Guorun Yao

Research profile

Research summary

Guorun Yao is a machine-learning researcher whose Evolve-reviewed output concentrates on one fast-moving problem: deciding what a long-horizon LLM agent should keep in context. Five 2026 arXiv papers — ICLR-based online reasoning pruning, GEM, StateComp, PaMER, and DRSR — propose complementary mechanisms: entropy-driven chain-of-thought pruning; geometry-aware compression cutting 21.4% of tokens at equal reward; learned compression timing cutting 52.27% at near-flat reward; pre-action memory-control signals with AUROC 0.831/0.765 and roughly 73% token reduction; and counterfactually supervised set-level deletion reporting +0.103 reward at -20.82% tokens. Evolve rates four papers medium and one low for reproducibility, repeatedly noting single-run evaluations, missing error bars, unpaired baselines, and model-generated labels. The OpenAlex record, resolved under ORCID 0000-0003-1762-0794, lists h-index 3 and i10-index 3, 21 distinct coauthors over three years, 4 works in the recent three-year window, and by-year counts of 1, 1, 3, and 4 for 2021, 2022, 2023, and 2026. Earlier topics — topic modeling, GAN image synthesis, anomaly detection, and multimodal ML — indicate a broader pre-2026 record, while every catalogued paper is middle-authored with no corresponding role. No institution or country is recorded in the snapshot.

Research keywords

Highlights

• StateComp: 52.27% token cut at near-flat reward

  StateComp learns when to compress agent history, reporting a 52.27% token reduction while reward stays near flat; Evolve judges the approach promising but notes single-run evaluations lack statistical rigor.

• Memory control signals emerge before action (AUROC .831/.765)

  The PaMER line of work reports that hidden states encode memory-control needs before the agent acts, with AUROC of .831 and .765, cutting tokens roughly 73% at near-par scores; Evolve notes LLM-made labels and thin statistics temper the claim.

• GEM trims 21.4% of agent tokens at equal reward

  The GEM paper reports cutting long-horizon agent tokens by 21.4% at equal reward, supported by a strong geometry-versus-evidence analysis of representation geometry and effective rank; Evolve flags that the 13-baseline comparison is unpaired and reward parity untested.

• DRSR learns set-level deletion risk via counterfactual supervision

  DRSR learns set-level deletion risk with counterfactual supervision and teacher-forced likelihood, reporting +0.103 reward at -20.82% tokens; Evolve rates reproducibility low because results lack error bars.

• Online reasoning pruning with ICLR for agent contexts

  This paper prunes low-entropy agent reasoning online with an ICLR criterion, cutting tokens while lifting reward on long-horizon tasks; Evolve notes the gains hinge on one domain and that a delete-all baseline nearly matches, with thin statistics.

Timeline

1. 2021OpenAlex records 1 work this year; no papers from this period appear in the Evolve-reviewed catalog.
2. 20221 work recorded in the bibliometric snapshot; still outside the reviewed catalog.
3. 20233 works recorded — the busiest year in the snapshot before 2026.
4. 2026All five Evolve-reviewed papers appear on arXiv, each on context compression or memory control for long-horizon LLM agents (ICLR pruning, GEM, StateComp, PaMER, DRSR), matching the count of 4 works in the recent three-year window.

Papers in the Evolve catalog

• 2026-09-24

  When Can Agents Forget Their Reasoning? ICLR for Long-Horizon Agent Context Compression

• 2026-09-23

  Stable Geometry with Divergent Task Evidence for Efficient Long-Horizon Agent Compression

• 2026-09-23

  StateComp: Learning When to Compress History in Long Horizon Agents

• 2026-09-23

  Memory Control Signals Emerge Before Action in Long Horizon Agents

• 2026-09-23

  DRSR: Learning Set-Level Deletion Risk for Efficient Long-Horizon Agents

Frequent co-authors