李晨悦

01 / Academic Research

Academic Research

Independent research, engineering explorations, and disclosure-safe updates, organized around questions, methods, and evidence.

Academic Research

Publication figure for ordered trajectory attribution: a six-action recovery DAG, action-level credit comparison, six-estimator aggregate table over forty cases, and Monte Carlo convergence across twelve seeds.
Swipe horizontally to inspect the full figure

MiniClaudeCode: Ordered Trajectory Credit & Agent Evaluation

01 / Independent project Offline estimator study
Research question
How should actions receive credit when tool trajectories contain prerequisites, interactions, and failure recovery?
Method
Constrain counterfactual replay with a precedence DAG, combining exact enumeration, uniform sampling of valid orders, and offline evaluation.
Key results
  • Under the controlled protocol, precedence-constrained estimates are closer to the prefix reference than vanilla Shapley.
  • Exact estimation, sampling convergence, and replay-validity analyses are publicly documented.
CodeProject page
Publication figure for offline recommendation: selection among five development candidates, Exact-versus-DIN results at three frozen-test cutoffs, and user-level paired differences, interval, wins, ties, and losses.
Swipe horizontally to inspect the full figure

KAI Offline RecSys Lab: Retrieval-to-Reranking Evaluation

02 / Independent project Public data · offline experiments
Research question
How can retrieval, negative sampling, and reranking be compared under the same population, data splits, and candidate sets?
Method
Freeze Two-Tower retrieval and Exact / HNSW candidates, compare DIN / DCN rerankers, and apply train-only calibration with paired user-level evaluation.
Key results
  • DIN reranking improves ordering within the same candidate sets under the frozen offline protocol.
  • Uniform negatives win this dev comparison; feature dependencies and ANN trade-offs remain visible.
CodeDemoProject page
Publication figure for multimodal data selection: Random and COINCIDE subset overlap, exact match for three paired seeds and the untuned base, plus the paired interval and improved-versus-harmed audit counts.
Swipe horizontally to inspect the full figure

Qwen2.5-VL: Data Selection Under a Small Fine-Tuning Budget

03 / Independent project Offline experiment · public summary
Research question
With the model, training budget, and held-out evaluation fixed, can data selection improve multimodal fine-tuning over random sampling?
Method
Compare Random-1K and faithful COINCIDE-1K on ScienceQA with the same Qwen2.5-VL-3B, LoRA / SFT settings, and paired seeds.
Key results
  • In the public summary, COINCIDE-1K trails Random-1K across all three paired seeds.
  • Negative results, a confidence interval crossing zero, and selection-mismatch analysis are retained.
DemoProject page

Research in progress

研究进行中｜AI Coding Benchmark 可复现性审计

检查公开数据、任务资产与评测程序在不同发布视图之间是否保持一致，并建立可追溯的离线验证流程。

Research Update · AI Coding · Agent Evaluation · Reproducibility · 2 min read

Research interests

01

Language models

Reasoning, adaptation, and the behavior of large language models.

02

Multimodal systems

Learning across language, vision, and other modalities.

03

Agents & evaluation

Tool use, reliable interaction, and evaluation beyond a single answer.