Yu Li

This work identifies and formalizes this failure mode, which is term destructive resource preemption: obtaining the resources required for a requested task by terminating, overwriting, evicting, or degrading an incumbent task.

The Cross-modal Perception Alignment Adapter (CMPA) is proposed, a manifold-aware framework designed to disentangle perceptual distortions from dominant semantics, and significantly outperforms state-of-the-art methods.

This work proposes UniRec, a Unified Cross-stage Recommendation Fusion model, and introduces a dual-axis preference alignment objective, and finds that unconstrained end-to-end fusion optimization can exploit imbalances in item attribute distributions, over-concentrating on high-reward regions at the cost of other objectives.

Multi-agent LLM training falters when updates are decoupled from joint state transitions; grouping interacting agent outputs into cardinality-normalized set actions solves credit assignment across both static and dynamically routed systems.

Multi-Tool Image Editing Attribution reveals that existing methods fail to keep pace with the complexity of modern image edits, but a new approach can accurately identify multiple tools at work.

Harness-policy co-evolution can reduce adverse safety responses by 3x while simultaneously boosting benign utility in LLM agents.

Silent data corruption can derail LLM training, but TrainSDC offers a targeted solution that maintains performance even under fault conditions.

StepGuard slashes attack success rates by over 77% while maintaining nearly all utility, setting a new standard for safety in LLM-based agents.

Task order can dramatically skew the performance of self-improving agents, revealing hidden dependencies that could undermine their reliability in real-world applications.

The shift from monomer folding to generative modeling could revolutionize our approach to understanding complex protein interactions and design.

Covert channels in LLM traffic can leak private information without any malicious intent, revealing a surprising vulnerability in seemingly benign interactions.

Evolving safety harnesses using trajectory data can reduce agent safety risks by over 3x while enhancing overall utility.

Over 1,500 submissions revealed stark differences in model performance across diverse domains, highlighting the challenges of generalizing egocentric video understanding.

Language's role in embodied agents is often overstated, with many claims lacking robust empirical support, revealing a critical gap in our understanding of its contributions.

AgentDoG 1.5 proves you can achieve GPT-5.4-level agent safety with open-source models trained on just 1k samples, slashing deployment overhead by two orders of magnitude.