training · stratify-rollouts-confidence-outcome-forming-rl-groups TEC-0213 · proposed

Stratify rollouts by confidence and outcome when forming RL

Within each RL step, bucket the step's own rollouts by length-normalized surprisal and reward into anchors, hard negatives and low-confidence successes, and assemble half the optimization groups from a fixed ratio of these buckets before cl

Proposed. Filed by the drafting stage because a paper introduces or tests it. Nobody has vouched for it. The claim drafted from that paper is its first evidence, and what the paper used to validate it is in that claim’s source note. Say whether it held up for you, below.

Within each RL step, bucket the step's own rollouts by length-normalized surprisal and reward into anchors, hard negatives and low-confidence successes, and assemble half the optimization groups from a fixed ratio of these buckets before cl Filed by the drafting stage from Stable-MM-R1: Anchoring Multimodal Reasoning Dynamics via Entropy-Guided Stratification (2609.07148) with status proposed: the paper introduces or tests it, and the claim drafted from the paper is its first evidence. The paper validated it by: Ablation on MathVista/MathVision/MMMU Pass@8: GRPO 60.7 avg, DAPO 62.0, +HSR alone 64.7, +PAQM alone 63.8, both 65.8; plus path-entropy curves staying ~0.35-0.60 nats versus baseline collapse. Baselines reported at peak pre-collapse. One run, no seed variance.. Nobody has vouched for it; add support or contest it on this page.

Addresses: Retaining skills and safety after fine-tuning

Does it work?

nothing measured0 supporting · 0 contesting sources

Efficacy claims — what this technique actually moves, under which conditions, and whether that has been contested.

No efficacy claim filed yet. The technique is catalogued; whether it moves the capability, and when, is a separate assertion that needs its own sources.

No search recorded either, so this says nothing about the literature — only that nobody has looked here yet.

Code

No repository linked yet. Contribute one.

Sources

Either way, bring evidence: a link, or what you saw.

Whether a technique works is a claim. What you file here becomes a claim about this technique under the capability it addresses, with your evidence as its source, and moves the standing above the same way a paper would.