/p/2026-08-21 · explainer
Paper explainer · 2608.18324 · Salas

Let the verifier pick the training targets.

Qwen3-14B on PlanBench replanning gets 1 of 80 plans accepted by VAL out of the box, 30 of 80 in thinking mode at ~66 seconds per plan, and 57 of 80 after 72 LoRA optimizer steps on its own thinking-mode attempts that VAL accepted — running in the cheap non-thinking mode at 1.2 seconds per plan. The ablation is where the paper earns its title: on 160 fresh cases with an identical candidate pool, model self-selection admits 69 plans, verifier selection admits 102, gap significant at p=2×10⁻⁶.

01 · The record

Five fields per accepted attempt — that is the whole training set

The paper's title is a shape, not a jargon. A governance record is a row you write every time an external verifier admits an attempt: what was asked, what the model produced, what the verifier said, which output was accepted, and where that output came from. Nothing about the model has to change to collect them — you probably have most of these fields already, if any part of your pipeline runs a verifier for safety, format, or business rules. Step through the fields.

Interactive · one record from the Qwen self-distillation runPlanBench replanning · VAL verifier · schema shown verbatim

      

The paper is careful about the gate field: it records not just that the verifier accepted the plan, but which of two acceptance gates it passed — semantic-only, or a stricter interface-cure gate that also checks token-level compatibility. The Qwen arm's interface-cure gate did not clear significance (p=0.25), and the paper reports that as a null rather than dropping it.

02 · The primary result

1 → 57 out of 80, at 1/56th the latency of the thinking-mode teacher

Qwen3-14B has two inference modes: cheap non-thinking (0.6 seconds per plan) and expensive thinking (66 seconds per plan). Base non-thinking gets 1 plan accepted by VAL out of 80. Base thinking gets 30. Sample thinking-mode attempts, keep the ones VAL accepts, run 72 optimizer steps of LoRA on the non-thinking mode, and the adapter gets 57 of 80 — at 1.2 seconds per plan.

Interactive · step through the three modesQwen3-14B · 80 held-out PlanBench replanning cases · VAL-acceptance is the metric
0/8040/8080/80

Schema validity moves the same way — 71/80 base non-thinking, 77/80 base thinking, 80/80 after training. The training does not just cover the semantic hole; it also stabilises the surface format the verifier accepts as a plan in the first place.

03 · The ablation

Same candidate pool, three ways to pick — the verifier picks the ones that work

The load-bearing experiment holds the candidate pool constant and varies only who selects the training targets. Same base model, same LoRA hyperparameters, same 24 targets, same 160 held-out cases. Only the selection method changes.

Chart · target-selection ablation on 160 fresh casesQwen3-14B · matched candidate pool · same LoRA recipe · non-thinking inference
0/16080/160160/160
69 / 160
model self-selection
the model grades its own attempts and trains on the ones it liked
102 / 160
verifier selection
the external checker admits attempts and only those become training targets

Pairwise difference +33 cases, p = 1.96 × 10⁻⁶. Schema-only selection (train on anything that parses as a plan) gets 55/160 — the parse gate alone helps a little; the verifier gate helps a lot.

The reading is not "verifiers are better than models at grading answers"; that would be the same claim you can make from any strong-verifier setting. The reading is that when the model grades its own attempts to pick supervision, the training signal it selects is drawn from the same distribution as the model, and the training tightens what the model already believes rather than what the verifier will accept. The verifier is the only party in the loop whose judgement is not entangled with the model's.

04 · The loop

Three self-distillation rounds, most of the gain in the first, saturation by the third

The training loop is small enough to spell out: thinking mode generates candidates for the current problem set, VAL admits or rejects each one, the adapter trains on the admitted ones for 72 optimizer steps, and the adapter's non-thinking mode is measured against the held-out set. Drag the round slider; watch acceptance climb and each successive round contribute less.

Interactive · self-distillation roundsQwen Generator A · 72 optimizer steps per round · one A100-80GB, BF16
VAL-accepted plans / 80
round cost

The stopping rule is not a token budget; it is that successive rounds add fewer than four accepted plans. Generator B stopped after two rounds. Generator C saturated in one round on its fixed 128-case measurement. The training loop is finding acceptance rather than a target scale.

05 · Your verifier

What this changes for a product that already runs a checker illustrative

Any pipeline that already runs a format checker, a business-rule validator, a code-compilation step, or a moderation gate is producing governance records for free — attempt in, verdict out. Pick how many requests a day pass your verifier, how often the model reaches the verified state on its cheap path, and what one thinking-mode call costs you. The paper's own gain (base non-thinking → adapter) is held per row.

Interactive · scale it to your trafficacceptance lift is the paper's · request volume, verifier hit rate and costs are yours
extra verifier-accepted requests per day
at the paper's base→adapter lift for the cheap non-thinking path
daily inference cost avoided versus 'thinking on every request'
at ~1/56th the thinking-mode per-plan cost the paper reports

The paper's own numbers are on 80 held-out PlanBench cases with a specific verifier and a specific model; the strength of the lift on your own pipeline will not be the paper's 56/80. What is transferable is the structural claim: the training signal that moves the model closer to what the verifier will accept is the one the verifier picks, not the one the model picks.

The numbers

What the paper actually reports

What this does not show

In practice