FSPO Beats GRPO and SDPO on Reasoning With Fewer Tokens via Temporal Self-Distillation
IITK · hf · 2026-10-27
A new paper proposes Function-Space Policy Optimization (FSPO), combining temporal consolidation with privileged adaptation to improve reasoning and tool-use policy optimization:
- Results: on Qwen3-8B and Olmo3-7B-Instruct, FSPO reaches avg@16 scores of 76.2 and 72.4 across five tasks (chemistry/physics/biology/materials/tool use) within a 5-hour training budget, beating GRPO and SDPO baselines (e.g., Qwen3-8B: 76.2 vs GRPO's 70.4)
- Efficiency: achieves highest accuracy with the fewest average generated tokens and reduced memory use
- Experiments ran on 4 NVIDIA GH200 GPUs following the SDPO training setup
More from Research
- Scientists argue research scaling is 80% one-pot protocols, not lab automation — clarejtbirch · 2026-10-27
- Yoav Artzi simulates 4 capped-reviewing mechanisms on real COLM data: 12-24% of accepted papers left out — yoavartzi · 2026-10-27
- Experiment finds 9 'gender neurons' in Qwen3.5-4B that flip its self-identity — burny_tech · 2026-10-27
- GPT-6-Astra embodied intelligence eval: near-100% on several of 12 benchmarks — chris_j_paxton · 2026-10-27
- LEANN open-source RAG project claims 97% vector DB storage savings: 201GB to 6GB — Promptmethus · 2026-10-27
- Claude Opus 5.5 claims ETH false, all-dimension Kakeya and Yang-Mills in one day — repligate · 2026-10-27