arxiv arXiv cs.AI · 90d ago · research

MAST Enables Selective Unlearning in RLVR-Induced Reasoning

from English

MAST, a mechanism-guided unlearning method, achieves targeted forgetting of RLVR-induced reasoning with minimal collateral damage. On Qwen2.5-Math-1.5B and Qwen3-1.7B-Base, it significantly reduces MATH performance (45/150 to 37/15-0) while preserving GSM8K accuracy by +0.8 points and maintaining MATH retention at -0.5 points. Results hold across seeds, objectives, and models, showing superior stability over full-parameter unlearning.

Importance 3/3 New harness with differentiators arXiv cs.AI Alibaba (Qwen) Evaluation & benchmarks Reasoning models Training methods

Read original

arxiv arXiv cs.CL · 90d ago · 3 views

HydraHead: Head-Level Hybrid Attention for Long-Context Performance

HydraHead introduces a head-level hybridization of Full and Linear Attention, leveraging interpretability to select retrieval-critical heads and fuse outputs via a scale-normalized module. Trained on 15B tokens, it achieves over 69% improvement over baseline at 512K context length, outperforming layer-wise hybrids and approaching Qwen3.5's performance on long-context tasks.

arxiv arXiv cs.LG · 90d ago · 2 views

MAST Enables Selective Unlearning in RLVR-Induced Reasoning

MAST, a mechanism-guided unlearning method, achieves targeted forgetting of RLVR-induced reasoning with minimal collateral damage. On Qwen2.5-Math-1.5B and Qwen3-1.7B-Base, it significantly reduces MATH performance (45/150 to 37/15-0) while preserving GSM8K accuracy by +0.8 points and maintaining MATH retention at -0.5 points. Results hold across different seeds, objectives, and models, showing superior stability over full-parameter unlearning.

arxiv arXiv cs.CL · 91d ago · 3 views

ZPPO: Teacher in Prompts, Not Gradients

Zone of Proximal Policy Optimization (ZPPO) integrates teacher knowledge directly into prompts rather than policy gradients. It uses Binary and Negative Candidate-included Questions to surface student failure modes and amplifies learning through a prompt replay buffer, achieving superior performance on hard questions across student scales, especially at smaller model sizes.

arxiv arXiv cs.CL · 62d ago · 29 views

GSM-Plus-BN introduces perturbation-based benchmark for Bangla math reasoning

The study introduces GSM-Plus-BN, a novel perturbed Bengali mathematical dataset derived from the English GSM-Plus benchmark and verified by human translators. This resource addresses the lack of systematic benchmarks for assessing model robustness in linguistically diverse regions like Bangladesh.

arxiv arXiv cs.CL · 63d ago · 13 views

CARE-PPO enables joint learning of accurate predictions and reliable confidence in language-based quantitative tasks

Researchers introduce CARE-PPO, a reinforcement learning framework that connects loss prediction for uncertainty estimation with actor-critic PPO fine-tuning. This approach allows for the joint learning of accurate numerical estimates and reliable confidence signals in language-based quantitative prediction from unstructured inputs.