Zhijiang Guo

Agentic RL agents can learn to efficiently manage context and subtasks across multiple sessions by estimating the marginal contribution of each session, outperforming naive advantage broadcasting.

A strong LLM can review grounded drafts from a weak model, retaining most task performance while halving latency and memory traffic.

Across reasoning, long-context understanding, and agentic tasks, SAS consistently outperforms trainable sparse attention baselines across attention budgets, with especially large gains under tight budgets, demonstrating more effective context ranking for downstream tasks.

Ockhamareto achieves a staggering 49.9% mutation score while using 44% fewer tests than the best existing method, revolutionizing unit-test generation efficiency.

MemoryCPT achieves a superior cost-performance trade-off for LLM agents by intelligently managing memory without overwhelming downstream models with context.

FormalRx transforms opaque autoformalization evaluations into clear, actionable insights, enabling targeted improvements in formal reasoning systems.

Learning high-level strategies can boost vulnerability reproduction success rates by over 20%, revolutionizing how we approach software security tasks.

The current RL checkpoint outperforms larger LLMs in redesigning training environments, revealing that iterative learning enhances diagnostic capabilities.

A single pair of boundary tokens transforms hidden-state reasoning into a trainable and interpretable framework, revealing causal insights that were previously obscured.

CoT fine-tuning can slash long-range recall by over 57% in hybrid LLMs, but a simple parameter restoration method can reverse this trend without additional training.

ReaLB achieves 1.29x faster multimodal MoE inference by dynamically adjusting expert precision, proving that real-time adaptation can overcome modality-induced load imbalances.

LLMs that ace code generation often fail to grasp intended program semantics, as evidenced by a stark performance decline when generating executable behavioral specifications on the new CodeSpecBench benchmark.