Search

48 results for “GSM8K” · page 3 of 3

1. Research · Safety & security

   Pretrained LLMs already contain refusal circuits that shift during alignment

   SafeEvo: Deciphering the Safety Alignment Mechanism and Evolution in Language Models

   SafeEvo identifies sparse refusal circuits in base models, traces their structural drift across safety checkpoints, and uses them to improve fine-tuning.

   HarmBench: 2.29 pts average ASR reduction (as reported)

   9 Oct 2026The University of Hong Kong, Chinese Academy of Sciences, Information Engineering University

2. Research · Language models

   Adaptive token computation narrows diffusion language models’ quality gap

   ALoDLM: Adaptively Looped Diffusion Language Models

   ALoDLM recurrently refines difficult token representations while committing easier predictions early, improving benchmark scores without giving up parallel decoding.

   GSM8K: 133.5 GFLOPs GFLOPs per generated token (as reported)

   8 Oct 2026University of Illinois Chicago, Amazon AGI, Korea University

3. Research · Reinforcement learning

   Residual advantage uses teacher-student disagreement as bounded step reward

   Residual Advantage: Student-Relative Teacher Guidance for RL with Verifiable Rewards

   Chen and Pang show that centering the teacher's probability residual within each response allows step-level credit redistribution while preserving verifier outcome labels, improving reasoning performance across multiple benchmarks.

   Macro: 44.75% Pass@8 (as reported)

   10 Oct 2026Harbin Engineering University, Tencent, Harbin Institute of Technology

4. Research · Efficiency & systems

   Canonical training improves language models under joint low-bit compression

   Few Bits, One Law: Toward W2A4KV2

   CanonQ uses fixed transforms and reusable codebooks, then jointly trains models to tolerate quantized weights, activations and attention caches.

   GSM8K: 47.8% final-answer exact match (as reported)

   8 Oct 2026Meta AI

5. Research · Language models

   Supervision reliability trumps alignment coverage in cross-tokenizer distillation

   Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability

   Strict token alignment already captures most learning signal; adding span-level supervision reduces accuracy due to weakly aligned gradients.

   ALFWorld: 30.34% success rate (as reported)

   9 Oct 2026Alibaba Cloud Computing

6. Research · Reinforcement learning

   Lexicographic distillation preserves priority-ordered capabilities in multi-objective RL

   Lexicographic Multi-Objective On-Policy Distillation

   Method routes student rollouts to highest-priority unmet specialist and projects corrections to avoid interfering with higher-priority experts.

   AIME 2024, AIME 2025, HMMT February 2025: 28% retained gain, composite conciseness (as reported)

   7 Oct 2026Cohere, Mila, Université de Montréal

7. Research · Reinforcement learning

   Outcome-guided self-distillation improves LLM reasoning by adapting supervision

   Outcome-Guided On-Policy Self-Distillation

   The authors propose OG-OPSD, which selects forward or reverse KL divergence based on answer correctness and shortens distillation on incorrect trajectories using teacher entropy, outperforming vanilla on-policy self-distillation.

   AIME 2024: 78.9% Avg@12 (as reported)

   6 Oct 2026Southeast University, Huawei Noah’s Ark Lab, Zhongguancun Academy

8. Research · Efficiency & systems

   Compressed looped models settle, not drift, so precise final loops recover them

   A Tilted Bowl Is Not a Slippery Slope: Compressing Looped Models

   Kolawole et al. show that rounding error in looped models shifts the fixed point rather than accumulating, enabling predictable failure and recovery with few high-precision steps.

   GSM8K: 45% accuracy (as reported)

   6 Oct 2026Carnegie Mellon University, ML Collective