Notes · Method note 04

System 1 verifiers for reasoning RL

A one-pass judge that returns a probability instead of text, used as the reward inside GRPO. Prior work: one figure per paper, from Cobbe's verifier to PRIME. What we did: DiffusionGemma run as a Jev-style verifier on one A100, compared with live Jev, then seven reward arms on GSM8K. Proposal: GRPO gets no gradient from a group whose rollouts all agree; a cheap judge can rank inside those groups and pick prompts before rolling out, if it is refit as the policy moves.

Seven things the field established, one figure each: verifiers rank samples, step labels beat answer labels, generating buys accuracy at a latency cost, step rewards can be had for free, proxies get gamed, and a small verifier as reward makes traces legible.

Prior work 1 · The loop

The verifier is the bottleneck box

policy samples 8 solutions per prompt solutions verifier reads a solution, returns a reward COST · DENSITY · CALIBRATION rewards GRPO group-relative advantage updates the policy · every step · 1,100 verifier calls per step in our runs what goes in this box decides what the policy learns
Every design choice in this deck is a choice about the green box: how much it costs per call, how many numbers it returns per solution, and whether those numbers mean what they say.

GRPO: Shao et al. 2024 (DeepSeekMath, 2402.03300) · RLVR: Lambert et al. 2024 (Tülu 3, 2411.15124)

Prior work 2 · 2021

Sample many, let a verifier pick

6Bgenerator solution 1solution 2…solution 100 verifier6B, scalar headtrained on correct / incorrect 0.91 ← pickrank, keep the best 6B + verifier over 100 samples ≈ a 175B model fine-tuned on the same data
The first verifier for reasoning was a classifier with one output, used at test time to rank samples. Verification scaled better with data than fine-tuning did. Not yet a training signal.

Cobbe et al. 2021, Training Verifiers to Solve Math Word Problems (2110.14168), introduces GSM8K

Prior work 3 · 2023

Label the steps, not the answer

OUTCOME SUPERVISION · ORM step 1step 2step 3answer ✗ one label per solution PROCESS SUPERVISION · PRM step 1step 2step 3answer ✓✓✗ 800,000 human step labels · PRM800K MATH subset, best-of-N 72.4 ORM 78.2 PRM the step where it went wrong is the label the model can learn from
Process supervision beat outcome supervision by six points, and the cost was the point: 800,000 human step labels. Every PRM paper since is an attempt to get the step signal without paying for it.

Lightman et al. 2023, Let's Verify Step by Step (2305.20050) · Math-Shepherd: Wang et al. 2024 (2312.08935), automatic step labels by rollout

Prior work 4 · 2024

Let the verifier think in tokens

problem + solution“Is the answer correct?” LLM verifier “Let's check step 2: 12 × 3 = 36, but the solution wrote 38, so the final answer …” → next token: Yes / No p(Yes) = 0.07vote over K rationales GSM8K, best-of-N with the verifier 73.0 discriminative 93.4 GenRM with CoT cost: a full generation per judgement, seconds each; the score is a token probability, not a calibrated one
Making the verifier generate buys accuracy and a rationale and pays for it in latency. The whole 2025 line of reasoning reward models (RM-R1, J1, DeepSeek-GRM) is this idea with RL on the verifier's own chain of thought.

Zhang et al. 2024, Generative Verifiers (2408.15240) · RM-R1 (2505.02387) · J1 (2505.10320) · DeepSeek-GRM (2504.02495)

Prior work 5 · 2024–25

Step rewards for free, then inside the RL loop

IMPLICIT PRM · YUAN 2024 train an ORM as r(y) = β · log π(y) / π_ref(y) step rewards fall out as partial sums r(step k) = Σ log-ratios over its tokens 1/38 of the dataof Math-Shepherd, better PRM PRIME · CUI 2025 · THE SAME PRM UPDATED ONLINE, INSIDE GRPO-STYLE RL policy implicit PRMretrained on the rollouts +15.1 % avg over SFT≈ 2.5× sample efficiency, 7B, math + code
The step signal without step labels, and the first dense verifier that is retrained on-policy during RL. Two things it does not give: a calibrated probability, and a verifier that is cheaper than the policy.

Yuan et al. 2024, Free Process Rewards without Process Labels (2412.01981) · Cui et al. 2025, PRIME (2502.01456)

Prior work 6 · 2022–23

Optimise a proxy long enough and the true score falls

distance from the initial policy, √KL reward proxy reward, what the policy sees gold reward, what you wanted peak bigger RM, or an ensemble: peak moves right Gao 2022: the peak's position scales smoothly with RM size and data; Coste 2023: ensembles + a KL penalty nearly remove the drop
Any learned verifier is a proxy, and a policy trained against it will find where the proxy and the truth part ways. Our arms did this within 50 steps. The fixes on record: bigger or more verifiers, conservative aggregation, and retraining the verifier on the policy's own samples.

Gao, Schulman, Hilton 2022 (2210.10760) · Coste et al. 2023 (2310.02743) · Adv-RM: Bukharin et al. 2025 (2504.06141) · Rafailov et al. 2024 for DPO (2406.02900)

Prior work 7 · 2024

Train for a small verifier, get legibility for humans

small verifierpredicts p(correct), ~1000× smaller helpful proverreward: correct and accepted sneaky proverreward: wrong and accepted solutions attacks verifier retrained each round on both humans under time pressure: accuracy up on helpful-prover solutions, down on sneaky ones, gap widens over rounds
The closest ancestor of what we ran and of the proposal: a cheap verifier as the reward, a prover shaped toward what it can check, and a sneaky prover keeping the verifier honest. Their verifier was a small LLM; ours is a one-pass model.

Kirchner et al. 2024, Prover-Verifier Games improve legibility of LLM outputs (2407.13692) · decoupled follow-up (2602.23248) · Generative Adversarial Reasoner: Liu et al. 2025 (2512.16917)

Prior work 8 · Lineage

Where that leaves the verifier menu

latency per judgement, log scale 1 ms 10 ms 100 ms 1 s 10 s 100 s RETURNS one scalar typed probabilities text + a verdict 5–20 ms ORM / PRM classifier heads Cobbe 2021 · Lightman 2023 · Yuan 2024 (implicit PRM) 34–500 ms System 1 models Jev (TypeSafe 2026, RLCD) · DiffusionGemma readout (djev) 3 s – 329 s generative judges: GenRM 2024 · RM-R1 · J1 · DeepSeek-GRM 2025 ×100 cheaper, still a probability, no rationale
Scalar heads are fast but return one number with no calibration story; generative judges reason and explain but cost seconds per call. System 1 models sit in the gap: a typed distribution per question, trained for calibration, at the price of a classifier.

Cobbe 2021 (2110.14168) · Lightman 2023 (2305.20050) · Yuan 2024 (2412.01981) · Zhang 2024 GenRM (2408.15240) · RM-R1 (2505.02387) · J1 (2505.10320) · DeepSeek-GRM (2504.02495) · Jev latency from typesafe.ai; 34 ms measured on one A100

Numbers from 21–22 Sep 2026 runs on one A100 unless a paper is cited. Vault: System 1 Verifiers bulletin, notes 01–03.