Xiaobing Tu

Research profile

Research summary

Xiaobing Tu is a researcher at Alibaba Group (China) working on the training and deployment stack for large language model agents. The Evolve-reviewed catalog contains four 2026 arXiv preprints, all with middle-author positioning and no corresponding-author role: HarnessBandit (bandit scheduling of training harnesses per GRPO step for agentic RL), TAC-Merge (curvature-graph tracing of cross-layer influence for multimodal model merging), BQ-LoRA (quotient-geometry compression of redundant agent trajectory updates under a fixed LoRA rank), and MemForest (EventTree partitioning and progressive merging for agent memory, reportedly 50% compression at 97–99.7% retention). Evolve's verdicts credit solid methods and plausible gains while flagging single small runs, missing signal ablations, absent significance tests, and no code release on two of four papers — consistent with a reproducibility mean of 38 across the four assessed papers. OpenAlex attributes 12 works over the last three years, all concentrated in 2026, with 24 distinct coauthors and recurring partners such as Xiantao Zhang, Yinggui Wang, and Jinkui Ren. The topical spread is wide — four field lines weighted 0.25 each, with recorded topics spanning GAN synthesis, few-shot learning, and robotics RL — and citation-based output and impact remain unscored.

Research keywords

Highlights

• HarnessBandit: bandit scheduling of training harnesses for agentic RL

  Jointly schedules learnability and transferability by picking the training harness per GRPO step, using gradient-sketch cosine similarity for curriculum-style scheduling. Evolve finds the gains over mixed batches plausible but grounded only in single small runs without signal ablations (medium reproducibility readiness).

• MemForest: EventTree-based agent memory management

  Partitions agent memory with EventTree and progressively merges nodes via maximum-spanning-tree token merging, reportedly achieving 50% memory compression at 97–99.7% retention with speedups. Evolve notes broad results but no significance tests, and AGPR gains that confound attribution (medium readiness).

• BQ-LoRA: quotient geometry for low-rank adaptation of LLM agents

  Introduces Behavior Quotient Learning, using a quotient manifold to balance and compress redundant agent trajectory updates under a fixed LoRA rank. Evolve rates the mathematics solid but the gains modest, with no code released (low readiness).

• TAC-Merge: cross-layer influence tracing for multimodal model merging

  Traces cross-layer multimodal influence with curvature (Ricci) graphs and tunes merge weights for LoRA adapters under trust-region control. Evolve calls the gains solid but notes that no code was released (low reproducibility readiness).

Timeline

1. 2026Publishes MemForest on EventTree-partitioned, progressively merged agent memory compression.
2. 2026Publishes BQ-LoRA on quotient-geometry low-rank adaptation for LLM agents.
3. 2026Publishes TAC-Merge on curvature-guided cross-layer influence for multimodal model merging.
4. 2026Publishes HarnessBandit, scheduling training harnesses for agentic RL with a learnability–transferability bandit.

Papers in the Evolve catalog

• 2026-09-12

  HarnessBandit: Joint Learnability-Transferability Scheduling for Multi-Harness Agentic Reinforcement Learning

• 2026-09-11

  Tracing and Coordinating Cross-Layer Influence for Multimodal Model Merging

• 2026-09-11

  Behavior Quotient Learning for Low-Rank Adaptation of LLM Agents

• 2026-09-08

  MemForest: Efficient Agent Memory Management via EventTree Partitioning and Progressive Merging

Frequent co-authors