AI Coding Live News

Tackling step-level verification in mathematical reasoning and in-context copying bottlenecks in Transformers, NYU/UCLA/Stanford researchers introduce LeanPRM (arXiv:2610.17697)—using the Lean 4 theorem prover to automatically label 152K steps across 10,634 natural-language proofs with zero human annotation, boosting Qwen3-4B step accuracy on LeanPRM-Bench from 48.2% to 74.7% (vs. 60.3% for PRM800K). Meanwhile, Cornell University's Copy-Paste Language Models (CPLMs, arXiv:2610.19247) add a lightweight Copy-Paste head (~3% compute overhead) that relieves the LM head gradient bottleneck, boosting LAMBADA accuracy by +2.5-4.0 points and allowing a w=128 sliding-window CPLM to match full-attention perplexity.

Zoom

In reinforcement learning with verifiable rewards (RLVR) for reasoning models, standard PPO critics fail to track token-level value evolution along trajectories under sparse terminal feedback. LINs-lab introduced TPPO (Tracking-Value PPO), demonstrating that successive prefix values adhere to a Bayesian decomposition with additive updates in log-odds space. By parameterizing cumulative value outputs summed into value logits with terminal-only supervision, TPPO boosts aggregate mathematical reasoning by 7.0 and 8.4 points on Qwen3-4B-Base and Qwen3-8B-Base, respectively, extending robustly to scientific coding.

Zoom

While reinforcement learning produces distinct specialized expert models excels at mathematics, code synthesis, or instruction following, real-world deployment requires a single unified foundation model possessing all of these capabilities. Multi-teacher on-policy distillation (MOPD) aims to merge these capabilities by routing student-sampled prompts to domain specialists for token-level supervision. However, researchers discover an acute imbalance across Qwen3.5 architectures: instruction-following teachers exhibit broad distribution dispersion, generating gradient magnitudes that overwhelm dense mathematical signals and degrading mathematical reasoning. The team introduces Domain-Normalized MOPD (DN-MOPD, arXiv:2609.35347), which normalizes each domain's supervisory signal by its measured empirical spread. Across six public benchmarks, DN-MOPD consistently outperforms standard MOPD across all model scales, recovering mathematical reasoning performance without sacrificing instruction compliance.

Zoom

On-policy self-distillation (OPSD) trains mathematical reasoning models by deploying a privileged teacher that observes ground-truth solutions to supervise student-sampled prefixes. Standard OPSD fixes the teacher parameters across all states, leaving valuable supervision untapped. Researchers from Tencent WeChat and Tencent AI Lab propose Neighborhood OPSD (N-OPSD). They discover that local parameter perturbations in the privileged teacher reveal complementary, reference-aligned corrections across distinct token positions. Offline greedy pruning compiles a compact pool of frozen neighbor experts, while an online MaxPeak and quantile router dynamically selects the optimal expert distribution. Across AIME 2024, AIME 2025, and HMMT 2025, N-OPSD improves Average@12 by up to 2.75 points on Qwen3-1.7B, 4B, and 8B models without any inference overhead.

Zoom