Human-in-the-loop or machine-executed: verifiable tasks turn agent traces into RL rewards
suragnair · x · 2026-09-25
Researcher suragnair explained on X how RL data generation works in practice:
- Coding version: give an agent a problem, it writes code, the machine executes, repeat until a final answer; if the answer is verifiable, the full trace gets a reward for post-training, improving LLM reasoning and coding.
- Human-in-the-loop version (e.g. biology): the agent proposes the next step, a human executes it, repeat until a verifiable final output (like checking experiment QC).
The upshot: verifiability is the dividing line — machine-executable domains close the RL loop automatically, while physical-world domains still need humans in the loop.
More from coding & agent
- Stream launches Agent Skills so coding agents can build chat, video and moderation apps from one prompt — amos_gyamfi · 2026-09-25
- NVIDIA Open-Sources Nemotron-Terminal: Boosts Qwen3-32B from 3.4% to 27.4% on Terminal-Bench 2.0 — _weiping · 2026-09-25
- 5 Ways to Handle Conflicting Information in AI Systems, From Source Ranking to Surfacing Disagreement — goyalshaliniuk · 2026-09-25
- TypeLLM adds type-safe, schema-guaranteed generation to LLMs without touching weights — kalyan_kpl · 2026-09-25
- kuberdenis talks harness engineering and dev workflows with alxfazio, visuals made by Opus 5.5 — tensorqt · 2026-09-25
- Five mistakes people make with agentic factories: micromanaging, no quality bar, and 'puppet shows' — hugobowne · 2026-09-25