Yifan Nie

Research profile

Research summary

Yifan Nie is an early-stage natural language processing researcher working on large language model (LLM) agents, evaluation, and reliability. The OpenAlex snapshot records four works in the past three years, all dated 2026, of which three arXiv preprints are reviewed in the Evolve catalog. Across them, Nie contributes as a middle author on two complementary fronts: benchmarking and system robustness. The TAM benchmark paper shows that GPT-5 near-fails long-horizon procedural reasoning over application manuals, with exact-match rates of only 1%–15.5%, though the review flags single-model coverage and an unquantified label-noise ceiling. A second study documents how synthetic-data LoRA fine-tuning induces catastrophic forgetting in skill retrieval for LLM agents, finding that regularized tuning preserves out-of-distribution behavior while low rank outweighs the regularizers. A third proposes control-data flow separation for stable prompt optimization in multi-agent LLMs, keeping output validity at 100% by construction, with modest and untested quality gains. OpenAlex topics span topic modeling (4 works), NLP techniques, multimodal machine learning, information retrieval, and multi-agent systems, with the field distribution split evenly between LLM agents and LLM evaluation. Collaboration is moderate, with 8 distinct coauthors recorded; no institution, country, or citation-impact data is yet available, so this profile should be read as an early record in progress.

Research keywords

Highlights

• TAM benchmark shows GPT-5 near-fails long procedural manuals

  Evolve's review of this 2026 preprint notes the benchmark reveals gaps in long-horizon procedural reasoning for language models, with GPT-5 achieving only 1%–15.5% exact match; the review flags single-model coverage and an unquantified label-noise ceiling as limitations.

• Synthetic-data LoRA tuning triggers forgetting in agent skill retrieval

  The study finds synthetic-data LoRA fine-tuning causes catastrophic forgetting in skill retrieval for LLM agents; regularized tuning preserves out-of-distribution behavior, but low rank outweighs the regularizers in Evolve's reading of the results.

• Control-data flow separation stabilizes multi-agent prompt optimization

  The paper freezes control channels so prompt optimization cannot corrupt multi-agent protocols, achieving 100% output validity by construction; Evolve rates its reproducibility readiness as high, while noting the quality gains are modest and untested.

Timeline

1. 2026Releases "Control-Data Flow Separation", freezing control channels so prompt optimization in multi-agent LLMs keeps 100% output validity by construction.
2. 2026Co-authors "When Synthetic Data Hurts", documenting catastrophic forgetting in skill retrieval for LLM agents under synthetic-data LoRA tuning.
3. 2026Co-authors "Tasks over Application Manuals" (the TAM benchmark), revealing that GPT-5 near-fails long manual procedures with 1%–15.5% exact match.
4. 2026The OpenAlex counts-by-year record logs 4 works for the year — the profile's entire visible output — alongside 8 distinct coauthors.

Papers in the Evolve catalog

• 2026-09-11

  Tasks over Application Manuals: Revealing Gaps in Long-Horizon Procedural Reasoning for Language Models

• 2026-09-09

  When Synthetic Data Hurts: On Catastrophic Forgetting in Skill Retrieval for LLM Agents

• 2026-09-01

  Control-Data Flow Separation: Stable Prompt Optimization in Multi-Agent LLMs

Frequent co-authors