ECHO-style SFT losses in RL dodge the elicitation-only criticism, researcher argues
stochasticchasm · x · 2026-11-23
In a discussion with YouJiacheng, stochasticchasm argues that supervised post-training losses like ECHO used during RL don't share the common "elicitation-only" criticism aimed at RL — that RL merely elicits capabilities already in the base model rather than teaching new ones. He also floats open questions about post-training older base models and whether a model with higher log-perplexity would still be good at tasks like cyber. A niche technical exchange on training methodology.
More from Research
- Base log perplexity is an extremely strong predictor of Epoch and OpenAI ECI, researchers note — stochasticchasm · 2026-11-23
- Idea: humans write high-level proofs, AI fills in lemma details in Lean — theshawwn · 2026-11-23
- Claude points to Lean blueprints, Lamport structured proofs, and Gödel's speedup theorem — theshawwn · 2026-11-23
- Two-person lab GSI claims 250+ breakthroughs across 8 fields with custom AI-built solvers, not brute-force compute — udmrzn · 2026-11-23
- Geoffrey Irving: statistical physics offers a route to obfuscated arguments in scalable oversight — AndrewCritchPhD · 2026-11-23
- Agent evals have two axes, not one: graders are the layers, repeated trials cut across all of them — bharadwajp · 2026-11-23