Search

66 results for “GSM8K” · page 4 of 4

1. Research · Reinforcement learning

   Self-privileged critic improves value estimation for RLVR

   Privy to the Foil: Recasting Value Estimation with a Self-Privileged Critic for RLVR

   πPPO reuses verified same-prompt rollouts as contrastive evidence, boosting value quality and policy performance on math reasoning.

   AIME 2024: 73.5% Pass@1 (as reported)

   30 Sept 2026Peking University, Tencent

2. Research · Agents

   Contrasting multiple trajectories sharpens skill learning for LLMs

   SkillCome: Group Contrast Skill Optimization with Dual Memory

   SkillCome generates several attempts per question and compares them to identify what separates success from failure, then stores patterns across rounds in a dual memory.

   ALFWorld: 95.9% accuracy (as reported)

   30 Sept 2026Fudan University, Ant Group, Shanghai Jiao Tong University

3. Research · Agents

   Multi-agent LLM scaling benefits depend on task type and aggregation method

   Multi-agent Scaling Across Disjunctive and Compensatory Tasks

   Disjunctive tasks see accuracy gains from revision but not from larger teams, while compensatory tasks show little benefit from averaging due to shared model biases.

   GSM8K: 54.44% pass@30 (as reported)

   28 Sept 2026

4. Research · Reinforcement learning

   Online distillation cuts needless verbosity in RL-trained models

   Mitigating the Length-Scaling Tax with Online Distillation

   Length Self-Distillation reduces excess response length on easy queries without sacrificing accuracy on hard ones.

   BrowseComp-Plus: 31.65% Pass@1 (as reported)

   1 Oct 2026ByteDance Seed, Tongji University, Peking University

5. Research · Agents

   Three-stage training instills world model planning into LLM agents

   Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning

   The authors show that injecting latent predictive capability through mid-training, then structuring it via supervised fine-tuning and calibrating with reinforcement learning, enables agents to verbalize future rollouts and success estimates, outperforming reactive baselines.

   25 Jun 2026Fudan University, Shanghai Innovation Institute, Tencent Youtu Lab

6. Research · Agents

   Letting language models decide when to compact their context windows

   Self-Compacting Language Model Agents

   SelfCompact scaffold pairs a compaction tool with a lightweight rubric, enabling adaptive context summarization without fine-tuning.

   competition math (IMO-Answerbench, HMMT Nov 25/Feb 26): 18.1% accuracy improvement over no-summarization baseline (as reported)

   22 Jun 2026Johns Hopkins University, AppleCode