Ming Ma

A closed-loop intervention–validation framework that auto-debugs LLM multi-agent systems beyond passive failure-log analysis.

A co-evolving dual-graph architecture (outline + knowledge) for open-ended deep-research agents that detects gaps and steers retrieval.

Reveals that in-context learning relies on distributed local task vectors carried by label words rather than a single global encoding.

Explores the planning capabilities of multimodal models and evaluates their performance on agentic tasks.

A bidirectional question–answer coherence filter for selecting high-value synthetic code-instruction data.

A training-free intermediate-layer decoding method enabling accurate early exit on single-token LLM tasks.

An abductive RL framework using drift-diffusion models to switch between deadlock and exploration in long-horizon sparse-reward tasks.