I work on LLM reasoning, reinforcement learning, and practical AI systems. My current research focus is how reasoning models use prompt echoes, attention refocusing, probabilistic anchors, and inference-time reminders during multi-step reasoning.
ICLR 2026 paper and codebase on LLM reasoning. The project studies Echo of Prompt behavior: when a model repeats or rephrases the original user question inside its reasoning trajectory. The work asks whether prompt echoes are only supervised fine-tuning artifacts, or whether they can act as anchors for attention refocusing and better reasoning.
- GitHub: https://github.com/hhh2210/echoes-as-anchors
- OpenReview: https://openreview.net/forum?id=vndn1Wrult
- Project page: https://hhh2210.github.io/projects/echoes-as-anchors/
- Keywords: LLM reasoning, chain-of-thought, Echo of Prompt, echoic prompting, attention refocusing, probabilistic analysis, reasoning probes.
- LLM reasoning and evaluation: mechanisms, prompting, probes, and analysis tools for reasoning trajectories.
- Agent and developer tooling: practical systems around AI coding, token-efficient command output, local model workflows, and Codex/Claude-style agent environments.
- Applied AI products: research ideas that can survive product constraints, user growth, and real usage.
- echoes-as-anchors - code for the ICLR 2026 paper on Echo of Prompt, probabilistic costs, and attention refocusing in LLM reasoning.
- rtk - a Rust CLI proxy that reduces LLM agent token consumption by filtering noisy command output before it enters model context.
- auto-skill - a few-shot skill induction prototype for turning examples into reusable AI-agent skills.
- CodexBar - a macOS menu bar utility for OpenAI Codex and Claude Code usage visibility.
- Date Match - zero-budget cold start of a Gen Z relationship-matching product, reaching 100K users in the first 10 days and 170K+ completed tests through organic sharing.
I am interested in LLM reasoning, reinforcement learning, agent systems, evaluation, and applied AI products. Reach out for research collaboration, prototype building, or practical AI system work.