ATRIUMsearch → argument graph
Article · 2026-07-08 · 6 moments

[AINews] Lilian Weng summarizes 35 papers on Harness Engineering for RSI

a quiet day lets us read some condensed insight ✦ AI generated

02
Claim

Even after many harness improvements are eventually internalized into the core model, the need to specify goals and context does not disappear.

Weng argues that no matter how much harness functionality gets absorbed into future base models, external goal- and context-specification will remain necessary, a point AINews calls a notable concession from a neolab cofounder.

transcript

Lilian Weng: Even when many harness improvement[s] get eventually internalized into core model, the need to specify goals and context will not disappear.

03
Data

FTPO (Final Token Preference Optimization) training, branded Antidoom, sharply cuts doom-loop rates in small reasoning models: LFM2.5-2.6B falls from 10.2% to 1.4% and Qwen3.5-4B falls from 22.9% to 1% under greedy sampling, alongside downstream eval gains.

Liquid AI's open-source Antidoom method relabels the token that triggers repetitive 'doom loops' and redistributes probability toward alternatives, yielding large measured drops in loop rates for LFM2.5-2.6B and Qwen3.5-4B.

transcript

Liquid AI: The reported reductions are substantial: LFM2.5-2.6B from 10.2% → 1.4% and Qwen3.5-4B from 22.9% → 1% under greedy sampling, with downstream eval gains. The method, FTPO (Final Token Preference Optimization), relabels the loop-triggering token and redistributes probability toward alternatives.

04
Prediction

It is hard to forecast how much the future of recursive self-improvement will rely on harnesses, but harness engineering will likely evolve toward self-improvement and enable auto-research, producing smarter systems in turn.

Lilian Weng frames RSI's future as likely driven by evolving harnesses (the scaffolding around models) rather than models rewriting their own weights, with harnesses potentially enabling auto-research.

transcript

Lilian Weng: It is hard to forecast how much the future of RSI will rely on harnesses. Likely harness engineering will evolve in the direction of self-improvement and enable auto-research, and, in turn, smarter

extends · 1

05
Mechanism

Muse Image and Video's self-refinement behavior emerged naturally during reinforcement learning rather than being hand-scripted, and their output quality improves as test-time compute is scaled up.

Meta Superintelligence Labs describes Muse Image/Video's agentic loop (planning, search, tool use, code execution, self-refinement) as a capability that arose emergently from RL rather than manual engineering, with quality scaling with test-time compute.

transcript

Meta Superintelligence Labs: The notable technical angle is not just image quality, but an explicitly agentic generation loop: planning, web search, tool use, code execution, and self-refinement before rendering. Meta also says performance improves with scaled test-time compute, and that self-refinement behavior emerged during RL rather than being hand-scripted in this follow-up.

Highlight slides
Related episodes