Princeton team's Coverage Principle paper explains why xent SFT underperforms for RL, proposes TailSFT
canondetortugas · x · 2026-09-12
An arXiv paper by Fan Chen, Sadhika Malladi, Dylan Foster et al. offers a theory of how pre-training enables post-training via coverage — the probability mass a model puts on high-quality responses. Coverage is necessary and sufficient for Best-of-N style post-training to work and predicts downstream performance better than cross-entropy, avoiding spurious length dependence. Follow-up work TailSFT replaces standard xent SFT with a lightweight, principled method that directly improves coverage and boosts post-RL performance.
Related event: DeepMind and Princeton Propose TailSFT to Boost RL Performance(3 posts)→
More from Research
- AI formal verification: Aristotle proves memory safety of OpenTitan root-of-trust C code — satnam6502 · 2026-10-31
- MIT CSAIL's 'Machine Studying': agents that pre-study corpora instead of searching on-the-fly — eshorten300 · 2026-10-31
- ARC Prize 2026 closes in 3 days as Yi-Chia Chen tops ARC-AGI-3 at 75.56% — GregKamradt · 2026-10-31
- NVIDIA Nemotron 3.5 scales LLM self-verification, +9-10 points on agentic benchmarks — sarahcat21 · 2026-10-31
- Unofficial OpenTargets MCP Server Brings Gene-Drug-Disease Data to LLMs — modelcontextprotocol · 2026-10-31
- Researchers find 5 personality structures in LLM weight space, Big Five-like — luke_drago_ · 2026-10-31