Yixin Cao

Research profile

Research summary

Yixin Cao is a machine-learning researcher at Fudan University studying how large language models reason, act, and improve after pretraining. The OpenAlex snapshot records 16 works in the last three years, all dated 2026, with topic clusters led by topic modeling (11), natural language processing techniques (6), explainable AI (5), adversarial robustness (4), and AI-based planning (3). Evolve has reviewed three 2026 arXiv papers spanning two evenly split subfields (0.5/0.5 between LLM agents/reinforcement learning and LLM interpretability/MoE routing). As last author, Cao's group shows enriched environment feedback (FEEs, AG-early/OE-late) lifting agent RL by an average 2.82% on SciWorld/BFCL — though Evolve notes 'consistently' is overstated and statistics reporting is thin. A second last-author study argues RLVR gains are largely recoverable by budgeted base-model search (BOPTR), an honest result whose edge over a greedy baseline sits within noise; it carries the catalog's highest reproducibility readiness. As a middle author, Cao contributed a spectral routing-effective-rank analysis uncovering a reproducible low-high-low arc in MoE reasoning cohorts, with rigorous statistics and honest nulls. Citation indices (h-index, i10-index) register 0, so output and impact remain unscored; reproducibility is mid-range at a mean of 55 over two papers, and collaboration scores 44 around 8 distinct coauthors. This profile is name-resolved from public bibliographic data.

Research keywords

Highlights

• Budgeted search recovers much of RLVR's reasoning gains

  As last author, Cao's BOPTR study argues RLVR improvements are largely recoverable by budgeted search over base-model rollouts. Evolve calls it a rigorous, honest study whose edge over a greedy baseline is within noise, and rates its reproducibility readiness high — the best in this catalog.

• Enriched environment feedback lifts agent RL by 2.82% on average

  As last author, Cao's group shows that feedback-enriched environments (FEEs, AG-early/OE-late) bootstrap self-evolving agents in long-horizon tasks, with an average 2.82% gain on SciWorld/BFCL. Evolve values the empirical study but finds 'consistently' overstated and the statistical reporting thin; reproducibility readiness is medium.

• Routing effective rank traces a low-high-low arc in MoE reasoning cohorts

  As a middle author, Cao contributed a spectral analysis defining routing effective rank in MoE reasoning cohorts, revealing a reproducible low-high-low arc. Evolve credits rigorous statistics and honest nulls while noting the gain over plain density is thin; reproducibility readiness is medium.

Timeline

1. 2026Last-author paper on environments as scaffold: FEEs bootstrap self-evolving agents in long-horizon tasks, averaging +2.82% on SciWorld/BFCL.
2. 2026Middle-author paper introduces routing effective rank, documenting a reproducible low-high-low arc across MoE reasoning cohorts.
3. 2026Last-author BOPTR paper argues RLVR reasoning gains are largely recoverable by budgeted base-model search, with the edge over greedy decoding within noise.
4. 2026By year-end snapshot, OpenAlex attributes 16 works to 2026 — the entirety of the recent three-year window — across agents, MoE routing, and post-training analysis.

Papers in the Evolve catalog

• 2026-09-08

  Environments as Scaffold: Enriching Feedback to Bootstrap Self-Evolving Agents in Long-Horizon Tasks

• 2026-09-06

  From Concentration to Differentiation and Back: Routing Effective Rank in MoE Reasoning Cohorts

• 2026-09-01

  From Base Rollouts to RL Reasoning: A Budgeted Search Perspective

Frequent co-authors