Junlan Feng

Research profile

Research summary

Junlan Feng's Evolve catalog centers on four 2026 arXiv papers that cut across four subfields in equal measure — speech synthesis (text-to-speech), LLM agents, efficient reasoning for LLMs, and NLP/AI safety (LLM alignment) — each accounting for 25% of the profiled field distribution. The work spans COT-TTS, an audio context-aware text-to-speech framework with chain-of-thought reasoning built on a 9M-sample pipeline; A*-Thought-V2, which compresses latent reasoning via geometric dynamics of LLM hidden states; EdgeMem, an LLM-free agent memory built on an evidence-preserving multi-anchor hypergraph; and a study of preventative steering in adversarial fine-tuning proposing PIS. Across 4 evaluated papers the reproducibility mean is 66/100, with EdgeMem rated high and the others medium. The collaboration score is 72/100 with 36 distinct coauthors, including frequent collaborators such as Fanyu Meng, Zeyang Cui, and Jiannong Cao. The focus score is 0/100, reflecting an even four-way split across fields rather than a concentrated agenda. Output, impact, and activity scores are unavailable from the current data; no citation or h-index figures are quoted here. This profile is name-resolved from public bibliographic data; on all four papers the recorded position is middle or last, with no corresponding-author role.

Research keywords

Highlights

• Audio context-aware chain-of-thought TTS with a 9M-sample pipeline

  Evolve's review describes a new context-aware CoT TTS task built on a 9M-sample pipeline with an 800-sample benchmark, where 0.6B/1.7B models rival 34-39B cascades on MOS; the reviewers note that the CoT benefit itself remains untested.

• Latent reasoning compressed via geometric dynamics of LLMs

  A*-Thought-V2 claims geometric latent chain-of-thought compression with gains of up to 2.6pts accuracy at roughly half the length; the review finds the direction promising but notes the absence of trained-latent baselines and significance tests.

• LLM-free agent memory via evidence-preserving multi-anchor hypergraph

  Evolve's review reports that this LLM-free hypergraph memory matches or beats generative baselines at far lower cost, with strong, well-controlled evidence — the highest reproducibility readiness in this catalog — though the headline gap lacks a confidence interval.

• Active adaptation in preventative steering during adversarial fine-tuning

  The paper claims preventative steering protects through active adaptation rather than static offsets and proposes PIS; the review credits a solid analysis while flagging single-run, judge-only evaluation.

Timeline

1. 2026COT-TTS: audio context-aware text-to-speech with chain-of-thought reasoning (arXiv preprint).
2. 2026Active Adaptation, Not Static Defense: temporal dynamics of preventative steering in adversarial fine-tuning, proposing PIS (arXiv preprint).
3. 2026A*-Thought-V2: efficient latent reasoning via geometric dynamics of LLMs (arXiv preprint).
4. 2026EdgeMem: LLM-free agent memory construction and retrieval via evidence-preserving multi-anchor hypergraph (arXiv preprint).

Papers in the Evolve catalog

• 2026-09-19

  COT-TTS: Audio Context-Aware Text-to-Speech with Chain-of-Thought Reasoning

• 2026-09-09

  Active Adaptation, Not Static Defense: Temporal Dynamics of Preventative Steering in Adversarial Fine-Tuning

• 2026-09-07

  A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM

• 2026-09-03

  EdgeMem: LLM-Free Agent Memory Construction and Retrieval via Evidence-Preserving Multi-Anchor Hypergraph

Frequent co-authors