Test-time self-distillation decoding beats beam search by 8.2 pts on MATH500 without training
hbouammar · x · 2026-11-12
Researchers propose Contrastive Beam Search (CBS), turning self-distillation into a decoding rule: a candidate's log-odds ratio under counterfactual "excellent vs wrong reasoning" contexts serves as a reward, approximated via beam search over the KL-regularized optimal policy — no training, demos, or verifier needed.
- Best or tied-best in 11 of 12 settings across MATH500, HumanEval, and GPQA
- +8.2 pts over beam search on MATH500 (Qwen3.5-0.8B, 48.0% accuracy)
- Ablation: neutral prompts do nothing; the excellent/wrong contrast adds +11.1 pts
- Paper: arXiv 2609.37041
More from Research
- ELBO explained in one diagram: how VAEs trade reconstruction against regularization — cneuralnetwork · 2026-11-12
- Acellera's AceFF-2-OMOL pushes the MLIP speed-accuracy Pareto frontier with NVIDIA-optimized TensorNet kernels — gdefabritiis · 2026-11-12
- DeepMind Podcast: AlphaGenome 2 Helps Scientists Screen Millions of DNA Changes for Rare Disease Causes — GoogleDeepMind · 2026-11-12
- Anthropic's 'Worse Together': single coordinator agent beats per-user agent teams in all 4 tests — ombre-ai · 2026-11-12
- 1,440 Synthetic Control Specs Show Huge Leeway in California $20 Minimum Wage Studies — RexDouglass · 2026-11-12
- banteg's Sokoban eval: community cuts 66-room LLM record from 3,819 to 3,063 keys in 24h — banteg · 2026-11-12