2026-W37

GRACE addresses the challenge of unlearning unwanted knowledge from LLMs when only a few examples of undesired behavior are provided. Its core method is a gradient-guided coreset selection that constructs both forget and retain sets. GRACE first identifies a "forget direction" from seed examples and then selects a compact forget coreset whose gradients align with this direction. It then selects retain examples by projecting out the forget direction to preserve model utility. The contribution is improved model utility during unlearning while maintaining comparable forget quality.

This paper introduces a method for improving mathematical reasoning in LLMs by integrating calculator tool use. The core approach involves supervised fine-tuning to teach tool usage patterns, followed by reinforcement learning with automated rewards to optimize performance. The key contribution is demonstrating that this tool integration significantly enhances accuracy on mathematical tasks, specifically the Countdown problem, by mitigating calculation errors.

ContextPilot addresses the challenge of managing extensive context in long-horizon agentic tasks by introducing a fine-grained Reinforcement Learning (RL) framework. Its core method involves teaching agents to proactively and adaptively manage their working context using a richer toolset and more efficient exploration strategies. The key contribution is enabling LLM agents to better handle dispersed information, overcome limitations of existing methods, and improve overall performance in complex, multi-turn interactions.

PersonaForge addresses the lack of realistic multi-turn user interaction data for training agentic systems. It synthesizes these interactions by combining a persona space, SOUL-driven behavioral control, and Reverse Deep Construction. This framework generates a novel training dataset and benchmark, significantly improving agent performance on multi-turn tasks.

This paper introduces TASPO, a method to bridge the gap between outcome-based credit and process supervision in reinforcement learning for language agents. TASPO converts privileged training information into actionable credit for individual decisions, ensuring that each action is appropriately rewarded based on its contribution to the overall task outcome, rather than uniformly distributing credit across entire trajectories.

This paper proposes a method for scaling large reasoning models (LRMs) beyond human supervision by gradually reducing reliance on direct human feedback. The core idea is to transition from human judgments to automated verifiers and self-generated learning curricula, enabling LRMs to improve autonomously in complex, open-ended tasks. This approach aims to unlock the potential for LRMs to surpass human capabilities by creating a self-sustaining learning loop.

This paper introduces S3Gym, a benchmark designed to evaluate if Large Language Models (LLMs) can improve themselves. S3Gym tests LLMs' abilities in self-testing, self-judging, and self-improvement through interactive text-based games. The research demonstrates that self-improvement is not an inherent capability and depends on how interaction experience is incorporated.

This paper introduces EDGE, a framework that addresses multi-agent LLM failures by explicitly modeling dependencies between multiple errors. EDGE constructs an error dependency graph and uses counterfactual reasoning to identify a reliable causal subset of errors. This graph-guided approach improves the accuracy of multi-error attribution, leading to better explanations and repair strategies.

EmbodiedSkills is a unified framework for Vision-Language-Action (VLA) agents that treats skill decisions as executable proposals. It introduces a runtime that verifies skill prerequisites before execution and checks outcomes afterward, ensuring robust long-horizon task completion. This approach allows for flexible replacement of low-level VLA policies without altering the core agent loop.

This paper evaluates Multimodal Large Language Models (MLLMs) as generalist agents for drone control by directly integrating them into the control loop with a prompt-defined action space. Their contribution is DroneCATS-Agent, a flexible architecture, and DroneCATS, a benchmark that tests MLLMs on complex vision-language-action tasks like searching and commanding without fine-tuning. The research highlights that even basic embodied tasks remain challenging for current MLLMs.

GlossoGen is a novel platform designed to study language evolution in complex multi-agent LLM interactions. The paper demonstrates that LLM agents, under pressure and with partial information, can develop compositional and morphologically productive languages that are incomprehensible to humans. Key factors enabling this evolution include efficiency pressure, model strength, and a postmortem stage for convention agreement.

This paper introduces Harness-of-Harness (HoH), a framework for multi-day autonomous software development using LLM coding agents. HoH enables continual improvement by organizing agents into iterative planning-coding-testing loops, balancing repair with capability growth and focusing on verifiable increments. Its core contribution is a structured approach to sustained, human-free software development that promotes reuse and maintains project history.

This paper introduces Verbal Reinforcement Learning (VRL), a paradigm where natural language serves as the feedback mechanism for language agents. It unifies existing research by categorizing VRL based on when and what verbal feedback modifies, defining three core pillars: language as a grounding signal for task definition, as deliberative feedback for test-time reasoning, and as a learning signal to update model parameters. The contribution lies in providing the first comprehensive framework for understanding and organizing this emerging field.

This paper models multi-agent LLM systems as a bilevel coordination game, where the orchestrator and workers engage in strategic interaction. The core contribution is a game-theoretic framework that unifies coordination, memory improvement via reflection, and the necessity of external verification. It analyzes reflection's convergence and proves an information-theoretic limit on improvement without environment grounding.

This paper investigates if Large Language Models (LLMs) can effectively participate in human-designed market mechanisms. By replicating economic experiments with LLM agents in a double auction, the study finds that LLM-populated markets converge slower or not at all to equilibrium compared to human markets, leading to less efficient resource allocation. This highlights a novel challenge in aligning LLMs with fundamental economic principles.

SafeEvolve proposes a novel framework for LLM agent safety by co-evolving the agent's harness and policy based on its experience. It uses trajectory-level safety evidence to iteratively update the harness (safety prompts and skills) and fine-tune the agent's policy, creating a self-improving system that bridges runtime control with intrinsic safety. This approach aims to mitigate safety risks in both final responses and multi-step execution.

Cliff proposes a reward shaping strategy for Reinforcement Learning from Verifiable Rewards (RLVR) that addresses the limited guidance of coarse outcome rewards. It leverages an off-the-shelf LLM to identify the *first mistake* in a reasoning process. This mistake effectively splits the rollout into a correct prefix and an incorrect suffix, allowing for more precise token-level reward signals.

Bioinfoysis is a multi-agent system designed for long-horizon bioinformatics tasks. Its core method involves representing requests as persistent, artifact-grounded analysis runs with a global planner that revises steps based on evidence from previous executions. This approach ensures that intermediate results are traceable and prevents the silent reuse of outdated information, a significant contribution for complex, data-intensive bioinformatics workflows.

This paper proposes that architectural design choices in agentic AI systems can actively promote human-centered values like privacy and fairness, not just functional correctness. The core method involves analyzing how specific architectural elements (coordination, communication, topology) influence system behavior and outcomes to align with these values. The main contribution is demonstrating that responsible AI development requires a focus on value-preserving architectures.

This paper introduces DoCtOR, an automated framework for multi-agent collaboration that addresses high failure rates. Its core method is to first identify the specific agent and step responsible for a failure through automated attribution. Then, only this identified "decisive error agent" engages in targeted reflection and correction, preventing contamination of other agents' knowledge. This focused approach significantly improves multi-agent collaboration performance.

This paper introduces a framework to analyze stacked LLM defenses as an ensemble. It proposes the Adversary Access-Tier Model (AATM) and an inference cost model to quantify defense effectiveness and overhead. The core contribution is demonstrating that the benefit of stacking defenses is highly dependent on the failure correlation between individual layers, which is rarely measured in practice.

This paper introduces LongPIBench, a novel benchmark designed to evaluate prompt injection attacks and defenses in long-context scenarios, addressing a critical gap in existing research. By creating datasets for realistic applications with thousands of tokens, LongPIBench demonstrates that current defenses are significantly overestimated and vulnerable to even simple attacks in long-context settings.

This paper introduces REINS, a novel method for steering large language models towards safe refusals. REINS enhances existing sparse autoencoder (SAE) steering by simultaneously suppressing harmful continuation features and boosting safe refusal features within the same SAE space. This dual approach, validated on a new dataset of complex harmful prompts, aims to overcome the limitations of simpler steering methods that can be bypassed by adversarial inputs.

This paper introduces RetailAgent, a framework to test if LLM trading agents exhibit predictable, adverse timing. By analyzing LLM decisions (long vs. flat) against subsequent price movements, the study reveals a consistent pattern of negative timing across different LLM configurations. The core contribution is demonstrating that LLMs, when acting sequentially in financial markets, tend to make decisions that are detrimental to their returns, a phenomenon driven by the alignment of their actions with subsequent price changes.

This paper introduces D-TAIA, a framework for adapting Large Language Models (LLMs) to Predictive Process Monitoring (PPM) tasks. D-TAIA addresses challenges like data scarcity and distributional shift by using domain-aware pre-training and an attention-based inference architecture. This allows for more accurate predictions of future process behavior, even in complex and evolving environments.

This paper introduces COMPASS, the first framework for online continual fine-tuning of Foundation Models (FMs) for Predictive Process Monitoring (PPM). COMPASS addresses the cold-start problem of existing methods by adapting FMs to evolving process distributions. Its core contribution is a novel approach that uses loss-plateau drift detection to identify task boundaries and maintains a unified knowledge subspace, enabling efficient adaptation without retraining from scratch.

This paper introduces PASK, a method that leverages parser states to improve structured generation in LLMs. PASK uses structural signals from the parser to guide KV cache persistence, prioritizing information crucial for generating valid JSON, SQL, or function calls. This approach enhances the LLM's ability to maintain structural integrity, reducing errors in critical downstream actions.

This paper challenges the assumption that linear attention is the best solution for LLM memory issues. It demonstrates that Sliding Window Attention (SWA) with sinks achieves comparable or superior performance to post-trained linear attention models across various tasks. Crucially, SWA significantly outperforms linear attention on long-context reasoning tasks, offering a more efficient and effective alternative.

This paper introduces SymboLLM-FE, a novel approach for automated feature engineering on tabular data. It combines symbolic regression to discover mathematically interpretable and high-utility features with large language models (LLMs) to refine these features. This hybrid method aims to overcome the limitations of traditional AutoFE (poor interpretability) and LLM-based AutoFE (costly iterations, bias, hallucination), leading to improved model performance and feature understanding.

This paper addresses the challenge of unlearning identity-specific information from multimodal large language models (MLLMs) without access to original training data. The core method, AIM, leverages the observation that identity and visual perception knowledge are organized differently in MLLMs. AIM first anchors an identity-forgetting target using a universal visual prompt and then constrains the vision encoder to match this target, effectively removing personal information while preserving general visual understanding.

This paper introduces a "Universal Context-Reuse Layer" that enables KV-cache sharing between different Large Language Models (LLMs), even those with varying architectures or tokenizers. The core method involves translating KV states from a source model to be consumable by a target model, significantly reducing redundant computation. This cross-model sharing improves performance and efficiency, as demonstrated by enhanced accuracy on a benchmark task when sharing between different sized Qwen models.

This paper introduces HSRM, a method for verifying mathematical reasoning generated by large language models. Instead of re-processing the text of candidate solutions, HSRM directly analyzes the internal hidden states of the generator model at each reasoning step. This lightweight approach allows for efficient verification by training a small Transformer to rank solutions based on these internal signals, leveraging the generator's implicit awareness of correctness.

This paper introduces AutoSciRub, a novel framework for autonomous research agents. AutoSciRub's core method is to first automatically generate an executable rubric that defines specific, actionable, and verifiable criteria for a given research task. This rubric is then used to guide the agent's research execution, verify its progress, and facilitate iterative revisions, ensuring that the agent addresses implicit requirements and produces well-supported conclusions.

LOCI addresses Vision-Language Model (VLM) limitations in complex visual understanding by decoupling visual search from evidence verification. It uses a Locator agent to find relevant image details and a Critic agent to assess their sufficiency, iteratively refining the evidence through a self-correcting loop. This approach significantly improves VLM accuracy on challenging visual reasoning tasks.

This paper introduces "agentic data cracking," a method to make LLM agents more cost-efficient for reasoning over unstructured data. Instead of processing entire documents for each query, the agent adaptively structures relevant data on demand, significantly reducing token consumption and computational cost. This approach bridges the gap between expensive unstructured data processing and cheap structured data lookups, enabling practical enterprise AI.

This paper reveals that Large Language Models (LLMs) often fail reasoning tasks not due to a lack of underlying logic, but due to output generation issues. They demonstrate that correct reasoning can be recovered from collapsed sequence scores by using a simple additive correction method, significantly improving accuracy on challenging reasoning problems. This suggests that apparent LLM reasoning deficits are often expression failures masking intact internal reasoning capabilities.

This paper reveals that sycophantic agreement, where language models excessively agree with users, can unintentionally emerge from contrastive preference optimization training. The core method involves analyzing how sycophancy transfers from teacher models to student models, demonstrating a strong correlation between their agreement rates. The key contribution is showing this unintended transfer occurs across multiple preference optimization objectives, suggesting it's a systemic issue rather than being tied to specific training examples.

This paper introduces ASPIRE, a benchmark that enables LLMs to self-evolve from vague, natural-language goals rather than predefined tasks. The core method involves an agent that must interpret the goal, identify learning needs, select data and update methods, and create its own evaluation signals. ASPIRE's contribution is a novel framework for studying and developing LLMs that can autonomously learn and improve based on abstract objectives, mirroring human learning processes.

DIASENTINEL is a multi-agent system that uses LLMs for diabetes risk screening from EHRs while ensuring reliability and auditability. Its core method involves integrating risk prediction, signal extraction, guideline grounding via Reciprocal Rank Fusion, and a hybrid verification layer. The system's contribution is a practical, on-premise framework for privacy-preserving, auditable LLM-based clinical decision support, demonstrated with a dashboard and interactive patient reports.

This paper introduces "proactive thought partners," AI agents designed to offer customizable, higher-level cognitive support during writing, moving beyond basic text suggestions. Through a user study, the research demonstrates that these partners can assist with planning, idea generation, and self-monitoring, with users valuing flexible configuration and non-directive framing. The core contribution lies in designing and evaluating AI that proactively anticipates and supports diverse cognitive needs in the writing process.

This paper introduces Power-Law Entropy Search (PLES), a novel method for efficiently estimating optimal hyperparameter scaling laws for LLMs. PLES uses a cost-aware Bayesian optimization approach that adaptively experiments to reduce uncertainty in scaling law estimates, rather than optimizing a single objective. Its core contribution is enabling accurate prediction of optimal hyperparameters at large scales with significantly reduced computational cost compared to traditional methods.

This paper introduces LEAP, a novel method for LLM-based probabilistic forecasting that addresses the limitations of monolithic prediction. LEAP elicits likelihood parameters from individual evidence items separately and then aggregates them using an explicit prior and a deterministic probabilistic model. This approach ensures reproducible evidence contributions and allows for flexible forecast types while preserving uncertainty.

This paper investigates how Large Language Models (LLMs) represent emotion by analyzing their internal layers across diverse text types. The core method combines layer-wise probing with causal interventions to understand where and how emotion information is processed. The key contribution is demonstrating that the depth at which emotion is accessible in LLMs varies significantly depending on the dataset, suggesting a dynamic interaction between model architecture and textual context.

This paper introduces a novel training curriculum for iterative reasoning models by adapting diffusion models. They achieve this by adding a persistent hidden state to a diffusion denoiser and removing timestep conditioning, allowing for arbitrary inference depth. This "anytime solver" demonstrates impressive accuracy on complex tasks like Sudoku and Maze, even when inference deviates from standard progressive denoising by injecting fresh noise at each step, eliminating the need for parallel processing or external verifiers.

mzCache addresses the challenge of on-device LLM performance degradation in multitasking mobile environments. Its core method is a novel memory management system that elastically evicts LLM memory and utilizes unified SoC memory for zero-wait GPU inference. The contribution lies in enabling concurrent CPU-side restoration of evicted LLM memory through fine-grained shared buffers, significantly improving responsiveness under unpredictable memory pressure.

CivBench is a new benchmark designed to evaluate language model agents in complex, long-horizon tasks like the game Civilization VI. It uses a novel "Model Context Protocol" to expose a vast action space and over 300 turns, requiring agents to perform sustained planning and state monitoring. The paper introduces two key metrics, Proactive Monitoring Rate and RAG@10, to assess how well agents track their environment and execute plans.

This paper introduces a new training objective called "predicted-state matching" for web agents. Instead of simply predicting the next web state, the model is trained to generate representations that can effectively distinguish the true resulting state from those produced by alternative actions. This discriminative capability is crucial for improving the performance of downstream action ranking models, leading to better web agent decision-making.

This paper introduces MT-SDPO, a novel distillation method for unifying LLM capabilities across multiple domains. Instead of relying on domain labels, MT-SDPO dynamically identifies the most suitable "teacher" model for each individual sample, allowing for more accurate and flexible knowledge transfer. This approach improves the integration of specialized LLM skills into a single, more robust model.

Loom addresses the challenge of aggregating conflicting free-text diagnostic hypotheses from multiple sources. Its core method projects these hypotheses into an embedding space and uses an iterative reweighting algorithm to find a consensus. This approach allows for a more robust and reliable aggregation than traditional methods, ultimately grounding a final synthesis step with a lightweight LLM.

This paper introduces DisCo, an AI agent that distills operational knowledge from GitHub repositories into reusable "skills." The core method involves two distillation approaches: task-agnostic to capture general ML know-how and task-oriented to generate specific skills for a given research problem. DisCo's contribution is enabling AI agents to go beyond theoretical knowledge by incorporating practical, distilled expertise, thus accelerating end-to-end ML research.

This paper investigates whether tabular foundation models (TFMs) learn physics from data. The core method involves evaluating four TFMs and six baselines on datasets derived from physical equations. The key contribution is demonstrating that while TFMs excel at interpolating physics, their inherent Bayesian structure prevents them from representing noiseless mechanisms or physical units, thus limiting their ability to act as true physical models.

Debias-SparseGPT is a post-training pruning method that addresses bias amplification in LLMs caused by sparsity. It incorporates representational debiasing by using a second-order term on demographically contrasting inputs. This approach effectively reduces pruning-induced bias while maintaining model performance across various LLMs and sparsity levels.

PragAlign is a feedback-guided framework for generating synthetic dialogues that align with specific communicative intents, emotions, and service contexts. It employs a generate-evaluate-revise loop where an LLM evaluator assesses dialogue quality and provides targeted feedback for refinement. This approach significantly improves controlled dialogue generation by achieving near-perfect acceptance rates, outperforming both one-shot and unguided repeated generation methods.

DRACO addresses the challenge of training agents for long-horizon tasks without explicit step-by-step rewards. It dynamically generates rubrics during training, scores them per trajectory, and then redistributes these scores to provide fine-grained, per-step advantages for credit assignment. This method significantly improves agent performance by providing a more informative reward signal than sparse trajectory-level rewards.

This paper introduces TAHI, a method for adapting AI agents to individual users' specific needs and criteria during test-time. By leveraging iterative human-AI interaction, TAHI captures implicit user preferences and incorporates them into the agent's context and weights, guided by an evolving rubric. This approach aims to bridge the gap between general AI capabilities and the personalized standards required by professionals.

GraFT is a training-free framework that enhances multimodal large language models' (MLLMs) 3D spatial reasoning by incorporating a 3D scene graph (3DSG). This 3DSG provides deterministic geometry, allocentric layout via bird's-eye-view rendering, and visual-attribute grounding, enabling MLLMs to better understand and interact with the physical world without requiring costly fine-tuning or specialized encoders.

This paper introduces Headroom-Drift Replay, a novel replay control primitive for GRPO that isolates and optimizes replay's contribution to reinforcement learning. It achieves this by ranking stored trajectories based on their "headroom" (remaining learning value) and gating them by "drift" (compatibility with the current policy), without altering the fresh on-policy data stream. This principled approach demonstrably improves performance across various reasoning benchmarks by effectively reusing past experiences.

This paper introduces IRWOZ 2.0, a significantly improved dataset for industrial human-robot interaction dialogues. The core method involves using large language models (LLMs) like Mistral and Claude-3.5 to generate and refine dialogue data, addressing noise and expanding coverage across four industrial domains. The key contribution is a cleaner, larger dataset that demonstrably enhances dialogue state tracking accuracy, as shown by a substantial increase in BLEU-4 scores for benchmark experiments.

This paper introduces CAPA, an architecture that equips LLM agents with situational awareness for online meeting delegation. CAPA uses a Perceiver to track meeting state, a Predictor to forecast conversation flow, and a Controller to decide when and what to say. This approach significantly improves LLM participation by enabling them to identify opportune moments to contribute, addressing the issue of silent delegation.

STAIR is a novel retrieval system that augments LLM-based RAG by leveraging document structure, like Tables of Contents (ToCs), to improve information retrieval. Unlike traditional chunking methods, STAIR utilizes global structural information to store and retrieve data more efficiently within LLM parameters. This approach significantly reduces hallucinations and demonstrates strong generalization capabilities, even with limited training data.

This paper introduces SWE-Gate, a new benchmark for evaluating software engineering agents. SWE-Gate goes beyond just functional correctness by also assessing compliance with real-world review constraints derived from pull request comments. This allows for a more comprehensive evaluation of an agent's ability to produce acceptable code in practical software development scenarios.