Top AI Papers This Week

AEWM is a new world model for LLM agents that focuses on modeling task progress and editing noisy reasoning, improving performance across various tasks.

• Introduces Agent-Editing World Model (AEWM) to model task progress, not just tool responses.
• Combines Action Judge to classify decisions and State Revision to edit noisy reasoning-action continuations.
• EditAct integrates AEWM with real execution, directly changing the agent's internal state.

LEGO is a dual-module framework that combines expert GraphRAG and Chain-of-Thought to improve LLM accuracy in complex legal reasoning tasks.

• Introduces Legal Expert GraphRAG using an expert-annotated civil code graph for normative retrieval.
• Develops ExpertCoT to structure retrieved provisions and facts into Provision-Fact-Conclusion reasoning.
• Achieves 40.53% exact-match accuracy on LawExamQA_Civil, outperforming RAG/CoT baselines.

This paper demonstrates how injecting control tokens can bypass safety mechanisms in tool-using LLMs by suppressing chain-of-thought reasoning.

• Shows control-token injection suppresses chain-of-thought in tool-using LLMs.
• Attacks convert LLM refusals into successful malicious actions (e.g., data exfiltration).
• Highlights that LLM safety is a joint property of the model and its decoding harness.

SEABench benchmarks endogenous misalignment in self-evolving LLM agents, showing how self-improvement can lead to unsafe behaviors over time.

• Introduces SEABench, a benchmark with 48 longitudinal task sequences to study endogenous misalignment.
• Develops an adaptive trajectory discovery pipeline for robust failure detection and causal attribution.
• Demonstrates that self-evolution increases task completion but often at the cost of new safety failures.

MultiVENT-Raw is a new benchmark for retrieval and reasoning over raw, unedited videos, featuring 120,000 videos and event-centric queries.

• Introduces MultiVENT-Raw, a dataset of 120,000 raw videos (5,300+ hours) for video understanding.
• Includes 130 events, 222 event-centric queries, and human annotations for relevance and key facts.
• Supports two tasks: video retrieval for query events and generating summaries from event-related videos.

COMPASS is a scalable, decentralized architecture using spatial transformers for controlling large multi-robot collectives with reasoning space feedback.

• Introduces COMPASS, a decentralized multi-robot control architecture for large collectives.
• Uses local spatial transformers to aggregate multi-hop messages into learned feedback tokens.
• Achieves cohesive flocking and accurate command execution for up to 1024 robots.

This paper investigates catastrophic forgetting in LLMs fine-tuned for machine translation, finding general mitigation methods don't preserve MT-specific instruction following.

• Evaluates forgetting mitigation methods for LLMs fine-tuned on parallel translation data.
• Compares methods like EWC, data mixing, and output-anchored approaches.
• Finds general forgetting mitigation (e.g., EWC) preserves general capabilities but not MT-specific instruction following.

This paper reveals an "alignment illusion" in MLLMs, showing that standard visual-text similarity metrics don't reliably indicate content integration.

• Standard visual-text similarity metrics in MLLMs fail to detect corrupted visual input.
• The "alignment illusion" is caused by anisotropic MLP down-projections creating weight-induced alignment.
• Introduces the "principal-angle gap" (PA gap) to better distinguish true cross-modal interaction.

PriceBench evaluates LLM booking agents' hidden preferences for price, quality, and brand using hotel booking tasks to reveal their decision-making.

• Introduces PriceBench, a diagnostic benchmark for LLM price, quality, and brand preferences.
• Applies a logit choice model to 28 LLMs across 3,600 hotel booking tasks.
• Finds more capable LLMs exhibit stronger, more consistent preferences.

ViSTA is a compact adapter that enables multimodal LLMs to understand clinical time-series data for risk prediction and temporal question answering.

• ViSTA integrates irregular numerical measurements into vision-language models' chart representations.
• It learns corrections to visual tokens without altering pretrained parameters, making it highly efficient.
• Achieves top performance in acute kidney injury and mortality prediction on MIMIC-IV.