Top AI Papers This Week

LLM agents struggle significantly with structural constraints in backend code generation, showing "constraint decay" as requirements accumulate.

• Systematic study on LLM agents' ability to handle structural constraints in multi-file backend code.
• Introduces "constraint decay": agent performance substantially declines as structural requirements accumulate.
• Agents perform better in minimal frameworks (Flask) but struggle in convention-heavy ones (FastAPI, Django).

GAP proposes a granular alignment paradigm to stabilize visual latent reasoning in MLLMs by addressing feature-space mismatches, improving performance.

• Identifies and addresses a feature-space mismatch in MLLMs causing unstable visual latent reasoning.
• Introduces GAP with feature-level alignment using a PCA-aligned latent head for input-compatible latents.
• Incorporates context-level alignment with auxiliary visual supervision and capacity-guided selective supervision.

OmicsLM is a multimodal LLM that connects quantitative omics data with natural language for biological reasoning, outperforming existing models.

• Introduces OmicsLM, a multimodal LLM linking quantitative omics profiles with natural-language biological tasks.
• Represents transcriptomic data as compact continuous representations within the LLM context for multi-sample processing.
• Trained on 5.5M examples across 70+ task types, covering diverse biological reasoning challenges.

OpenSearch-VL provides an open-source recipe for training frontier multimodal deep search agents, achieving state-of-the-art performance.

• Curated high-quality training data (SearchVL-SFT-36k, SearchVL-RL-8k) reducing retrieval shortcuts.
• Developed a diverse tool environment integrating text/image search, OCR, and image processing.
• Introduced a fatal-aware GRPO algorithm to manage cascading tool failures in multi-turn interactions.

Open-SAT improves open-vocabulary satellite image retrieval by using LLMs to refine query embeddings at inference time, achieving significant F1 score gains.

• Refines VLM-generated query embeddings using LLMs for better alignment with satellite imagery.
• Operates as a training-free, inference-time algorithm, avoiding additional model training.
• Leverages contextual information from LLMs about objects and their surroundings for enhanced retrieval.

This paper introduces a framework to evaluate LLM research agents' source attribution, revealing high link validity but low factual accuracy in citations.

• Developed a novel framework using an AST parser to evaluate LLM citation quality at scale.
• Evaluates citations across three dimensions: link accessibility, content relevance, and factual accuracy.
• Benchmarked 14 LLMs, finding high link validity (>94%) and relevance (>80%) but low factual accuracy (39-77%).

CyBiasBench reveals LLM cyber-attack agents exhibit inherent biases, concentrating efforts on specific attack families regardless of prompts.

• Introduced CyBiasBench, a 630-session benchmark to quantify attack-selection bias in LLM cyber agents.
• Found explicit, agent-specific biases in attack family allocation, independent of attack success rates.
• Identified a "bias momentum effect" where agents resist steering, without improving attack performance.

This paper proposes the Causally Emergent Alignment Hypothesis, showing that causal emergence in RL agents predicts final reward and aligns with learning.

• Introduces the Causally Emergent Alignment Hypothesis for RL agents.
• Shows causal emergence (CE) in latent spaces predicts final reward early in training.
• Demonstrates CE dynamics align with reward improvement across diverse RL tasks.

Meow-Omni 1 is the first quad-modal MLLM for feline ethology, fusing video, audio, physiology, and text to achieve SOTA intent recognition.

• Introduces Meow-Omni 1, the first open-source quad-modal MLLM for computational ethology.
• Fuses video, audio, physiological time-series, and text for deeper feline intent understanding.
• Achieves state-of-the-art 71.16% intent-recognition accuracy on the new MeowBench benchmark.

LITMUS benchmarks LLM agent behavioral jailbreaks in real OS environments, revealing critical safety gaps and a new "Execution Hallucination" phenomenon.

• Introduces LITMUS, a benchmark for LLM agent behavioral jailbreaks in real OS environments.
• Utilizes semantic-physical dual verification and OS-level state rollback for robust testing.
• Reveals agents execute dangerous operations (40.64%) and suffer "Execution Hallucination."