Lei Li · page 2
Papers
- ALGO: Synthesizing Algorithmic Programs with LLM-Generated Oracle Verifiers (2023 · 19 citations)
LLMs handle code generation well when a task spells out exactly what to implement. They do much worse on competitive-programming problems, where the model must first choose an efficient algorithm, and their output still needs human checking. ALGO tackles both issues. It builds on an observation: writing a slow… - The Role of Deductive and Inductive Reasoning in Large Language Models (2025 · 17 citations)
Prompting methods such as Tree of Thought improve LLM reasoning by exploring many output paths, but this is expensive, and fixed prompt templates adapt poorly to harder instances. - Could Thinking Multilingually Empower LLM Reasoning? (2025 · 10 citations)
LLMs are usually assumed to reason best in English. Recent evidence shows that some non-English languages can outperform English on reasoning benchmarks. This paper asks how much accuracy could be gained by having a model reason over the same question in many languages and pooling the results. - Mending the Holes: Mitigating Reward Hacking in Reinforcement Learning for Multilingual Translation (2026 · 2 citations)
LLM translation quality drops sharply for low-resource languages, and the usual post-training fixes require parallel or preference data that these languages often lack. - CodeEvo: Interaction-Driven Synthesis of Code-centric Data through Hybrid and Iterative Feedback (2026 · 1 citation)
Training code-generation LLMs requires large numbers of instruction–code pairs. Manual curation is expensive, and popular automatic methods such as Self-Instruct and Evol-Instruct depend on fixed prompting heuristics. As a result, they often produce vague tasks and code that fails to compile or pass its tests.… - How Chain-of-Thought Works? Tracing Information Flow from Decoding, Projection, and Activation (2026 · 1 citation)
Chain-of-Thought (CoT) prompting reliably improves LLM reasoning, but earlier explanations of why it works were mostly intuitive and not tied to model internals. - A Practical Evaluation Method for Long-Form Simultaneous Speech-to-Speech Translation (2026)
The paper addresses the lack of a reproducible way to evaluate simultaneous speech-to-speech translation (SimulS2ST) on long, unsegmented speech. - DIMAS-OMOP: A Deliberative Intelligence-Based Multi-Agent System for Chinese Medical Text Standardization toward OMOP (2026)
Converting free-text Chinese radiology/imaging reports into the OMOP CDM v5.4 is difficult. Rule-based tools break on abbreviations and phrasing variants, BERT-style extractors need costly expert annotation, and LLMs give inconsistent outputs or require cloud APIs that conflict with patient-privacy rules. The authors… - Hierarchical Policy Optimization for Simultaneous Translation of Unbounded Speech (2026)
Simultaneous speech translation (SST) systems built on LLMs can handle unbounded, hours-long audio efficiently when framed as a multi-turn dialogue that reuses the KV cache. - When “Correct” Is Not Safe: Can We Trust Functionally Correct Patches Generated by Code Agents? (2026)
Code agents are usually judged by whether their patches resolve an issue and pass the test suite. This paper argues that passing tests does not mean a patch is secure. The authors define a threat class called functionally correct yet vulnerable (FCV) patches: fixes that pass every test but still contain a…
This page in Markdown → https://md.research.elamur.ai/free/en/authors/lei-li/page/2/