@thunderous777

Undergraduate at SJTU ACM Honors Class

thunderous77.github.io/
Joined October 2023
  • Tweets17
  • Following63
  • Followers24
5Media
Xuanchang Zhang retweeted
Bingxuan Li
@bingxuan_lOct 13
What does it take to turn VLMs for cost-efficient zero-shot robot manipulation, without tools, without task-specific training, without external modules like VLA-as-tool or SAM3 or IK Solver ? 🤖 Introducing MotorMind ! 💡 Zero-shot Manipulation: No Fine-tuning or Coding agent 💡 VLM-only: No external perception models, No tools, No VLA-as-tool, No IK-Solver 💡 Local Deployable: Qwen3.8-Next-Flash as backbone model for cost-efficient local deployment 💡 Strong Generalization: 83.3% success on LIBERO-PRO 💡 Real-Robot Transfer: 95% success on real-robot tasks without embodiment-specific training Check out more: 📄 Paper: arxiv.org/pdf/2609.38078 🌐 Project: motor-mind.github.io/ 💻 Code: github.com/Motor-Mind/MotorM…
6
29
3
99
11,057
Xuanchang Zhang retweeted
Xuanchang Zhang@thunderous777Oct 16
1/7 Can an LLM learn from an older rollout model without importance weights? Introducing Gibbs-Gap Policy Optimization (GGPO). Sync rollouts every 8 -> 128 optimizer updates: accuracy stays at 46.58% -> 46.39% across five math benchmarks (Qwen3-1.7B-Base).
GGPO accuracy stays near 46.5 percent as the rollout synchronization interval grows from 8 to 128 optimizer updates. Other tested methods decline more at 128 updates.

ALT GGPO accuracy stays near 46.5 percent as the rollout synchronization interval grows from 8 to 128 optimizer updates. Other tested methods decline more at 128 updates.

4
13
1
20
1,145
Xuanchang Zhang retweeted
WANG Ruida@RickyRDWangOct 16
🚀 Introducing Lean4AgenticRL framework. The first method to apply formal verification in an open-ended domain for agentic RL. We apply the Lean4Agent(xitter.cf/RickyRDWang/status/206…) framework to provide process rewards for agentic RL. We train LAR-9B/27B using Lean4AgenticRL. The LAR-9B models achieve a 20.2% improvement on SWE-Bench-Verified, bringing the score to 64.2%. The LAR-27B model achieves 77.3% on TerminalBench-2.1, improving the base model by 7.7%. The LAR-9B/27B outperforms all the known baselines of similar size in SWE-Bench-Verified, TB-2.0, and TB-2.1. Welcome to read our paper and try our model! [1/n] Paper: github.com/RickySkywalker/Le… Model: huggingface.co/amazon/LAR-27… and huggingface.co/amazon/LAR-9B Code: github.com/RickySkywalker/Le… Website: lean4agent.github.io/
WANG Ruida@RickyRDWangJun 9
🚀 Introducing the Lean4Agent framework. The first framework 🎉 that unifies the formal verification of agent workflow and executions. It views the agent workflow as “a computer program in natural language.” 🤖 The Lean4Agent provides FormalAgentLib that treats agent workflows like programs. We develop a predicate system that elegantly constrains the agent's execution step via preconditions and postconditions. ✅ This framework makes agent workflows’ self-consistency verifiable before execution, and enables the detection of which step of the workflow breaks the consistency for fine-grained fixes. 🔁 When a verified workflow fails on a specific problem, the predicate system, combined with the LeanEvolve proposed in the Lean4Agent framework. We can dynamically update the workflow to enhance its performance. [1/n] Paper: arxiv.org/abs/2606.06523 Huggingface: huggingface.co/papers/2606.0… GitHub Repo: github.com/RickySkywalker/Le…
2
17
4
31
118,411
1/7 Can an LLM learn from an older rollout model without importance weights? Introducing Gibbs-Gap Policy Optimization (GGPO). Sync rollouts every 8 -> 128 optimizer updates: accuracy stays at 46.58% -> 46.39% across five math benchmarks (Qwen3-1.7B-Base).
GGPO accuracy stays near 46.5 percent as the rollout synchronization interval grows from 8 to 128 optimizer updates. Other tested methods decline more at 128 updates.

ALT GGPO accuracy stays near 46.5 percent as the rollout synchronization interval grows from 8 to 128 optimizer updates. Other tested methods decline more at 128 updates.

4
13
1
20
1,145
more replies
6/7 No fresh rollouts at all? With every response generated by the initial Qwen3-1.7B-Base model, GGPO's MATH-500 accuracy rises from 32.96% to a peak of 59.79% at update 240. It later declines to 48.31% at update 400.
Fixed-pool offline learning on MATH-500. GGPO starts at 32.96 percent, peaks at 59.79 percent at 240 updates, then declines to 48.31 percent at 400 updates.

ALT Fixed-pool offline learning on MATH-500. GGPO starts at 32.96 percent, peaks at 59.79 percent at 240 updates, then declines to 48.31 percent at 400 updates.

1
1
2
77
7/7 Paper, code, and a visual walkthrough: Blog: thunderous77.github.io/GGPO/ Paper: arxiv.org/abs/2610.20814 Code: github.com/thunderous77/GGPO The blog develops the primal-dual intuition, the Best-of-K view, and connections to AGRO and MaxRL.
2
3
105
Xuanchang Zhang retweeted
Jack Bai@jackbot_csOct 14
1/11 Can a single-token importance weight preserve the target of full-sequence correction? With Future-KL, we prove that both updates share the same unique stationary policy. We also contribute the promising empirical results we observed. arxiv.org/abs/2610.18738
3
23
1
114
6,661
Xuanchang Zhang retweeted
Chenlu Ye@jobmarket@ye_chenluFeb 19
1/5 Happy CNY🎊 Still bothered by RL off-policy instability in LLM? Introducing a new way💡Adaptive Layerwise Perturbation (ALP)💡, a simple but robust fix that outperforms GRPO/MIS/Bypass, achieves better stability (KL, entropy) and exploration! 🔗 Blog: beneficial-curiosity-d98.not…
4
27
5
145
25,390
Excited to present my work at #EMNLP2024 in Riverfront Hall on Nov 13 at 10:30 a.m.! 🎉✨ I’ll be introducing a novel, gold label-agnostic approach for prompt evaluation and optimization 🔍💡. Through self-consistency evaluation and mutual-consistency refinement, we achieve accurate, unsupervised prompt assessments. Can’t wait to share our insights! #NLP #PromptLearning
1
1
7
245
Load more