RLVR bounds on Qwen3.5-4B land within 8–13% of training accuracy
ddkang · x · 2026-07-21
New RLVR generalization bounds on Qwen3.5-4B stay within 8–13% of training accuracy
Bridgewater AIA Labs, UIUC, and MIT report what they describe as the first non-vacuous generalization bounds for reasoning LLMs on real-world tasks.
- The framework targets RLVR and uses a compression-based PAC-Bayes approach.
- Because RLVR training is based on stochastic decoding, they first reparameterize decoding with the Gumbel-max trick.
- To make the bound non-vacuous, they introduce Progressive RLVR with:
- on-policy distillation
- TinyLoRA for extreme parameter efficiency
- quantization
In experiments on Qwen3.5-4B across Math, Code, General Knowledge, and Text-to-SQL, the bounds are:
- 8–13% away from the model’s training accuracy
- 17–51% above the base model’s accuracy
Their ablation shows each ingredient matters: removing distillation or training directly with TinyLoRA makes the bound looser, and replacing TinyLoRA with standard LoRA makes the bounds vacuous or meaningless on the smaller model.
More from Research
- arXiv paper "Agent Meltdowns": overly helpful agents turn benign errors into unsafe actions — alex_verem · 2026-10-27
- Cornell paper: 64.7% of agents take unsafe actions after hitting ordinary errors — alex_verem · 2026-10-27
- LightMTP: lightweight latent multi-token prediction boosts LLM planning in pretraining — gdm3000 · 2026-10-27
- EMNLP study: AI literary translation reads fine but varies more, even within one book — tuvllms · 2026-10-27
- Lab monkeys at $30-80K each: a 4-year breeding cycle is the real preclinical drug bottleneck — zebird0 · 2026-10-27
- OpenAI Foundation funds UNC tumor antigen project, hiring senior experimental scientist — iskander · 2026-10-27