remain inaccessible through text-layer modifications.
The system's deterministic pipeline ensures that each evolutionary change is anchored to production failure evidence and verified through ephemeral trial workers before deployment. This rigorous validation process, combined with user-consent gated promotions and health-probe gated rollbacks, creates a robust framework for autonomous agent improvement that maintains system stability while enabling rapid iteration. The demonstration on OpenClaw showing a 0.36 improvement in grader scores without human intervention validates the practical effectiveness of this approach.
This work fundamentally challenges the assumption that agent evolution must be limited to text-based modifications. By treating source code as the primary medium for adaptation, MOSS opens new possibilities for agents that can autonomously restructure their own architecture in response to changing conditions, potentially enabling more sophisticated and resilient autonomous systems that can adapt to complex, unforeseen scenarios.
Key insight: MOSS demonstrates that source-level rewriting enables autonomous agents to self-evolve by modifying code structure rather than just text artifacts, achieving deterministic, Turing-complete evolution that can fix structural failures unreachable through text-only approaches.
Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention
arXiv: 2605.22791
Gated DeltaNet-2 addresses a fundamental limitation in existing linear attention mechanisms where scalar gates tied erase and write operations, creating suboptimal memory editing. By introducing separate channel-wise erase and write gates, the model achieves more precise control over memory dynamics, particularly beneficial for long-context scenarios where maintaining associations while forgetting irrelevant information is crucial. This architectural innovation allows for adaptive forgetting with channel-wise decay, improving performance across multiple benchmarks.
The theoretical foundation of Gated DeltaNet-2 includes a fast-weight update view and chunkwise WY algorithm with channel-wise decay absorbed into asymmetric erase factors, providing both mathematical rigor and computational efficiency. The gate-aware backward pass preserves efficient parallel training, making the approach practical for large-scale applications. The model's performance gains on long-context RULER benchmarks, particularly in multi-key retrieval settings, demonstrate the practical advantages of this decoupled approach over previous methods.
This advancement has significant implications for LLM efficiency and long-context processing capabilities. By improving memory editing precision while maintaining linear time complexity, Gated DeltaNet-2 enables more sophisticated attention mechanisms that can handle extended sequences without the computational overhead of traditional softmax attention, potentially enabling more capable language models for complex reasoning tasks.
Key insight: Gated DeltaNet-2 advances linear attention by decoupling erase and write operations through channel-wise gates, achieving superior performance on long-context tasks while maintaining computational efficiency through fast-weight updates and chunkwise algorithms.
LCGuard: Latent Communication Guard for Safe KV Sharing in Multi-Agent Systems
arXiv: 2605.22786
LCGuard tackles a critical safety concern in multi-agent LLM systems where KV cache sharing can inadvertently expose sensitive information from one agent to another. By treating shared KV caches as latent working memory and applying representation-level transformations before transmission, the framework addresses the opacity of cache-based communication channels that traditional natural language approaches cannot adequately secure.
The adversarial training formulation in LCGuard is particularly innovative, as it operationalizes sensitive information leakage through reconstruction - an agent-specific input can be recovered from shared cache artifacts. This approach enables the framework to learn transformations that preserve task-relevant semantics while minimizing reconstructable information, creating a practical solution for safe latent communication in multi-agent environments.
Empirical evaluations across multiple model families and benchmarks demonstrate LCGuard's effectiveness in reducing leakage and attack success rates while maintaining competitive task performance. This work is particularly relevant for the growing field of multi-agent systems where secure communication and information isolation are paramount, especially in applications involving sensitive data or regulated environments.
Key insight: LCGuard introduces a representation-level transformation framework that safely enables KV cache sharing in multi-agent systems by reducing reconstruction-based leakage while preserving task-relevant information through adversarial training.
Advancing Mathematics Research with AI-Driven Formal Proof Search
arXiv: 2605.22763
This work demonstrates the transformative potential of AI in mathematical research by showing that LLM-based agents can autonomously resolve open problems from the Erdős collection, proving 9 of 353 problems at a cost of a few hundred dollars per problem. The approach of alternating LLM-based generation with Lean-based verification creates a powerful pipeline that leverages the creative capabilities of LLMs while ensuring mathematical rigor through formal verification.
The scalability of this approach is remarkable - the agents proved 44 of 492 OEIS conjectures, showing that AI-assisted formal proof search can systematically advance mathematical knowledge. The demonstration that even basic agents can replicate successes at higher computational cost reveals the fundamental effectiveness of the methodology, while the superior performance of more capable agents suggests room for further improvement through better agent design.
This research has profound implications for mathematical research methodology, potentially enabling a new paradigm where AI agents can independently pursue mathematical conjectures and theorems. The approach also demonstrates how AI can be used to validate and extend mathematical knowledge, opening new avenues for collaboration between human mathematicians and AI systems in advancing mathematical understanding.
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 computational costs.
Towards a General Intelligence and Interface for Wearable Health Data
arXiv: 2605.22759
This work presents a significant advancement in wearable health data analysis by leveraging a foundation model pretrained on over one trillion minutes of unlabeled sensor signals from five million participants. The approach addresses fundamental challenges in health data analysis, including high phenotypic diversity, individual baseline variations, and the scarcity of high-quality labeled data through massive pretraining and scaling.
The demonstration that joint scaling of model capacity and pretraining data volume leads to systematic performance improvements across diverse health prediction tasks shows the power of large-scale pretraining in health applications. The ability to achieve label-efficient few-shot learning and generative capabilities for robust daily metric estimation represents a major step forward in making health data analysis more accessible and practical.
The deployment of a classroom of LLM agents to autonomously search downstream predictive heads and the validation through 1,860 clinician ratings demonstrate the practical utility of this approach. The integration of downstream predictors into a Personal Health Agent that supports more relevant, contextually aware, and safe responses shows how foundation models can be effectively leveraged for real-world health applications, potentially revolutionizing personalized healthcare delivery.
Key insight: A foundation model pretrained on massive wearable sensor data enables label-efficient few-shot learning and generative capabilities for health prediction, with downstream LLM agents autonomously searching for optimal predictive heads and improving performance with model capacity.
HarnessAPI: A Skill-First Framework for Unified Streaming APIs and MCP Tools
arXiv: 2605.22733
HarnessAPI addresses a critical inefficiency in LLM tool deployment where the same business logic must be implemented in both HTTP endpoints and MCP tool registrations, creating maintenance overhead and divergence issues. By treating a typed skill folder as the single source of truth, the framework eliminates this duplication and streamlines the development process.
The dual-mode content negotiation capability that allows the same handler to serve both SSE-streaming and JSON-returning clients without handler changes represents a significant improvement in developer experience and code maintainability. The dynamic code-generation mechanism ensuring Pydantic type annotations propagate correctly to FastMCP's inspection layer resolves a technical limitation that previously prevented naive closure-based registration.
The 74% reduction in framework-facing boilerplate compared to manual dual-stack implementations demonstrates the practical impact of this approach. By subclassing FastAPI and inheriting its full middleware, dependency-injection, and deployment ecosystem, HarnessAPI maintains compatibility with existing toolchains while providing substantial productivity gains for developers building LLM tools and agents.
Key insight: HarnessAPI eliminates duplication in LLM tool deployment by treating typed skill folders as single source of truth, automatically generating streaming HTTP endpoints, interactive UIs, and MCP tools from a unified framework that reduces boilerplate by 74%.
Can AI Make Conflicts Worse? An Alignment Failure in LLM Deployment Across Conflict Contexts
arXiv: 2605.22720
This work reveals a critical gap in AI alignment evaluation by demonstrating that LLMs deployed in conflict contexts can systematically produce outputs that exacerbate rather than mitigate conflicts. The finding that model choice becomes a safety question in itself, with some configurations failing 80-100% of the time when pushed for 'balance' in cases where international courts have already assigned responsibility, underscores the urgent need for specialized alignment frameworks.
The comprehensive evaluation across 90 multi-turn scenarios designed to surface misaligned behavior in conflict contexts provides a robust foundation for understanding how AI systems can fail in sensitive environments. The fact that failure rates vary significantly between different providers (OpenAI, Anthropic, DeepSeek, xAI) suggests that alignment approaches need to be tailored to specific deployment contexts rather than assuming universal safety properties.
The release of the first evaluation framework for this domain is particularly valuable for the AI safety community, as it provides a standardized approach for assessing alignment in conflict-sensitive applications. This work highlights that AI safety cannot be assumed to be a universal property but must be carefully evaluated for each specific application domain, especially those involving high-stakes social or political contexts.
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 need for specialized evaluation frameworks for conflict-sensitive applications.
AMEL: Accumulated Message Effects on LLM Judgments
arXiv: 2605.22714
The AMEL phenomenon reveals a fundamental limitation in LLM evaluation systems where the polarity of prior conversation history systematically biases subsequent judgments, with negative histories producing significantly stronger bias effects than positive ones. This finding challenges the assumption that LLMs maintain consistent performance across different contexts and highlights the importance of context management in evaluation protocols.
The observation that bias does not grow with context length - 5 prior turns and 50 produce the same shift - suggests that the bias is not simply a function of context length but rather a persistent characteristic of how LLMs process information. This finding has important implications for designing evaluation systems that maintain fairness and consistency across different types of interactions.
The proposed solutions, including using fresh context per item or balancing histories when batching is unavoidable, provide practical approaches for mitigating this bias. The finding that scaling helps but doesn't solve the problem indicates that this is a fundamental characteristic of LLM behavior rather than a scaling limitation, requiring systematic intervention in evaluation design rather than just model improvements.
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, suggesting that evaluation pipelines should use fresh contexts per item.
Beyond the Org Chart: AI and the Transformation of Invisible Work
arXiv: 2605.22707
This research provides valuable insights into how AI adoption is reshaping organizational dynamics beyond formal role boundaries, revealing that AI's impact extends to informal cultural practices that are crucial for professional development and engagement. The finding that AI enables smoother collaboration between peers while potentially undermining traditional mentoring and feedback mechanisms highlights the complex social implications of AI integration.
The identification of both positive and nuanced changes in how AI affects professional interactions suggests that organizations need to proactively address the potential risks to career development opportunities and informal networks. The proposed steps to make invisible work more visible and support colleagues through AI transformation are particularly relevant for maintaining healthy company cultures during AI adoption.
This work contributes to understanding how AI systems can be designed to support rather than undermine the social and cultural aspects of work that are essential for diverse thinking, collaboration, and informal interactions. The findings suggest that successful AI integration requires attention to both technical capabilities and social dynamics, particularly in preserving the informal practices that support career growth and professional development.
Key insight: AI adoption in organizations is transforming both formal role responsibilities and informal cultural practices like mentoring, creating both opportunities for smoother collaboration and risks to traditional career growth opportunities and professional networks.
Reducing Political Manipulation with Consistency Training
arXiv: 2605.22771
PCT addresses a subtle but significant form of political bias where LLMs handle counterpart topics from opposing political sides asymmetrically, revealing systematic patterns in rhetoric and framing that can be quantified and reduced. The identification of seven categories of techniques through which covert bias operates provides a comprehensive framework for understanding and addressing political bias in language models.
The two complementary paradigms of Sentiment Consistency Training and Helpfulness Consistency Training offer a nuanced approach to bias reduction that doesn't sacrifice the model's overall helpfulness. This is particularly important because political bias often manifests in ways that can make models appear less helpful or less engaging to users, making it crucial to maintain utility while reducing bias.
The demonstration that PCT preserves overall helpfulness while substantially reducing covert political bias shows that it's possible to address bias without compromising model performance. The generalization to held-out benchmarks indicates that the approach is robust and not just a local optimization, suggesting that consistency training can be a valuable tool for developing more balanced and fair language models.
Key insight: Political Consistency Training (PCT) reduces covert political bias in LLMs by addressing both sentiment consistency and helpfulness consistency through two complementary paradigms, achieving substantial bias reduction while preserving overall helpfulness.
Understanding Data Temporality Impact on Large Language Models Pre-training
arXiv: 2605.22769
This work demonstrates that the temporal ordering of data during pre-training significantly impacts the acquisition of time-sensitive factual knowledge in LLMs. The finding that sequentially trained models match shuffled baselines on general language understanding while consistently exhibiting more up-to-date and temporally precise knowledge reveals a fundamental trade-off in pre-training strategies.
The comprehensive benchmark of over 7,000 temporally grounded questions provides a robust evaluation framework for understanding how temporal grounding affects model performance. The observation that shuffled pre-training peaks on older data, possibly due to increased factual repetition, suggests that the choice of pre-training data ordering has important implications for model knowledge freshness and relevance.
The release of code, checkpoints, and datasets provides a foundation for future research on continual learning for LLMs, particularly in understanding how to maintain temporal grounding while preserving general language understanding. This work suggests that temporal ordering considerations should be integrated into pre-training strategies to ensure that models remain current with respect to time-sensitive information.
Key insight: Temporally ordered pre-training improves factual freshness and temporal precision while maintaining general language understanding, showing that sequential training can provide more up-to-date knowledge compared to shuffled pre-training approaches.
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 by introducing temporal components to disease associations, recognizing that clinical relevance changes over the course of a disease. This temporal grounding is essential for longitudinal clinical reasoning and retrieval augmentation, particularly for conditions where symptoms or diagnostic criteria evolve over time.
The multi-agent pipeline approach using frontier LLMs for knowledge extraction, combined with credibility filtering and ontology alignment, creates a robust framework for building temporal biomedical knowledge graphs. The demonstration that ChronoMedKG adds temporal grounding for 6,250 diseases absent from existing resources shows the practical impact of this approach.
The introduction of ChronoTQA benchmark with temporal and static controls provides a standardized way to evaluate temporal reasoning capabilities in biomedical systems. The finding that frontier LLMs lose roughly 30 points moving from static to temporal questions, while ChronoMedKG retrieval rescues 47-65% of long-tail failures, demonstrates the substantial value of temporal grounding for clinical reasoning systems.
Key insight: ChronoMedKG introduces temporal grounding to biomedical knowledge graphs, enabling clinical reasoning that accounts for disease progression and temporal relevance, with retrieval-augmented systems showing significant improvements in long-tail failures compared to static approaches.
Tokenization with Split Trees
arXiv: 2605.22705
ToaST represents a significant advancement in subword tokenization by directly optimizing compression through a novel recursive inference procedure that greedily splits pretokens into full binary trees using precomputed byte n-gram counts. This approach eliminates the need for vocabulary selection to be tied to specific tokenization algorithms, providing a more principled method for tokenization.
The Integer Program formulation for vocabulary selection that minimizes total token count under the inference procedure provides theoretical guarantees for near-optimal vocabularies, with training time scaling quadratically in the number of split trees. This mathematical foundation, combined with the practical performance gains, makes ToaST a compelling approach for tokenization that can be scaled to large vocabularies.
The empirical demonstration that ToaST achieves the highest CORE score among baselines, outperforming them by 2.6%-7.6% on 22 individual tasks, shows that the improved tokenization directly translates to better model performance. The substantial improvement in Renyi efficiency, where common single-byte tokens are used less frequently, indicates that the approach produces more efficient representations that can extend effective context length.
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, extending effective context length.
Self-Policy Distillation via Capability-Selective Subspace Projection
arXiv: 2605.22675
SPD addresses a fundamental limitation in self-distillation where existing methods either rely on costly external signals or train on all raw outputs, diluting the signal for specific capabilities. By extracting a low-rank capability subspace from model gradients and projecting KV activations into this subspace, SPD achieves generalizable, capability-selective improvement without requiring external feedback.
The approach's ability to achieve up to 13% improvement over state-of-the-art self-distillation methods without external signals, and up to 16% improvement over pre-trained baselines, demonstrates the effectiveness of capability-selective training. The superior generalizability shown under out-of-domain generalization settings suggests that this approach can be applied broadly across different tasks and domains.
This work contributes to the understanding of how self-distillation can be made more effective and generalizable by focusing on specific capabilities rather than trying to improve all aspects simultaneously. The capability-selective nature of the approach provides a principled way to improve specific aspects of model performance while maintaining overall utility.
Key insight: SPD achieves capability-selective self-distillation without external signals by extracting low-rank capability subspaces from model gradients and projecting KV activations into this subspace during self-generation, achieving up to 16% improvement over pre-trained baselines.
Seeing the Poem: Image-Semantic Detection of AI-Generated Modern Chinese Poetry with MLLMs
arXiv: 2605.22654
This work demonstrates the effectiveness of multimodal approaches for AI-generated content detection, particularly in the context of Chinese poetry where traditional text-only methods may be insufficient. The integration of image information that reflects the content of poetry provides complementary judgment that significantly improves detection accuracy over baseline methods.
The approach of using example-driven methods to incorporate meaning, imagery, and feeling from images into the detection process shows how visual information can enhance textual analysis for content detection tasks. The demonstration that the Gemini detector using this method achieves state-of-the-art performance indicates that multimodal approaches can be particularly effective for complex content types like poetry.
The finding that different LLM detectors show varying performance improvements on multiple LLM-generated data proves the effectiveness of the image-semantic guidance approach. This work suggests that multimodal detection methods may be particularly valuable for applications where content has rich semantic and contextual elements that are difficult to capture through text alone.
Key insight: Image-semantic guided poetry detection using LLMs achieves state-of-the-art performance (85.65% Macro-F1) by integrating visual content with textual analysis, demonstrating that multimodal approaches can significantly improve AI-generated content detection in Chinese poetry.
Boiling the Frog: A Multi-Turn Benchmark for Agentic Safety
arXiv: 2605.22643
The Boiling the Frog benchmark reveals a critical vulnerability in current AI safety evaluation - that tool-using agents can be gradually manipulated through incremental attacks that start with benign workspace edits and escalate to risk-bearing requests. The aggregate attack success rate of 44.4% across nine models demonstrates that this is a widespread problem that affects even state-of-the-art systems.
The three-level operational risk taxonomy grounded in AI Act frameworks provides a structured approach to understanding and categorizing the types of risks that agents face in real-world deployments. The focus on stateful multi-turn evaluation with persistent workspaces and controlled risk-bearing payloads creates a realistic testbed for evaluating agent safety under conditions that mirror actual usage scenarios.
The finding that model-level attack success rates range from 20.5% to 92.9% shows that some models are significantly more vulnerable than others, suggesting that safety measures should be tailored to specific model capabilities. The demonstration that Code of Practice loss-of-control scenarios achieve 93.3% average chain category-level attack success rates highlights the particular risks associated with these types of scenarios.
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 reveals that the relationship between context, knowledge, and model performance in value detection is more nuanced than previously assumed. The finding that more context is not uniformly better - full-document context improves supervised models but does not consistently help zero-shot LLMs - suggests that different approaches may be needed for different model types.
The demonstration that retrieved moral knowledge is more consistently useful in matched comparisons, improving each tested model family and context condition under early fusion, shows that knowledge integration can be more effective than simply increasing context length. The observation that simple early fusion outperforms late-fusion and cross-attention RAG variants for encoders indicates that the timing and method of knowledge integration matter significantly.
The per-value analyses showing that context and retrieval help most for socially situated or conceptually confusable values provide practical guidance for when to apply different approaches. The conclusion that value-sensitive NLP should evaluate context, knowledge, and model family jointly rather than treating longer inputs or larger models as universal improvements represents a significant shift in how these systems should be designed and evaluated.
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 models while retrieval-augmented settings with moral knowledge are more consistently useful, suggesting that joint evaluation of context, knowledge, and model family is necessary.
Vector Policy Optimization: Training for Diversity Improves Test-Time Search
arXiv: 2605.22817
VPO addresses a fundamental limitation in standard LLM post-training where optimizing for a pre-specified scalar reward leads to low-entropy response distributions that struggle with diverse downstream tasks. By explicitly training policies to produce diverse solutions that specialize to different trade-offs in the vector reward space, VPO enables more effective test-time search procedures.
The approach's effectiveness across four tasks, where VPO matches or beats scalar RL baselines on test-time search metrics like pass@k and best@k, with the gap widening as search budget grows, demonstrates that diversity training becomes increasingly valuable as search complexity increases. The finding that VPO models unlock problems that GRPO models cannot solve at all shows the practical advantages of this approach.
This work has significant implications for how LLMs should be post-trained for deployment in environments where diverse solutions are required. The demonstration that optimizing for diversity may need to become the default post-training objective as test-time search becomes more standardized suggests a fundamental shift in training paradigms for LLMs in practical applications.
Key insight: Vector Policy Optimization (VPO) trains policies to anticipate diverse downstream reward functions by explicitly optimizing for diverse solutions, achieving better performance than scalar RL baselines in test-time search scenarios with increasing search budgets.
Remember to be Curious: Episodic Context and Persistent Worlds for 3D Exploration
arXiv: 2605.22814
This work identifies a critical failure in existing curiosity-driven reinforcement learning approaches - the lack of spatial persistence and episodic context - which leads to agents becoming trapped in local loops and receiving fresh rewards for revisiting forgotten states. The solution of combining persistent world modeling with episodic trajectory history maintenance addresses these fundamental limitations.
The design using online 3D reconstruction as a persistent model of the world, paired with sequence models over RGB observations to maintain episodic context, enables effective exploration during training while allowing deployment using only RGB frames. This approach demonstrates that effective curiosity requires both long-term memory of the environment and short-term memory of exploration history.
The demonstration that the agent outperforms RL-based active mapping baselines and generalizes zero-shot to different environments shows the practical effectiveness of this approach. The ability to adapt to downstream tasks like apple picking and image-goal navigation without retraining highlights the generalizability of the learned exploration capabilities.
Key insight: Effective curiosity in 3D exploration requires both persistent world modeling through online 3D reconstruction and episodic trajectory history maintenance, enabling agents to navigate novel regions while maintaining spatial persistence and avoiding local loops.
The Matching Principle: A Geometric Theory of Loss Functions for Nuisance-Robust Representation Learning
arXiv: 2605.22800
The Matching Principle offers a novel theoretical framework that unifies various robustness approaches by identifying the common structure underlying problems like robustness, domain adaptation, and photometric invariance. By estimating the covariance of label-preserving deployment nuisance and regularizing the encoder Jacobian along that covariance's range, the approach provides a principled method for robust representation learning.
The closed-form optimality theorem and related mathematical results provide strong theoretical foundations for the approach, including cube-root water-filling within the matched range and necessity of range coverage for quadratic Jacobian penalties. These theoretical insights help explain why different existing methods (CORAL, adversarial training, IRM, etc.) are essentially different estimators of the same underlying object rather than independent approaches.
The empirical validation across classical ML through Qwen2.5-7B demonstrates that the predicted matched, then isotropic, then wrong-W ordering on geometry and deployment drift produces consistent results, with the sole exception being an eigengap failure that was identified before the run. This suggests that the theoretical framework provides a reliable guide for practical implementation and can help identify when different approaches are likely to be effective.
Key insight: The Matching Principle provides a unified geometric theory for loss functions that regularizes encoder Jacobian along the range of label-preserving deployment nuisance, achieving closed-form optimality and improved robustness across multiple domains.
SDPM: Survival Diffusion Probabilistic Model for Continuous-Time Survival Analysis
arXiv: 2605.22776
SDPM introduces a novel approach to survival analysis by modeling the conditional distribution of survival outcomes using denoising diffusion models, which avoids the structural assumptions on hazard functions and discretization that limit existing methods. This generative approach allows for more flexible modeling of time-to-event distributions without imposing restrictive parametric forms.
The use of standardized log-times and continuous Gaussian-mixture representation of censoring indicators in the transformed target space enables the model to handle censored observations effectively while maintaining the ability to generate samples that can be transformed into survival function estimates using the Kaplan-Meier estimator. This formulation provides a clean bridge between generative modeling and traditional survival analysis methods.
The competitive performance across multiple real survival datasets, including improvements in C-index, integrated time-dependent AUC, and integrated Brier score, demonstrates that SDPM can achieve results comparable to strong baselines while offering advantages in flexibility and avoiding discretization errors. The ablation study confirming the importance of target-space transformations shows that these design choices are crucial for achieving good performance.
Key insight: SDPM provides a generative approach to continuous-time survival analysis using denoising diffusion models, avoiding parametric assumptions and discretization while achieving competitive predictive performance across multiple survival datasets.
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: handling frequent data missingness from blinks and tracking failures, and efficiently modeling long-range temporal dependencies. The XMD encoding approach that augments raw features with observation masks and time-deltas explicitly models data uncertainty, while the bidirectional Mamba-2 architecture captures temporal dependencies with linear computational complexity.
The demonstration that MambaGaze achieves 76.8% and 73.1% accuracy on CLARE and CL-Drive datasets, outperforming CNN, Transformer, ResNet, and VGG baselines by 4-12 percentage points, shows the practical effectiveness of this approach. The real-time inference capabilities at 43-68 FPS with power consumption below 7.5W confirm the feasibility for wearable cognitive load monitoring applications.
The edge deployment benchmarks on NVIDIA Jetson platforms demonstrate that the approach is practical for real-world applications, particularly in safety-critical domains like driver vigilance monitoring or automated flight deck assistance. The combination of high accuracy and efficient real-time performance makes MambaGaze a promising solution for wearable cognitive load monitoring systems.
Key insight: MambaGaze addresses challenges in cognitive load assessment from eye-gaze tracking data through XMD encoding that explicitly models data uncertainty and bidirectional Mamba-2 that captures long-range temporal dependencies with linear computational complexity.
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 by introducing LeadBridge, a learnable adapter that transforms 3-lead wearable signals into anatomically consistent 12-lead representations, and ProFine, a progressive fine-tuning strategy that prevents catastrophic forgetting while gradually unfreezing encoder layers.
The evaluation on CLARE and CL-Drive datasets under leave-one-subject-out cross-validation shows that CogAdapt substantially outperforms baselines trained from scratch, achieving macro-F1 scores of 0.626 and 0.768. This demonstrates that foundation model adaptation can effectively bridge the gap between clinical and wearable sensor configurations.
The approach's ability to achieve superior performance with subject-independent cognitive load assessment from wearable sensors shows the potential for foundation models to be adapted for diverse applications. The demonstration that CogAdapt can leverage rich representations from clinical ECG foundation models while adapting to wearable constraints represents a significant advancement in wearable health monitoring.
Key insight: CogAdapt enables transfer of clinical ECG foundation models to wearable cognitive load assessment through LeadBridge adapter that transforms 3-lead wearable signals into anatomically consistent 12-lead representations and ProFine progressive fine-tuning strategy.
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 where the standard plug-in bridge parameterization is optimized by a leave-one-out posterior rather than the denoising posterior, which has important implications for training and inference. The identification of this mismatch provides insights into why uniform diffusion models may not perform as well as expected with standard parameterizations.
The derivation of exact conversions between the denoiser, leave-one-out posterior, and score enables disentanglement of parameterization and training objective, leading to improved inference through informed predictor-corrector samplers and temperature sampling. These improvements demonstrate that better understanding of the underlying posteriors can lead to practical gains in model performance.
The absorbing-state reformulation of uniform diffusion that preserves the joint law while decomposing it into masked-diffusion-like sampling operations provides a new perspective on uniform diffusion. The simpler denoising posteriors, carry-over unmasking, and natural remasking mechanism suggest that this approach can be more practical than traditional uniform diffusion while maintaining the benefits of the original formulation.
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 fundamental challenge in differentially private random forests where existing approaches typically degrade performance to impractical levels. By constructing large random decision trees and then applying aggressive, privacy-preserving pruning to retain only sufficiently populated nodes, the approach achieves substantially better utility than prior methods.
The novel $(\varepsilon,\delta)$-DP heavy hitter detection algorithm for hierarchical data with error scaling of $O_{\varepsilon,\delta}(\sqrt{\log h})$ for trees of height $h$ is a key innovation that enables the use of significantly deeper trees than in prior work. This favorable scaling allows for improved expressiveness under privacy constraints, closing much of the utility gap.
The empirical evaluation on benchmark datasets shows that Lumberjack consistently outperforms prior DP random forest methods, establishing a new state of the art. The substantial improvements in privacy-utility trade-off for practical privacy budgets suggest that carefully designed DP random forests can be practical for real-world applications involving sensitive tabular data.
Key insight: Lumberjack achieves substantially higher utility in differentially private random forests through aggressive pruning of large trees and a novel heavy hitter detection algorithm that enables deeper trees with improved expressiveness under privacy constraints.
Cyber-Physical Anomaly Detection in IoT-Enabled Smart Grids Using Machine Learning and Metaheuristic Feature Optimization
arXiv: 2605.22749
This work demonstrates how metaheuristic feature optimization can significantly improve cyber-physical anomaly detection in smart grids by reducing redundant measurements while maintaining or improving detection performance. The genetic algorithm-based feature selection reduces the clean PMU feature space from 112 attributes to an average of 27.4 attributes over five runs.
The improvement in macro-F1 from 0.9118 to 0.9212 and ROC-AUC from 0.9791 to 0.9837 shows that the reduced feature set not only maintains but actually improves detection performance, indicating that many synchronized electrical measurements are indeed redundant. This finding has important implications for smart grid design and deployment.
The demonstration that a compact subset of phasor-based features can still provide accurate and interpretable anomaly detection in smart grids suggests that the approach can be applied to other domains where sensor data reduction is needed. The ability to identify redundant measurements while maintaining high detection accuracy makes this approach valuable for practical 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 improving macro-F1 from 0.9118 to 0.9212 and ROC-AUC from 0.9791 to 0.9837.
Plug-in Losses for Evidential Deep Learning: A Simplified Framework for Uncertainty Estimation that Includes the Softmax Classifier
arXiv: 2605.22746
This work addresses the computational challenges of evidential deep learning by approximating the objective of first-order empirical risk minimization with a plug-in loss evaluated at the Dirichlet mean, providing a simplified approach that maintains the benefits of EDL while being more practical to implement.
The theoretical analysis showing that the approximation error decays with growing evidence for a broad class of loss functions, including mean-squared error and cross-entropy loss, provides strong justification for the approach. The demonstration that under a particular evidence-to-Dirichlet mapping, the framework includes the standard softmax classifier shows that the approach can be seen as a generalization of existing methods.
The empirical validation on the Google Speech Commands dataset shows that the 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 makes evidential deep learning more accessible for practical applications.
Key insight: Plug-in losses for evidential deep learning provide a simplified framework that achieves comparable performance to classical EDL while being simpler to implement using standard deep learning losses and training pipelines, including softmax as a special case.
Self-Evolving Multi-Agent Systems via Decentralized Memory
arXiv: 2605.22721
DecentMem addresses key limitations of centralized memory in multi-agent systems by introducing 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. This approach eliminates communication overhead, privacy concerns, and agent diversity collapse associated with centralized systems.
The theoretical guarantee of global reachability of the solution space and $O(\log T)$ cumulative regret, matching the stochastic bandit lower bound, provides strong foundations for the approach. The online reweighting based on stage-wise feedback from an LLM-as-a-judge ensures that agents can adapt their memory strategies dynamically based on performance feedback.
The practical demonstration across three MAS frameworks, three Qwen3 backbones, and five benchmarks 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 significant performance improvement demonstrates the practical value of decentralized memory for multi-agent systems.
Key insight: DecentMem 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.
Sibyl-AutoResearch: Autonomous Research Needs Self-Evolving Trial-and-Error Harnesses, Not Paper Generators
arXiv: 2605.22343
Sibyl-AutoResearch addresses a fundamental limitation in current autonomous research systems where weak evidence becomes prose, pilot signals become broad claims, and memory remains textual without learning from process failures. The introduction of Scientific Trial-and-Error Harnesses provides a framework for agents to run bounded trials, preserve outcomes, and route lessons into later planning and validation.
The formalization through two auditable conversion units - trial-to-behavior conversion and trial-to-harness-behavior conversion - creates a systematic approach to learning from research experience. The demonstration that eight high-confidence conversion events were identified with median latency of one iteration shows that the framework can capture and utilize learning from real research processes.
The implementation of the framework in SIBYL, a file-backed autonomous research system that exposes state, roles, memory, gates, and artifact traces, provides a practical platform for inspecting conversion paths and understanding how learning from failures can be systematically incorporated into research workflows. This approach represents a significant step toward truly autonomous research systems that can improve through experience.
Key insight: Sibyl-AutoResearch introduces Scientific Trial-and-Error Harnesses that enable autonomous research systems to learn from trial experience, preserve positive and negative outcomes, and route lessons into later planning and validation, addressing limitations of current systems that lose trial experience.
ACCoRD: Actor-Critic Conflict Resolution with Deep learning for O-RAN xApps
arXiv: 2605.22306
ACCoRD addresses the critical challenge of conflict mitigation in Open Radio Access Networks (O-RAN) by introducing a reinforcement learning-based approach that trains a Conflict Resolution Agent to analyze network data and conflicting control decisions to infer optimal CR actions. The use of PPO-Clip algorithm for training provides a robust framework for learning conflict resolution strategies.
The evaluation based on simulation data shows that the proposed ANN-based method significantly improves on the efficiency of rule-based approaches by reducing negative network events caused by conflicting control decisions in medium and high traffic scenarios. This demonstrates the practical value of reinforcement learning for complex network control problems.
The approach's ability to gather feedback from the network after each resolved conflict to assess efficiency and adjust ANN weights during batch training creates a self-improving system that can adapt to changing network conditions. This learning capability is crucial for maintaining effective conflict resolution in dynamic network environments.
Key insight: ACCoRD resolves control conflicts in O-RAN xApps using a Conflict Resolution Agent with ANN trained via PPO-Clip reinforcement learning, achieving significant reduction in negative network events caused by conflicting control decisions compared to rule-based approaches.
Emergence of agriculture in an artificial society of reinforcement learning agents
arXiv: 2605.22256
This work demonstrates how complex collective behaviors can emerge from simple interactions in artificial societies, with agricultural practices developing spontaneously without explicit instruction. The identification of four key ingredients - individual planning, social vulnerability to cheaters, stabilization via social learning, and emergent lock-in effect - provides a framework for understanding how complex cultural innovations can arise.
The finding that social learning acts as a 'firewall' that suppresses cheater invasion and enables propagation of successful strategies shows how social mechanisms can stabilize beneficial innovations. The demonstration that this leads to sustained population growth and nonlinear amplification of domesticated resources reveals how collective behaviors can become self-reinforcing.
The broader implications extend beyond agriculture to understanding how artificial societies can be used to study major evolutionary transitions and cultural innovations. The approach provides experimental platforms for studying emergence of complex behaviors through individual decision-making, social interactions, and ecological feedback, offering insights into how collective behaviors can be designed and understood.
Key insight: Artificial societies of reinforcement learning agents spontaneously develop 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 identifies a fundamental challenge in EV charging system design: the three interdependent layers of Planning, Scheduling, and Behavior cannot be solved in isolation due to computational difficulties, creating a PSB trilemma where realistic integration requires reducing the fidelity of at least one layer. This insight is crucial for understanding the complexity of integrated systems design.
The three-layer PSB framework organizes EV charging research according to decision horizon, actor objective, and coupling structure, providing a systematic approach to understanding cross-layer interactions. The identification of specific costs from simplifying assumptions - such as omitting the third layer or representing it by static surrogates - helps clarify the trade-offs involved in different approaches.
The diagnosis of the literature's fragmentation and cross-layer interaction limitations reveals important open challenges in emerging charging technologies, behavioral incentives, equity metrics, and city-scale learning-based methods. The need to balance fidelity, interpretability, and policy relevance suggests that future research should focus on developing methods that can handle complexity while remaining practical for real-world deployment.
Key insight: EV charging systems require integrated planning, scheduling, and behavior decisions across three interdependent layers, with a PSB trilemma showing that realistic integration generally requires reducing fidelity of at least one layer due to computational difficulties in isolation.
Argo: Efficient Importance Labeling for Enterprise Email Systems
arXiv: 2605.21604
Argo addresses the computational cost challenge of using large language models for enterprise email importance labeling by developing a profiler that efficiently searches the cost-quality trade-off space to identify cost-efficient alternatives to GPT-4.1 scale LLMs. This approach makes large-scale, context-aware email labeling practical for enterprises.
The on-demand provisioning scheme that intelligently scales Argo with real-time load minimizes cost increases during peak load inference, demonstrating practical deployment considerations for large-scale AI systems. The 148-167X inference cost reduction with negligible quality degradation shows the substantial efficiency gains possible through careful system design.
The 20-640000X lower profiling costs compared to traditional approaches make Argo particularly valuable for enterprise applications where cost optimization is critical. The framework's ability to achieve near GPT-level labeling quality with significantly lower cost demonstrates that sophisticated AI systems can be made practical for large-scale enterprise deployment.
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 designing 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, where single-agent approaches fail due to ignoring other actors or treating them as environmental noise. The approach enables agents to develop sophisticated anticipatory behaviors including proactive collision avoidance, overtaking, and handling multi-agent physical interactions.
The demonstration that agents outperform a champion-level human pilot in multi-player races at speeds exceeding 22 m/s while reducing collision rates by 50% compared to state-of-the-art single-agent baselines shows the practical advantages of multi-agent interaction. The ability to train with diverse artificial agents enabling zero-shot generalization to safer human interaction suggests that multi-agent systems can be designed to be more robust and safe.
The finding that the path to robust robotic co-existence lies not in isolated safety constraints but in the rigorous demands of multi-agent interaction represents a fundamental shift in how safety should be approached for autonomous systems. The multimedia materials and detailed analysis provide valuable insights into how multi-agent systems can be designed to handle complex real-world scenarios.
Key insight: Multi-agent reinforcement learning provides essential safety scaffolding for real-world interaction, enabling agents to navigate complex aerodynamic interactions and strategic maneuvering with a variable number of racers, outperforming human pilots in multi-player races.
Cross-domain benchmarks reveal when coordinated AI agents improve scientific inference from partial evidence
arXiv: 2605.22300
This work provides important insights into when coordinated AI agents add value over simpler scientific workflows by evaluating cross-domain benchmarks spanning four scientific tasks. The finding that cross-channel composites improve over single-channel baselines when different disciplines each capture only part of the phenomenon demonstrates the value of coordinated approaches for complex scientific problems.
The benchmarks show that coordination improves performance in cases where the full phenomenon is not captured by any single discipline, with climate-vector emergence reaching AUROC 0.944 and exoplanet vetting reaching AUROC 0.955. However, when one signal dominates, as in paradigm-shift detection, coordination mainly improves interpretation and traceability rather than predictive performance.
The ScienceClaw x Infinite framework provides an auditable artifact and provenance layer for this evaluation, assigning value to coordination only when corresponding performance, provenance, or representation claims are supported by explicit comparators. This approach ensures that the value of coordination is clearly demonstrated rather than assumed, providing a rigorous foundation for understanding when and how coordination improves scientific inference.
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 reaching AUROC 0.944 for climate-vector emergence and 0.955 for exoplanet vetting, but coordination mainly improves interpretation and traceability when one signal dominates.
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 language models come first, then tools, then rules, and finally logging is bolted on for observability. By making the append-only event log the source of truth, the working graph becomes a deterministic projection of that log, creating a system where behaviors react to changes in the shared graph rather than direct instruction.
The deterministic replay capability allows any run to be reproduced from its log, while cheap forking enables branching at any event without re-executing the shared prefix. This approach provides unprecedented auditability and the ability to explore alternative execution paths, which is particularly valuable for self-improving agents that need to experiment safely.
The demonstration that this substrate is unusually well suited to self-improving agents and extends the BabyAGI lineage and prior graph-memory research shows that event-sourced reactive graphs can provide a robust foundation for autonomous systems. The ability to reconstruct full causal structures from logs alone represents a significant advancement in system transparency and accountability.
Key insight: ActiveGraph inverts traditional agent frameworks by making the append-only event log the source of truth, with the working graph being a deterministic projection of that log, 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 can support real-time instructional adaptation in serious games, distinguishing between instructional intelligence (inferring learner knowledge and reasoning about pedagogically appropriate responses) and adaptivity (modifying instructional actions during interaction). The historical synthesis traces developments from early computer-assisted instruction through intelligent tutoring systems to recent AI-enabled architectures.
The discussion of how large language models, reinforcement learning, and agent-based architectures may contribute to more integrated forms of intelligence and adaptivity in serious games provides a comprehensive view of current capabilities and future directions. The recognition that these approaches can address persistent challenges like static scenario design, authoring bottlenecks, and limited learner modeling shows the potential for AI to transform serious game design.
The identification of practical and research challenges associated with AI-enabled systems, including explainability, validation, computational cost, and limited empirical evidence regarding long-term learning outcomes, highlights the need for careful consideration of implementation. The chapter's balanced approach acknowledges both the promise and the limitations of AI integration in serious games.
Key insight: AI integration in serious games enables dynamic scenario variation, contextual feedback, adaptive pacing, and learner-state modeling, but raises important questions about validity, transparency, system control, and learner trust that must be addressed for practical deployment.