Peng Xia

Multimodal models can achieve significant image generation improvements by learning from their own critiques during test-time compute, effectively acting as their own recursive self-improvement mechanism.

Regularized Recursive Self-Improvement of Agent Harnesses (RRSI) is introduced, which incorporates the principles of regularizations into harness self-improvement by constraining the evolution candidate proposal and selection.

Static environments can be transformed on-the-fly to better suit agent learning, resulting in up to a 9.0-point performance boost with fewer execution steps.

Targeted optimization of normalization affine parameters can dramatically enhance low-bit quantization performance, breaking the limits of conventional training methods.

Pre-conditioning full-precision models before low-bit quantization can significantly boost performance, revealing a critical oversight in conventional PTQ approaches.

VisualClaw slashes API costs by 98% while boosting accuracy, transforming how VLMs can operate in real-time environments.

Current AI agents struggle to maintain accurate beliefs in evolving information environments, with performance varying significantly based on both model capability (15.4% range) and framework design (9.2%).

Forget hyperparameter tuning – autonomous research reveals that bug fixes and architectural tweaks unlock far greater gains in multimodal agent memory.

LLM agents can now learn on the fly and adapt to evolving user needs without disruptive downtime, thanks to a novel meta-learning framework that synthesizes new skills from failure trajectories and optimizes the base policy during inactive periods.

Even the best multimodal agents struggle with realistic visual scenarios, achieving only 27% accuracy on the new AgentVista benchmark that demands long-horizon tool use across web search, image search, and code.