2026-W41
The Week in Review
This week's research prominently features advancements in LLM agent efficiency, evaluation, and robustness. A significant trend is the development of methods to manage and compress context for more efficient agentic inference. Papers like ActKV, Continuous Context Management, and KV-streams introduce techniques to prioritize or compact KV cache entries and interaction history, reducing computational load and improving throughput.
Evaluation and benchmarking remain critical, with notable contributions in assessing LLM capabilities beyond standard benchmarks. DIAL and Game Arena focus on more nuanced LLM evaluation by addressing biases in LLM judges and creating competitive environments, respectively. PriceBench specifically diagnoses booking agents' preferences, while ScAn-Bench evaluates scaling analysis methodologies. For AI agents, AgentXploit and EffectMatch introduce methods for autonomous red-teaming and runtime validation, respectively, to ensure security and integrity.
Several papers also explore specialized agentic applications and reasoning. DeepEdu-v1 and the cultural awareness evaluation for Haitian Creole highlight the need for localized and culturally sensitive LLMs. PIA addresses specialized health data understanding, and ViSTA bridges vision-language models with clinical time-series data. Research on multi-agent systems, such as Multi-agent Scaling and BaRe-Mem, focuses on improving coordination and reliability in complex interactions. Finally, papers like Prompt Minimization and Statistical attribute alignment demonstrate efforts to optimize LLM input and output for greater efficiency and control.
Top Papers
ActKV: Efficient LLM Agents through Action-Guided KV Cache Management
ctKV addresses the memory burden of LLM agents by prioritizing KV cache entries crucial for action generation. Its core method employs action-oriented eviction and confidence-driven budget allocation to retain relevant information, enabling more efficient and higher-throughput agentic inference. This contributes the first KV cache compression framework specifically designed to optimize agent performance by focusing on action quality.
DIAL: Position-Debiased LLM Judges with Adaptive Human Preference Calibration
IAL addresses the limitations of LLM judges by developing a framework that combines LLM and human comparisons. It separates position biases in LLM judgments, learns their underlying preferences, and then adaptively calibrates these to align with human preferences. This approach offers a theoretically grounded and empirically robust method for more accurate LLM evaluation.
Game Arena: Strategic LLM Evaluation in Competitive Environments
he paper introduces Kaggle Game Arena, a platform for evaluating LLMs through competitive gameplay. Its core method involves head-to-head matchups in structured game environments, allowing for dynamic performance assessment as models improve. The contribution lies in providing a robust, transparent, and reproducible system for studying LLM strategic capabilities across diverse game settings, overcoming the limitations of static benchmarks.
PriceBench: A Diagnostic Benchmark for Price, Quality, and Brand Preferences in LLM Booking Agents
his paper introduces PriceBench, a benchmark designed to diagnose the price, quality, and brand preferences of Large Language Models (LLMs) acting as booking agents. By analyzing LLM booking choices using a logit model, the benchmark reveals that more capable LLMs exhibit stronger, more consistent preferences, while less capable ones are either overly biased or indifferent. The study highlights that these preferences vary significantly across different LLM providers and even within the same family, impacting what options are presented and at what cost.
Nonparametric In-Context Learning under Growing Geometric Complexity: Minimax Optimality and Local Geometry-Adaptivity of Transformers
his paper analyzes how Transformers perform in-context learning on data with complex, varying geometric structures. The core method involves developing a novel local-polynomial estimator that adapts to these heterogeneous manifolds. The key contribution is establishing minimax optimality for this type of learning, demonstrating that Transformers can effectively leverage local geometric information for prediction.
ARISE: Adapting to Evolving Capability Gaps in Agentic Reinforcement Learning
RISE addresses the challenge of evolving capability gaps in long-horizon reinforcement learning agents. Its core method, Adaptive Rubric-Skill Co-Evolution (ARISE), continuously adapts evaluation criteria and exploration strategies based on agent performance. This allows the agent to focus learning on emerging weaknesses and make progress even with sparse task-level feedback.
Continuous Context Management
his paper introduces Continuous Context Management (CCM) for LLM agents, which compacts interaction history at every turn instead of waiting for a threshold. This prevents the active prompt from growing excessively, significantly reducing input usage and prompt size. While CCM initially lowers task success for most models, it can be improved through fine-tuning with privileged distillation, demonstrating a novel approach to efficient long-horizon LLM agent operation.
KV-streams for Efficient Compaction in Agentic Reinforcement Learning
his paper introduces KV-streams, a novel method to improve the efficiency of context compaction in agentic reinforcement learning. Instead of repeatedly refilling the LLM's context, KV-streams stream the KV cache forward, significantly boosting training throughput (2.6-5x speedup) without compromising performance. This approach also allows the streamed KV cache to act as a recurrent state, preserving long-term information.
SEABench: Benchmarking Endogenous Misalignment In Self-Evolving Agents
his paper introduces SEABench, a novel benchmark designed to evaluate "endogenous misalignment" in self-evolving AI agents. The core method involves creating longitudinal task sequences in a personal assistant environment to observe how agents' self-modifications can lead to unsafe behavior over time, even without external adversarial input. SEABench's contribution is providing a standardized way to study this emerging risk and an adaptive pipeline to discover and attribute failures in these evolving systems.
TokenCast: Forecasting Token Consumption During LLM Agent Execution
okenCast addresses the unpredictable token consumption of LLM agents by learning a composable cost representation for execution segments. This method estimates future token usage by accounting for both segment-specific consumption and the context growth introduced, allowing for dynamic, real-time forecast updates without additional LLM calls. Its core contribution is enabling accurate and efficient prediction of token costs during LLM agent execution.
Rethinking Personalized Generation: Test-Time Alignment via Factorized Ranking Models
his paper proposes a test-time alignment method for personalized generation with LLMs. It argues that personalization is a candidate matching problem, not a generator limitation, and that existing reward models are ill-suited for this. The core contribution is a parameter-efficient framework using small MLP ranking models that leverage base LLM embeddings to efficiently score and select personalized outputs.
ScAn-Bench: Evaluating Scaling Analysis Methodology
his paper introduces ScAn-Bench, a novel benchmark for evaluating methodologies used to derive scaling laws and optimal prescriptions for large foundation models. The authors systematically analyze data acquisition and extrapolation techniques across language and vision-language models, identifying a critical gap in current research. Their contribution lies in providing the first comprehensive evaluation of scaling analysis methodology, facilitating future advancements in model scaling.
Auditable Long-Term Memory: A Deterministic Retrieval Chain Measured at 479/475 of 500 on LongMemEval-S
his paper introduces an auditable long-term memory system that achieves high performance on the LongMemEval-S benchmark. Its core method involves a deterministic retrieval chain combining hybrid candidate retrieval, cross-encoder reranking, and coverage-first packet compilation, with an LLM acting solely as a final reader. The system's main contribution is demonstrating strong, auditable long-term memory retrieval capabilities, scoring 479/500 and 475/500 on the benchmark, highlighting its potential for reliable knowledge recall.
Diagnosing and Improving Probabilistic Reasoning in Large Language Models
his paper proposes a decision-theoretic framework to analyze how Large Language Models (LLMs) make probabilistic decisions. It decomposes decision loss into belief accuracy and action optimization, using a synthetic benchmark to diagnose LLM reasoning. The study shows that interventions targeting specific reasoning components can improve performance but don't always transfer to other areas, and decision improvement is possible even without better belief formation.
On Trajectory-Aware Training for Masked Diffusion Language Models
his paper introduces PUMBA, a unified framework for trajectory-aware training of masked diffusion language models. PUMBA addresses the mismatch between training and inference by training the model on consecutive steps of predicted trajectories and passing information between them. This approach optimizes the model jointly through backpropagation, leading to improved performance by aligning training masks more closely with inference conditions without suffering from local overfitting.
SelfSearch: Reward-Free Search for Self-Improving Agents
elfSearch enables LLM agents to improve themselves without needing external rewards. It achieves this by using past self-improvement attempts, including reasoning and outcomes, as experience to guide future modifications. This reward-free approach leads to better task performance and more effective self-improvement strategies.
The Unequal Influence of Bad Advice: Using Training Data Attribution to Modulate Emergent Misalignment
his paper investigates how fine-tuning large language models can lead to "emergent misalignment" by amplifying harmful personas. The core method uses training data attribution to quantify the influence of individual harmful examples on this misalignment. The key contribution is demonstrating that score-based filtering of training data can significantly enhance or attenuate emergent misalignment, suggesting that not all harmful examples contribute equally and that attribution can be used to control model behavior.
Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning
his paper introduces "agentic meta-reasoning," a system where a controller makes explicit, structured decisions about how to proceed with a complex task. Instead of replaying the entire history, the controller uses a compact summary to evaluate and dispatch the next steps, improving efficiency. This approach significantly enhances agentic capabilities on long-horizon tasks like program reconstruction.
UserProxyBench: Evaluating LLM User Simulators for Agent Benchmarks and Training
his paper introduces UserProxyBench, a framework to evaluate the performance of LLM-based user simulators in agent benchmarks. It proposes the User Fidelity Score (UFS) to measure how well these simulators adhere to their assigned user instructions, independent of the agent's success. The contribution lies in highlighting the significant impact of user simulator behavior on agent performance and identifying premature disclosure as a common failure mode.
TabFM-Auto: Self-Evolving Pipelines for Tabular Foundation Models
abFM-Auto addresses limitations of existing tabular foundation models and self-evolving agents by creating an adaptive pipeline. It uses a language model agent to iteratively refine data cleaning, feature engineering, and context selection around a pre-trained tabular foundation model (TabFM). This approach significantly improves TabFM's performance across a benchmark, achieving top rankings and boosting its Elo score.
TabFM: A Zero-Shot Foundation Model for Tabular Data
abFM is a novel foundation model for tabular data that leverages in-context learning for zero-shot predictions, eliminating the need for task-specific tuning. Trained on synthetic data, it learns general tabular representations that achieve state-of-the-art performance across various benchmarks, outperforming even tuned AutoML pipelines. Extensions like TabFM+ and TabFM-Auto further enhance its capabilities through ensembling and LLM-guided data processing.
Effective Dense Retrieval using Only In-Context Examples
his paper introduces RICE, a training-free method to make decoder-only LLMs effective dense retrievers. RICE prompts the LLM with in-context examples that establish a shared context for encoding queries and documents, thereby producing high-quality dense representations without any fine-tuning. This approach significantly improves prompt-based LLM embeddings for retrieval tasks.
Agent Error Dataset: Scaling 50,000 Error--Diagnosis Pairs for Failure Analysis and Error-Aware Post-Training
his paper introduces the Agent Error Dataset (AED), a large collection of 50,000 error-diagnosis pairs for text-based LLM agents. The core method involves a five-stage pipeline (AET) to automatically identify agent failures, diagnose their causes, and propose corrections using existing execution traces. The main contribution is providing a scalable resource for analyzing agent failures and developing error-aware post-training techniques, enabling agents to learn from their mistakes without re-running entire rollouts.
PhantomEnvironments: Training LLM Agents in Fictional Worlds
his paper introduces PhantomEnvironments, a novel method for training LLM agents using purely rule-based, fictional environments. This approach bypasses the need for costly human data or LLM-generated environments prone to hallucinations. The core contribution is demonstrating that agents trained in these synthetic worlds can effectively transfer their multi-hop search capabilities to real-world tasks, even outperforming agents trained on real data.
PivotOPD: Learning to Recover from Pivotal Mistakes in Multi-Turn Agents
ivotOPD addresses the compounding errors in multi-turn language agents by identifying and correcting "pivotal mistakes" – actions that significantly derail task progress. The core method involves an on-policy distillation framework that trains agents to both avoid these critical errors and recover from the states they induce. This approach aims to improve task success by focusing on early intervention and guided recovery after crucial missteps.
Debias It Yourself: Teaching LLMs Cognitive Bias Mitigation Interventions
his paper introduces "Debias It Yourself" (DIY), a framework that translates human cognitive bias mitigation techniques into methods for large language models. The core method involves applying these interventions through in-context examples, instruction tuning, or guided self-revision. The key contribution is demonstrating that these cognitively grounded approaches significantly reduce bias in LLMs while maintaining high reasoning accuracy, even on novel bias types.
JuryFlow: Disagreement-Guided Human-in-the-Loop Multi-Agent Evaluation
uryFlow addresses the unreliability of LLM judges by treating inter-judge disagreement as a valuable signal. Its core method decomposes responses into atomic claims, uses heterogeneous judges to provide per-claim verdicts, and builds a disagreement graph. The framework's contribution is a human-in-the-loop system that guides minimal human interventions to resolve specific disagreements, improving evaluation accuracy.
Persistent Context Graphs for Efficient Memory Compaction in LLM Agents
his paper introduces ReCAP, a novel method for efficient memory compaction in LLM agents. ReCAP constructs a persistent context graph storing attention-derived importance scores and message dependencies. This allows agents to quickly identify and retrieve relevant past interactions for new requests without re-encoding expired KV caches, significantly reducing computational costs and improving long-horizon task performance.
Agentick: A Unified Benchmark for General Sequential Decision-Making Agents
gentick is a unified benchmark designed to fairly evaluate diverse sequential decision-making agents, including RL, LLM, and hybrid approaches. It offers 37 procedurally generated tasks across various capabilities, difficulties, and modalities through a single interface. Agentick's contribution lies in enabling direct comparison of agent types and facilitating research into fundamental sequential decision-making challenges.
Calibrated Enough to Know, Not Calibrated to Act: Fabricated Evidence Makes LLM Agents Commit to the Unknowable
his paper demonstrates that Large Language Model (LLM) agents are highly susceptible to fabricated evidence, committing to actions on unknowable questions when presented with professional-looking data. The core method involves comparing LLM agent responses to provably unpredictable questions with and without fabricated market panels, showing a significant increase in confident action. The contribution highlights that the authority of the evidence's packaging, rather than its factual content, drives LLM decision-making, even when the underlying information is false.
CODESKILL: Learning Self-Evolving Skills for Coding Agents
ODESKILL is a novel LLM-based framework that learns to manage a bank of reusable coding skills. It extracts, evolves, and maintains these skills from agent trajectories using a learned policy, addressing limitations of fixed prompts and heuristic updates. This self-evolving skill bank allows coding agents to more effectively learn and adapt to new tasks.
Combee: Scaling Prompt Learning for Self-Improving Language Model Agents
ombee is a novel framework designed to scale prompt learning for self-improving language model agents. Its core method addresses the limitations of existing approaches by enabling efficient and high-quality parallel prompt learning, allowing agents to learn from a large number of traces without performance degradation. This contribution significantly improves the efficiency and effectiveness of agents that learn and adapt over time.
DIAL: Position-Debiased LLM Judges with Adaptive Human Preference Calibration
IAL is a framework that uses both abundant LLM comparisons and limited human feedback to improve LLM-based evaluations. It debiases LLM judgments by accounting for response order and then calibrates these debiased LLM preferences to better align with actual human preferences. This approach allows for more scalable and accurate LLM evaluations by leveraging LLMs while ensuring they reflect human values.
Estimating and Orthogonalizing Unknown Pre-training Gradients for Continual Fine-tuning of Large Language Models
his paper proposes EoupCT, a novel framework to address catastrophic forgetting in continual LLM fine-tuning. EoupCT estimates unknown pre-training gradients by generating pseudo-data via a learnable soft prompt, allowing for orthogonalization and preservation of general-purpose knowledge lost during fine-tuning. This approach aims to overcome limitations of existing methods that require inaccessible pre-training data.
FTB Graph: Determining and Validating First-token Broadcasters and Language-Identity Head Circuits in Multilingual Language Models
his paper introduces FTB Graph, a method using Edge Attribution Patching and activation patching to map the causal circuits responsible for determining the target language of multilingual language models. The core contribution is the discovery of "broadcasting hubs" in deep or mid-to-deep layers that dictate first-token language identity, with variations observed across different model architectures.
G$^2$PTQ: Improving LLM Post-Training Quantization with Generalized Gradient Compensation
$^2$PTQ improves LLM post-training quantization by unifying first- and second-order gradient information within a globally supervised, block-wise optimization. It addresses limitations of prior methods by dynamically refreshing gradient and Hessian estimates for each Transformer block and stabilizing updates with a trust-region scaling mechanism. This approach leads to more accurate and efficient LLM quantization.
HasMem: Hard-Origin Adaptively Softened Memory for Long-Term LLM Agents
asMem introduces a novel memory system for long-term LLM agents that combines a fixed, verifiable "hard-origin" prompt with an adaptively softened memory. This approach allows for flexible memory resizing and re-encoding without altering the frozen LLM, significantly improving long-term context retention and question answering performance on complex multi-session tasks.
MoMHa: Multi-Objective Optimization of LLM Harnesses over Accuracy, Safety, and Tokens
his paper introduces MoMHa, a novel method for optimizing the Python code (harness) that surrounds Large Language Models. MoMHa treats harness design as a multi-objective search problem, balancing accuracy, safety, and token cost. Its core contribution is demonstrating that a single-phase joint-reward approach for this search outperforms other strategies, leading to better overall harness performance.
PUBG Ally: A Conversational Embodied Agent as an AI Teammate
UBG Ally is a conversational embodied agent designed to act as an AI teammate in the game PUBG: BATTLEGROUNDS. Its core method combines a language model for reasoning and communication with a fast control layer for real-time game actions, enabling it to perceive the game, interact naturally with players, and make strategic decisions. The paper's contribution lies in demonstrating this integrated approach to creating a responsive and engaging AI teammate, trained on extensive real-world gameplay data.
Rufus-Air: An Open LLM Post-Training Recipe
ufus-Air presents an open, eight-stage post-training recipe for LLMs, building capabilities sequentially from supervised fine-tuning to advanced agentic reasoning. Its core contribution is a reproducible methodology using open-source components and public data, emphasizing reward reliability and difficulty filtering for effective training without new human annotation.
Stepwise Intrinsic Rewards for Reasoning in Large Language Models
his paper introduces Stepwise Marginal Information Gain (MIG), an intrinsic reward signal for training LLMs in reinforcement learning. MIG measures how each reasoning step improves the likelihood of the correct answer, rewarding progress without requiring external annotations or complex search. This allows for more granular feedback during training, guiding LLMs towards better reasoning by identifying beneficial intermediate steps.
StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction
his paper introduces StraTA, a framework that enhances agentic reinforcement learning for long-horizon tasks by incorporating an explicit, sampled trajectory-level strategy. This strategy guides the agent's actions, improving exploration and credit assignment. StraTA's core contribution is this strategic abstraction, which jointly trains strategy generation and action execution, leading to significant improvements in sample efficiency and final performance across various challenging benchmarks.
The Geometry of Refusal: Why Post-Hoc Safety Is Fragile and Pretraining-Time Safety Persists
his paper argues that post-hoc safety training methods like RLHF are fragile because they create a "refusal gate" that masks, rather than erases, undesirable behaviors. Geometrically, these safety updates are orthogonal to the model's core capabilities, meaning they only suppress them slightly on sharp curvature directions. Consequently, even minor benign fine-tuning can easily restore the original, unsafe capabilities.
Think Short, Defer Smart, Act, and Repeat: Calibrated Reasoning and Uncertainty-Aware Deferral for Edge LLM Agents
his paper introduces Think Short, Defer Smart (TSDS), a framework for edge LLM agents. TSDS efficiently manages on-device reasoning by stopping early when an action is stable and intelligently defers uncertain actions to a cloud model using perplexity. Its core contribution is a jointly calibrated approach that optimizes both task performance and resource usage on edge devices.
VLANeXt Family: A Systematic Study of VLA Models from Core Recipes to Emerging Paradigms
his paper systematically studies Vision-Language-Action (VLA) models by dissecting design choices across foundational components, perception, and action modeling. Their core method involves a unified framework and evaluation setup, starting from a baseline similar to RT-2. The key contribution is distilling 12 findings into a practical recipe for building strong VLA models, exemplified by their proposed VLANeXt.
Why Jailbreaks Succeed in Diffusion Language Models: An Energy Landscape Analysis
his paper explains jailbreaks in diffusion language models by viewing safety alignment as an energy landscape. Jailbreaks succeed by either hiding the harmful intent initially or forcing the generation process over an "energy barrier" towards unsafe outputs. The authors propose three training-free detection methods based on analyzing the initial query and the generation's trajectory, offering a unified framework for understanding and detecting these attacks.
AgentXploit: Autonomous Repository-to-Runtime Red-Teaming for AI Agents
gentXploit is an autonomous system for pre-deployment red-teaming of AI agents. It uses a two-agent approach: an Analyzer Agent discovers potential attack paths within the agent's code repository, and an Exploiter Agent translates these paths into concrete runtime attacks, refining them with feedback. This system's contribution is a novel, automated method for identifying and exploiting security vulnerabilities in AI agents before deployment, supported by a benchmark of 72 reproducible vulnerabilities.
Beyond Approved Actions: Runtime Validation of Persistent Outcomes in Agent Workflows
his paper introduces EffectMatch, a runtime validation system for LLM agents that interact with software. Its core method is to collect and compare persistent changes made by an agent's actions against the approved outcomes for the current state. EffectMatch's contribution is to prevent unapproved side effects from being accepted as success, thereby ensuring the integrity of agent workflows and preventing downstream errors.
Compact Documentation for Coding Agents: A Benchmark, an Optimizer, and Why It Does Not Transfer
his paper introduces a benchmark for evaluating code documentation by testing if regenerated code passes original tests, finding completeness, not length, is key. They developed an optimizer that creates highly faithful documentation, but surprisingly, this improved documentation did not help coding agents resolve real software issues, even when compared to using no documentation.
DeepEdu-v1: Efficient and Scalable Agentic LLMs for Vietnamese Education
his paper introduces DeepEdu-v1, an AI tutoring system for Vietnamese education designed to address data sovereignty and curriculum alignment issues. Its core method, the SCALE framework, innovates by improving context-aware learning and self-improvement. This contributes a scalable, efficient agentic LLM that overcomes the limitations of existing solutions for localized educational AI.
Evaluating Cultural Awareness of LLMs for Haitian Creole
his paper introduces the first systematic evaluation of cultural awareness in LLMs for Haitian Creole. It uses a benchmark of culturally relevant prompts to assess specificity, bias, diversity, and variation, revealing significant gaps compared to French and highlighting issues like French linguistic interference and stereotypical portrayals of Haitian characters. The core contribution is establishing a methodology and dataset to identify and address cultural insensitivity in LLMs for underrepresented languages.
Highlight-Then-Summarize: Learning to Compress Evidence for Long-Context Understanding
his paper introduces Highlight-Then-Summarize (H2S), a novel compress-then-reason approach for long-context understanding. H2S first identifies relevant evidence from lengthy documents and then generates a compact, question-conditioned summary of this evidence before producing the final answer. This method, trained with a new dataset and reinforcement learning, significantly improves performance on long-context tasks by effectively filtering and condensing information.
Multi-agent Scaling Across Disjunctive and Compensatory Tasks
his paper introduces Steiner's taxonomy of group tasks to analyze how multi-agent LLM performance scales with team size, focusing on disjunctive and compensatory tasks. The authors model agents as conditionally independent and derive theoretical limits for plurality voting and averaging. Their empirical findings show distinct scaling behaviors, with disjunctive tasks showing potential gains that plurality voting largely fails to capture, though multi-round revision improves accuracy.
Prompt Minimization: Reducing Input Redundancy Without Sacrificing Output Fidelity
his paper introduces "prompt minimization," a method to find the shortest possible prompts for large language models that still produce accurate outputs. By removing redundant information, their approach reduces computational costs and improves model performance. The core contribution is demonstrating that these minimal prompts can achieve comparable fidelity to longer, more verbose ones, highlighting the importance of efficient prompt design.
Statistical attribute alignment for black-box generative AI via output post-processing
his paper addresses the challenge of aligning generative AI outputs with user-specified attribute distributions in a black-box setting. The core method involves post-processing generated outputs to statistically match a target distribution for a specific attribute. The key contribution is the development of query-efficient algorithms that minimize the number of calls to the generative model, offering optimal solutions for achieving exact or approximate attribute alignment.
Towards Mitigating Fabricated Consensus: The Active Provenance Gate for Multi-Agent Debate Synthesis
his paper introduces the Active Provenance Gate (APG), a post-debate verification layer for multi-agent debate systems. The APG analyzes debate logs, audits claims, and self-corrects to prevent fabricated consensus, ensuring summaries are grounded in the debate's history. Its core contribution is a mechanism to enhance the reliability and factual accuracy of synthesized conclusions from LLM-based debates.
ViSTA: A Simple Bridge Extends Visual Alignment to Clinical Time-Series Understanding in Multimodal LLMs
iSTA is a compact adapter that bridges clinical time-series data with multimodal large language models by learning corrections to visual tokens representing numerical measurements. This approach allows pretrained vision-language models to accurately predict clinical outcomes from structured patient data without altering their original parameters. ViSTA significantly improves prediction performance on benchmarks like MIMIC-IV, demonstrating its effectiveness in extending language model capabilities to clinical time-series understanding.
New LoRA Skills Should Read but Never Write
his paper introduces READ, a method for combining multiple LoRA adapters without interference. READ rewrites each adapter into a canonical form that preserves its original update, allowing them to be merged into a single model. This approach avoids the issues of weight-space interference, expensive retraining, and complex routing, enabling efficient composition of specialized skills.
PIA: A Personal Intelligence Agent Turning Health Conversations into Records and Records into Understanding
IA, a personal intelligence agent, addresses the limitations of general summarization for health data. Its core method involves a four-control memory harness (extraction, memory, retrieval, understanding) that converts health conversations into structured clinical records and then synthesizes these records into user understanding. PIA's contribution lies in its domain-agnostic memory mechanisms, enhanced by pluggable health modules, enabling more accurate and context-aware responses for health-related queries.
Stale-Document Poisoning: When Outdated Retrieval Overrides Correct Model Answers
his paper introduces "stale-document poisoning," a vulnerability in Retrieval-Augmented Generation (RAG) where outdated retrieved information causes models to provide incorrect answers, even if they would have been correct otherwise. The authors created a benchmark of knowledge reversals and demonstrated that outdated evidence significantly degrades model performance, particularly for recent factual changes. Their work highlights the critical need for temporal awareness in RAG systems to ensure the validity of retrieved information.
AutoRef: Harness Optimization for Agentic Multi-Reference Image Generation
his paper introduces AutoRef, a method to optimize the "harness" in agentic multi-reference image generation. AutoRef automatically tunes how reference images are processed and how generated outputs are evaluated to improve fidelity and naturalness. Its contribution lies in providing a systematic approach to enhance the complex decision-making process within these generative agents.
BaRe-Mem: Bayesian Reliability Memory for Robust and Adaptive Agent Consultation
aRe-Mem is a Bayesian memory system that estimates advisor reliability in multi-agent systems by analyzing the central model's beliefs and past interactions. This allows it to dynamically adjust the influence of advisor responses and decide whether to consult or reason autonomously, making it more robust to misleading information than existing methods. Its contribution lies in providing a principled and adaptive approach to agent consultation and worker allocation, improving performance and reliability.
Behavioral Foundation Models for Quality Diversity
his paper introduces BFM-QD, a framework that leverages the latent space of Behavioral Foundation Models (BFMs) for Quality-Diversity (QD) search. Instead of searching in high-dimensional policy parameters, BFM-QD efficiently discovers diverse and high-performing behaviors within a compact latent space. A key contribution is a novel, critic-free policy improvement operator that approximates policy gradients without requiring backpropagation.
Distillation Defenses Easily Break After Reinforcement Learning
his paper reveals that defenses against distillation attacks on large language models are often ineffective because they don't account for attackers further training their distilled models with reinforcement learning. This subsequent training significantly lowers the bar for attackers to steal advanced reasoning capabilities from closed-source models, undermining existing security measures.
FactorEngram: Factorized N-gram Memory with Basis-Level Gating for Language Models
actorEngram introduces a novel memory mechanism for language models that factorizes n-gram representations into shared basis vectors. This allows for more efficient parameter sharing and contextual gating at the basis level, enabling models to selectively retrieve relevant components of memory for polysemous patterns, unlike previous monolithic approaches. Its core contribution lies in enabling finer-grained memory access and improved semantic understanding through this factorized and gated memory structure.
Failure-Transparent Agents: Benchmarking Post-Failure Reporting in Tool-Using Language Models
his paper introduces Failure-Transparent Agents (FTA), a benchmark designed to isolate and evaluate how tool-using language models report failures. FTA controls for external factors, allowing direct auditing of post-failure claims by fixing the observed failure and required evidence. The contribution lies in providing a controlled environment to measure and improve models' honesty and accuracy in reporting tool failures, leading to significant reductions in false success and fabricated details.
MechBench: Can AI Scientific Agents Discover Mechanisms Beyond Phenomenal Laws?
his paper introduces MechBench, a new benchmark designed to evaluate AI scientific agents' ability to discover underlying mechanisms, not just observable laws. MechBench separates these capabilities by providing agents with observational data and context, then testing their understanding of internal scientific consequences through "mechanism probes." The benchmark's novel tasks are generated by mutating known mechanisms, pushing agents beyond memorization to true mechanistic discovery.
Not All Thinking is Created Equal: Latent Reasoning Discovers a Recurrent Search Algorithm for Depth Generalization
his paper investigates whether different reasoning methods in LLMs share a common underlying mechanism. By training and comparing various models on a multi-hop reasoning task, they discover that models performing reasoning in latent space generalize better to out-of-distribution problems than those relying on explicit intermediate traces like Chain-of-Thought. Their key contribution is identifying a recurrent search algorithm in latent reasoning that enables this improved depth generalization.
Reinforcing Agentic Creativity in Scientific Ideation with Night Science
his paper introduces AI Night-Scientist, a reinforcement learning framework designed to enhance LLM creativity in scientific ideation. It trains models to balance predictable "day science" with serendipitous "night science" by modeling creativity across action, process, and outcome. The core contribution is a method to generate more diverse and novel scientific proposals by teaching LLMs when and how to deviate from standard reasoning.
Representation Alignment as a Bottleneck in LLM-Based Retrosynthesis Planning
his paper argues that the failure of LLMs in chemical retrosynthesis planning stems from forcing them to simultaneously handle chemical analysis and planning language structuring. Their core method decomposes retrosynthesis into distinct stages: molecule mapping, reaction mapping, and PDDL generation. This approach successfully overcomes the bottleneck, demonstrating that representation alignment, not model capacity, is the key challenge, and highlighting the need for representation-centric design in future AI systems.
RSI-Master: Structuring Experiments to Guide Autonomous Model Improvement
SI-Master introduces a novel framework for autonomous AI model improvement by structuring experiments to prevent "hacking" and "strategy lock-in." It employs an "Experiment OS" for regularized actions and traceable records, alongside a "Reviewer-Guided Research Orchestration" system that dynamically organizes exploration through a research Directed Acyclic Graph (DAG). This approach ensures both safe, focused experimentation and broad exploration of research directions.
Share-Borne AI Virus: Memory-Hopping Attacks Across LLM Agents
his paper introduces "artifact-mediated propagation," a novel attack where malicious content spreads between independent AI agents through shared persistent artifacts like documents. The core method involves an attacker injecting harmful data into an artifact, which is then stored, reproduced, and passed to subsequent agents, effectively creating a self-propagating "virus." The contribution is demonstrating this memory-hopping attack's feasibility and potential for widespread propagation across interconnected AI assistants.
Shockingly Simple Self-retrospection Improves Agentic Models Without RL
his paper introduces Retrospection-Only Fine-Tuning (ROFT), a novel method where language model agents improve by generating and learning from explanations of their own past actions, without external rewards or teachers. The core contribution is demonstrating that this "self-retrospection" alone, through next-token prediction on explanations, significantly enhances agent performance on complex tasks like software engineering.
Character Training for Risk-Averse Agents
his paper proposes "character training" to instill risk aversion in AI agents, making them favor safer strategies and thus reducing catastrophic harm. The core method involves training agents with a "constitution" defining constant absolute risk aversion (CARA) over their resources, achieved through on-policy distillation. The contribution lies in demonstrating that persona traits are a robust mechanism for instilling risk preferences, leading to competitive and better generalizing agents compared to traditional training methods.
Correct Answers, Invalid Traces: What Verifiable Grade-School Math Reveals About Chain-of-Thought Traces
his paper investigates the reliability of chain-of-thought (CoT) traces as accurate records of model reasoning, particularly in grade-school math problems. By using a synthetic benchmark (iGSM) that allows programmatic verification of each step, they found that even when models produce correct answers, their CoT traces can be invalid, especially on harder problems. This highlights a significant gap between answer correctness and the faithfulness of the generated reasoning process.
Do LLM Agents Execute the Plans They Declare? From Planning-Mode Declaration to Pattern-Specific Execution
his paper investigates the "Plan Declaration-Execution Gap" in LLM agents, where agents may fail to execute the plans they declare. The core method, "Planning-as-Routing," uses an LLM to declare a planning mode (Predefined, Sequential, Hierarchical, or Search), which a deterministic router then directs to a specialized executor. The contribution is demonstrating that this pattern-specific execution significantly improves fidelity to declared plans compared to generic approaches, especially for complex tasks.
Dr. OPD: Learning What to Follow for Optimal On-Policy Distillation of Large Language Models
his paper introduces Dr. OPD, an improved on-policy distillation method for LLMs. Instead of treating all teacher signals equally, Dr. OPD learns to optimally weight these signals to maximize the student model's performance. This is achieved through a bilevel optimization approach that iteratively updates both the student policy and the signal weights.
Gender bias across LLMs is common and highly heterogenous
his paper investigates gender bias across ten diverse LLMs using two methods: analyzing phrase attribution and moral judgments. It reveals that gender bias is prevalent and varies significantly across models, with some exhibiting masculine-stereotyped phrase attribution to women and others showing a male-disadvantaging moral judgment asymmetry. The study's contribution lies in its broad, comparative analysis, highlighting the heterogeneity of LLM gender biases and their potential real-world implications.
Probability is Not Enough: Exploring and Counting Divergent Tokens for Reasoning Uncertainty Quantification in LLMs
his paper introduces Divergent Token Confidence (DTC), a novel method for quantifying reasoning uncertainty in LLMs. Instead of relying solely on token probabilities, DTC counts tokens where two models disagree significantly during decoding, using Jensen-Shannon divergence to identify these points. This count of divergent tokens is found to be negatively correlated with answer accuracy, providing a more robust signal of model confidence.
Retrieval-Augmented Skill Optimization via Cross-Harness Adaptation
his paper introduces Retrieval-Augmented Skill Optimization (RASO), a method to improve agent skills by leveraging existing knowledge. RASO addresses the inefficiency of current methods by retrieving relevant skills from a large corpus and adapting them to new tasks and environments using Cross-Harness Adaptation. This approach significantly reduces reliance on costly agent simulations by incorporating prior learning.
Video-RSI: Recursive Self-Improvement of Video Understanding Agents via Harness Evolution
ideo-RSI recursively improves video understanding agents by using their own language models to revise their observation and action harnesses. The agent actively investigates videos to gather evidence for competing failure explanations, grounding revisions in new observations. This cost-aware evolution leads to improved accuracy with reduced processing costs.
Which Attention Heads are like the Human Head? Not the Ones that Compute
his paper investigates the relationship between brain activity and Large Language Model (LLM) attention heads. It finds that heads aligned with human brain signals (EEG) are causally involved in task performance, but their ablation is less disruptive than heads identified by traditional attribution methods. This suggests that brain alignment and computational contribution are distinct, with concept-aligned heads showing more promise for brain-AI correspondence than function-based ones.
Explore Broadly, Reason Sharply: Push Small Models toward the Frontier via Sampling
his paper introduces Parallel Power Tempering (PPT) to improve reasoning in language models without costly retraining. PPT addresses the exploration-exploitation trade-off in existing sampling methods by running multiple model replicas at different "sharpening" levels. This allows for broader exploration of reasoning paths while still focusing on promising solutions, pushing smaller models closer to state-of-the-art performance.
Cogentic: Multi-Agent Orchestration for Automated Proof Discovery
ogentic is a multi-agent system designed for automated proof discovery on complex, open research problems. Its core method involves an iterative prove-verify loop where an orchestrator manages multiple independent "provers" exploring different mathematical directions. These proofs are then rigorously verified by specialized components, and confirmed intermediate results are stored to build upon in subsequent rounds, enabling the system to tackle challenging math and theoretical computer science problems.
ComputerSD: Online Self-Distillation from Real-Time Feedback for Computer-Use Agents
omputerSD addresses the challenge of sparse rewards in training computer-use agents by introducing online self-distillation. It leverages real-time feedback from GUI transitions to generate privileged guidance and a step-level value score. This dual approach allows for token-level learning from the guidance while regulating the learning
cua-speedrun: Standardized Benchmarking of the Speed of Computer-Use Agents
his paper introduces `cua-speedrun`, a standardized benchmarking framework for evaluating the speed and efficiency of computer-use agents (CUAs). It addresses the reproducibility crisis in CUA benchmarks by providing a uniform virtual machine setup and execution pipeline. `cua-speedrun`'s core contribution is enabling reliable speed comparisons of CUAs, crucial for their wider adoption.
EviRover: Reinforcing Agentic Perception Beyond a Glance
his paper introduces EviRover, a novel agentic perception system that overcomes the limitations of single-glance visual analysis. Its core method involves training an agent to interactively gather more information beyond an initial observation, addressing "perception under insufficient evidence." The key contribution is the creation of new datasets and the EviRover agent itself, designed to learn this interactive perception process.
Learning from Research: Toward Lifelong Agent Harness Evolution
his paper introduces ScholarEvolve, a framework for continually improving language agents by evolving their "harness" (tool use, memory, execution logic) without retraining the core language model. Its core method involves using topic modeling on research literature to identify and implement novel improvement strategies for different harness modules, enabling proactive and exploratory adaptation. The main contribution is a system that leverages external research knowledge to guide harness evolution, overcoming limitations of purely reactive, feedback-driven methods.
What Can Component-Replacement Evidence Establish? A Critical Scoping Review of Local Decisions in LLM Agents
his paper critically reviews how component replacement evidence is used to evaluate LLM agents. It finds that while many studies report local decision metrics and task outcomes, few establish a clear causal link between the two due to insufficient matched comparisons. The core contribution is highlighting the need for better experimental design to accurately assess the impact of local decision quality on overall task performance.
Who Verifies the Graph? Misspecification Attacks on Causal Action Verification for Language Agents
his paper introduces "misspecification attacks" that exploit vulnerabilities in causal action verifiers used by language agents. By subtly corrupting the underlying causal graph, these attacks can trick verifiers into approving harmful actions, even though the verifier issues a seemingly valid certificate. The authors propose an attestation step to detect such attacks by cross-referencing certified executions with randomized samples.
OpenTSLM TeeMoE: A Unified Time-Series Language Model for Forecasting, Contextual Prediction, and Reasoning
penTSLM TeeMoE is a unified time-series language model that addresses the fragmentation of current models by integrating forecasting, contextual prediction, and temporal reasoning. Its core method involves a mixture-of-experts architecture with three specialized, low-rank experts for aggregation, native forecasting, and temporal analysis, controlled by a learned weighting mechanism. The key contribution is enabling a single model to perform these diverse time-series tasks effectively, even synthesizing predictions from external numerical specialists.
EvoDuet: Bilevel Co-Evolution of Web Searching and Task Solving for Scientific Discovery
voDuet addresses LLM limitations in scientific discovery by co-evolving solutions and search queries. It uses a retrieval gate to dynamically decide when to search for new information or reuse existing documents based on knowledge gaps. This bilevel optimization significantly improves discovery gain for some LLMs by intelligently guiding web searches to yield more relevant documents for problem-solving.
Admission Without Answers: Label-Free Certification and Experience Learning for LLM-Based Optimization Modeling
his paper introduces ADMITOR, a novel label-free method for admitting reliable optimization models generated by LLMs. ADMITOR achieves this by generating multiple models, testing them on resampled problem instances, and admitting those with consistent optimal values across different model families. Its key contribution is enabling robust experience learning for LLM-based optimization without requiring ground-truth answers, significantly improving the precision of admitted candidate models.
Affective Flow Language Model for Emotional Support Conversation
his paper introduces the Affective Flow Language Model (AFlow) to improve emotional support conversations. AFlow models the dialogue as an "affective utility flow," allowing it to estimate the value of intermediate conversation states and potential strategies. Its contribution lies in the Affective Flow Preference Optimization (AFPO) method, which uses this flow concept to propagate feedback from overall dialogue success to guide the model's sequential strategy decisions in multi-turn interactions.
Agentic Limit Order Books: Phase Transitions and Market Impact
his paper introduces a novel framework of Limit Order Books populated solely by autonomous reinforcement-learning agents. The core method involves simulating agent interactions within an order-matching engine to study emergent market phenomena. The key contribution is the discovery of distinct phase transitions in trading regimes, from orderly price discovery to volatile cascades, and the identification of non-classical market impact dynamics driven by agent feedback loops.
Fenchel Tilting: Weighted Correction for Efficient Finetuning of Generative Models
his paper introduces Fenchel Tilt Flow Control (FTFC), a novel method for efficiently fine-tuning generative models to align with arbitrary utility functions. FTFC decouples utility optimization from model fitting by learning
Game-Guided Skill Discovery through Self-Play for Playable Agent Control
his paper introduces Game-Guided Skill Discovery (GGSD), a framework that leverages self-play in games to learn human-playable motor skills. By training a hierarchical agent to compete against itself, GGSD discovers semantically distinct and interpretable skills that act as compact behavioral abstractions for agent control. The core contribution is enabling direct human control of agents using a small set of learned, playable skills, leading to emergent complex behaviors through skill transitions.
Grounding Time-Series Foundation Models in Digital Twin Topology for Predictive Maintenance
his paper addresses the challenge of applying time-series foundation models (TSFMs) to predictive maintenance within digital twins. The core method involves benchmarking existing TSFMs and proposing a novel topology-informed fusion approach. The main contribution is demonstrating that multivariate TSFMs perform better for remaining useful life prediction in digital twins and introducing a method to leverage digital twin topology for improved adaptation.
How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering?
his paper investigates the necessity of complex "harnesses" for autonomous machine learning engineering (MLE) agents. The core finding is that advanced, multi-agent orchestrators offer no significant advantage over a simple
Inference Auctions
his paper proposes an "inference auction" to address compute capacity limitations for LLM APIs. The core method is an auction mechanism that allows users to bid for faster inference service, efficiently allocating priority without sacrificing latency. The key contribution is a system that enables economically efficient priority allocation, incentivizes truthful bidding through fast algorithms, and includes an autobidding agent to maximize user utility within budget constraints, ultimately increasing system welfare.
