Process Reward Models: Step-Level Verifiers for Reasoning

A process reward model scores intermediate reasoning steps instead of only terminal answers, giving denser verifier feedback for reranking and search while remaining a learned proxy.

published · difficulty 4/5 · 24 min read

Intuition

Build the mental picture first so the rest of the page has something to attach to.

PredictName the object in plain language, then predict what should change.Leave with one reusable mental picture before notation appears.

Canonical sources: Lightman et al., "Let's Verify Step by Step", the OpenAI prm800k dataset release, Uesato et al., "Solving math word problems with process- and outcome-based feedback", Cobbe et al., "Training Verifiers to Solve Math Word Problems", and Snell et al., "Scaling LLM Test-Time Compute Optimally".

An outcome reward model asks one terminal question: did the final answer look correct?

A process reward model asks a local question at every step: given the prompt and the reasoning prefix so far, is this next step valid?

That changes the shape of feedback. Instead of treating a whole solution as one indivisible completion, the verifier sees a path through intermediate states. A lucky final answer with broken reasoning can receive high outcome reward and low process reward. A trace with one early algebra mistake can be rejected before the mistake compounds.

The important caveat is that a process reward model is still a learned proxy. It can localize errors better than terminal-only feedback, but its step labels can be ambiguous, expensive, domain-limited, or wrong. PRMs do not solve reward hacking; they move the proxy from the leaf of the reasoning tree onto the edges.

Section prompt

Copy-only prompts — each action copies a page-grounded prompt to your clipboard. Nothing is sent by this site.

Math

Translate the story into symbols, assumptions, and a derivation you can inspect.

InspectTrack the same object through the notation and check each symbol.Leave with the invariant the equations preserve.

Equation 1Inspect equation objecth=(a_1,\dots,a_t),\qquad c(h,a)\in\{0,1\},\qquad P_\phi(c=1\mid h,a)=\sigma(r_\phi(h,a)).Equation 2Inspect equation objectp_i=P_\phi(c=1\mid h_i,a_i),\qquad \mathcal L_{PRM}(\phi) = - \sum_i [ c_i\log p_i + (1-c_i)\...Equation 3Inspect equation object\tau=(a_1,\dots,a_T).

Fix one prompt $x$ . In this page's binary teaching notation, a process reward model sees a reasoning prefix and a candidate next step, then predicts whether that step is locally correct. Lightman et al.'s PRM800K labels include positive, negative, and neutral ratings; here we collapse the idea into correct vs. not-correct to expose the verifier mechanics.

h=(a_1,\dots,a_t),\qquad c(h,a)\in\{0,1\},\qquad P_\phi(c=1\mid h,a)=\sigma(r_\phi(h,a)).

For labeled steps $(h_i,a_i,c_i)$ , the binary verifier can be trained with cross-entropy:

p_i=P_\phi(c=1\mid h_i,a_i),\qquad \mathcal L_{\mathrm{PRM}}(\phi) = - \sum_i \left[ c_i\log p_i + (1-c_i)\log(1-p_i) \right].

A complete solution is a terminal trace

\tau=(a_1,\dots,a_T).

Outcome supervision observes only a terminal label

z(\tau)\in\{0,1\}.

For a terminal trace, one simple additive score is the sum of step logits:

S_\phi(\tau) = \sum_{t=1}^{T} r_\phi(h_{t-1},a_t).

Other aggregations, such as mean step probability, product of step probabilities, or minimum step score, are design choices rather than a universal PRM definition. The key definition is local verification of steps.

Separate from Lightman et al.'s best-of-N evaluation, this page also shows a finite KL-style RLHF bridge. A KL-regularized trajectory update is

\pi_\beta(\tau\mid x) = \frac{ \pi_{\mathrm{ref}}(\tau\mid x)\exp(S_\phi(\tau)/\beta) }{ \sum_{\tau'} \pi_{\mathrm{ref}}(\tau'\mid x)\exp(S_\phi(\tau')/\beta) }.

This is the same finite-action probability-shaping pattern as RLHF, but the learned reward decomposes over reasoning steps. Smaller $\beta$ pushes harder on the verifier score. If the step verifier is wrong, optimization can still select verifier error.

Section prompt

Copy-only prompts — each action copies a page-grounded prompt to your clipboard. Nothing is sent by this site.

Code

Keep the implementation aligned with the notation so the algorithm is legible.

TraceMatch variables to symbols before reading the implementation.Leave with a runnable witness for the math.

Code witness 1import math import numpy as np def logit(p): p = min(max(p, 1e-6), 1 - 1e-6) return math.log(...python

This witness uses four fixed traces for the equation $2(x+3)=14$ . The outcome verifier prefers a trace with the correct final answer but invalid reasoning. The accurate process verifier prefers the clean trace. If we inject a false-positive process error, the process verifier can be fooled too.

import math
import numpy as np

def logit(p):
    p = min(max(p, 1e-6), 1 - 1e-6)
    return math.log(p / (1 - p))

def policy(scores, ref, beta=0.7):
    logits = np.log(ref) + np.asarray(scores) / beta
    logits = logits - logits.max()
    weights = np.exp(logits)
    return weights / weights.sum()

traces = ["clean", "lucky_final", "slip", "wrong"]
ref = np.array([0.30, 0.22, 0.28, 0.20])
outcome_p = np.array([0.86, 0.92, 0.16, 0.12])
step_p = [
    [0.93, 0.91],
    [0.16, 0.24],
    [0.93, 0.22],
    [0.20, 0.35],
]

def process_scores(step_probs):
    return [sum(logit(p) for p in probs) for probs in step_probs]

outcome_winner = traces[int(policy([logit(p) for p in outcome_p], ref).argmax())]
process_winner = traces[int(policy(process_scores(step_p), ref).argmax())]

hacked_steps = [list(probs) for probs in step_p]
hacked_steps[1] = [0.98, 0.97]
hacked_winner = traces[int(policy(process_scores(hacked_steps), ref).argmax())]

print(outcome_winner, process_winner, hacked_winner)
assert outcome_winner == "lucky_final"
assert process_winner == "clean"
assert hacked_winner == "lucky_final"

Section prompt

Copy-only prompts — each action copies a page-grounded prompt to your clipboard. Nothing is sent by this site.

Interactive Demo

Use direct manipulation to connect the explanation to a moving system.

ManipulateChange one control and predict the visible response before reveal.Leave with the observed invariant or a repaired model.

Live Concept Demo

Explore Process Reward Models: Step-Level Verifiers for Reasoning

The stage is code-native and interactive. Use it to test the explanation against the mechanism.

difficulty 4/5undergraduatecode-aligned

Demo inquiry checkpoint

Manipulate one control and predict the visible change.

01Choose lensTrace a quantity

02ObserveDemo state pending

03GroundName the equation, invariant, or control that explains it.

04CarryNext: Test-Time Compute: Spending Inference Budget on Search

Choose what to inspect in Process Reward Models: Step-Level Verifiers for Reasoning. This shared fallback is an observation guide, not evidence of learning.

Loading interactive demo...

Use the demo as a diagnostic instrument. Switch between outcome-only scoring and process scoring, change the process aggregation, inject verifier errors, and lower $\beta$ to increase optimization pressure. Watch how probability mass moves across complete traces, and whether the selected trace has a correct final answer, valid local steps, or merely a high verifier score.

Section prompt

Copy-only prompts — each action copies a page-grounded prompt to your clipboard. Nothing is sent by this site.

Object

4/4 sections ready

Concept: Process Reward Models: Step-Level Verifiers for Reasoning

What is the smallest example that makes Process Reward Models: Step-Level Verifiers for Reasoning click without losing the math?

BeforeRLHF: Reward Modeling + KL-Regularized Policy OptimizationNow4/4 sections readyTryManipulate one control and predict the visible change.NextTest-Time Compute: Spending Inference Budget on Search

Object contextAlignment

StateAlignment

1 source attachedChecking local snapshot

concept:alignment/process-reward-models

Math3 equation objectsh=(a_1,\dots,a_t),\qquad c(h,a)\in\{0,1\},\qquad P_\phi(c=1\mid h,a)=\sigma(r_\phi(h,a)).CodeProcess Reward Models: Step-Level Verifiers for Reasoning code witness 1assert outcome_winner == "lucky_final"DemoProcess Reward Models: Step-Level Verifiers for Reasoning predictionManipulate one control and predict the visible change.

ConceptLearner lens

Process Reward Models: Step-Level Verifiers for Reasoning

What is the smallest example that makes Process Reward Models: Step-Level Verifiers for Reasoning click without losing the math?

Mode questionCan I say the mechanism back in one sentence before I reveal anything?

Start with the prediction checkpoint, then compare the reveal to the mental model.

Take this move

Study modes

Keep the object fixed; change the lens.

01PredictCommit before reveal.02ExplainName the invariant.03Ground1 source link nearby.04ExperimentChange one knob.05CarrySave the next question.

Route back through the notebook

Carry the same object through intuition, math, code, and demo.

4/4 sections ready

Carry inRLHF: Reward Modeling + KL-Regularized Policy Optimization

Bring the mental model from RLHF: Reward Modeling + KL-Regularized Policy Optimization; this page will reuse it instead of restarting from zero.

Work hereProcess Reward Models: Step-Level Verifiers for Reasoning

A process reward model scores intermediate reasoning steps instead of only terminal answers, giving denser verifier feedback for reranking and search while remaining a learned proxy.

Carry outTest-Time Compute: Spending Inference Budget on Search

The next edge should feel earned: use the demo prediction here before following Test-Time Compute: Spending Inference Budget on Search.

After The First Pass

Turn the concept into an inspected object.

The lower panels are one second act: keep the object fixed, inspect it visually, check source boundaries, practice transfer, then attach the research question.

ConceptProcess Reward Models: Step-Level Verifiers for ReasoningAlignment

01MechanismInspect motionSee how the idea changes form before adding more explanation.02VisualChoose a cueBind one visible mark to one mathematical role.03SourcesCheck boundary1 source link attached to the object.04ClaimAudit evidenceTreat the page claim as something to verify.05PracticeAttempt transferWrite before feedback so the model has something to repair.06RoomAsk groundedAttach the question to the same exact object.

Mechanism Storyboard

See the idea move before the page explains it

A process reward model scores intermediate reasoning steps instead of only terminal answers, giving denser verifier feedback for reranking and search while remaining a learned proxy.

Demo notes open01 / Intuition

Editorial alignment illustration of a branching reasoning trace with step-level verifier scores and a fading invalid path.

Prediction lens

Start with the picture, metaphor, or geometric mechanism.

Commit first

Before reading further, choose the kind of change Process Reward Models: Step-Level Verifiers for Reasoning should make visible.

Visual Inquiry

Make the image answer a mathematical question

A process reward model scores intermediate reasoning steps instead of only terminal answers, giving denser verifier feedback for reranking and search while remaining a learned proxy.

4/4 stages readyDemo notes connected

Visual cueWhich visible object should carry the first intuition?

Inspection depth2/4

Prediction

Which visible object should carry the first intuition?

Commit first

Pick the cue that should make Process Reward Models: Step-Level Verifiers for Reasoning easier to reason about before the page gives the answer.

Source Grounding

Canonical references for the mechanism on this page.

Object - ConceptProcess Reward Models: Step-Level Verifiers for ReasoningQuestion

What is the smallest example that makes Process Reward Models: Step-Level Verifiers for Reasoning click without losing the math?

concept:alignment/process-reward-models

Boundary

sources: lightman-2023-verify-step-by-step

Check

Open the closest source note before trusting the local explanation.

Evidence

1 selected-object source shown first; 1 reference total.

Next move

Audit the claim boundary, then ask from the same selected object.

selected object source · paper · 2023Let's Verify Step by StepLightman et al.

Located CF editorial boundary

Grounds process supervision and PRM800K as step-level feedback for mathematical reasoning traces.

Used here as

Lightman et al. contrast ORMs using final results with PRMs receiving feedback for each step, train PRMs to predict step correctness, reduce step scores for solution ranking, and evaluate...

Caveat

Binary Bernoulli/BCE is a teaching reduction; Lightman use positive/negative/neutral labels and token log-likelihood. Excludes exact aggregation, calibration, KL/RL generator training, un...

Open source

Claim Review

A process reward model scores intermediate reasoning steps instead of only terminal answers, giving denser verifier feedback for reranking and search while remaining a learned proxy.

Object - ConceptProcess Reward Models: Step-Level Verifiers for ReasoningQuestion

What is the smallest example that makes Process Reward Models: Step-Level Verifiers for Reasoning click without losing the math?

concept:alignment/process-reward-models

Boundary

sources: lightman-2023-verify-step-by-step

Check

Treat every claim as provisional until source support and a local witness agree.

Evidence

1 structured claim check on this concept.

Next move

Run the prediction or practice transfer before asking for a grounded review.

1 CF editorial source-scope review recorded

Publisher-side editorial review is not independent replication. Claims without it still need exact source-support review. 1 reference and 3 local witnesses are available for inspection.

Process reward models are step-level verifiers rather than only final-answer judges: for prefix h and next step a, a binary formulation estimates P_phi(c=1|h,a)=sigma(r_phi(h,a)), enabling reranking or search while remaining a learned proxy.

Registered CF editorial boundary

Let's Verify Step by StepGrounds process supervision and PRM800K as step-level feedback for mathematical reasoning traces.

Used here as

Lightman et al. contrast ORMs using final results with PRMs receiving feedback for each step, train PRMs to predict step correctness, reduce step scores for solution ranking, and evaluate by best-of-N search...

Local witness

source-span-lightman-2023-verify-step-by-step math-object-1 math-object-2 code-witness-1 interactive-demo

Equation 1

h=(a_1,\dots,a_t),\qquad c(h,a)\in\{0,1\},\qquad P_\phi(c=1\mid h,a)=\sigma(r_\phi(h,a)).

Equation 2

p_i=P_\phi(c=1\mid h_i,a_i),\qquad \mathcal L_{\mathrm{PRM}}(\phi) = - \sum_i \left[ c_i\log p_i + (1-c_i)\log(1-p_i) \right].

Caveat

Binary Bernoulli/BCE is a teaching reduction; Lightman use positive/negative/neutral labels and token log-likelihood. Excludes exact aggregation, calibration, KL/RL generator training, universal gains, cheap...

Review stateCF editorial source-scope reviewClaim metadata: source checkedPublisher-side editorial review only; not independent replication. Check caveats and exact source scope.

Lightman et al. support PRMs as process-supervised reward models trained on step-level labels and evaluated by best-of-N selection of the highest-ranked generated solution. The page's binary Bernoulli/BCE notation is a didactic collapse of their positive/negative/neutral token-likelihood setup. PRM reliability/proxy limits are supported by their reward-model mistake/false-positive examples. KL reweighting is outside this reviewed claim.

Reviewer: codex+oracle; reviewed 2026-05-07

Practice notebook

Use the idea, then test it somewhere new

A process reward model scores intermediate reasoning steps instead of only terminal answers, giving denser verifier feedback for reranking and search while remaining a learned proxy.

AttemptNo learning claim inferred

Object - ConceptProcess Reward Models: Step-Level Verifiers for ReasoningQuestion

What is the smallest example that makes Process Reward Models: Step-Level Verifiers for Reasoning click without losing the math?

concept:alignment/process-reward-models

Boundary

sources: lightman-2023-verify-step-by-step

Check

Use one state from Process Reward Models: Step-Level Verifiers for Reasoning to explain what changes, why it changes, and which assumption the explanation needs.

Evidence

No learner move yet; no learning state is inferred.

Next move

Write first, use only the help you need, then try a new case without it.

Explain

Use one state from Process Reward Models: Step-Level Verifiers for Reasoning to explain what changes, why it changes, and which assumption the explanation needs.

Hint 1

Reveal when your model needs a nudge.

Hint 2

Reveal when your model needs a nudge.

Hint 3

Reveal when your model needs a nudge.

Your answer canvas

Attempt context

Learner move

Write an attempt before asking the companion.

Assistance used

0 of 3 progressive hints opened.

Evidence boundary

This draft and any AI response do not establish mastery; a later unassisted case can.

Misconception check

Grounded object roomClose

Selected object routeAsk from this object; carry one invariant back.sources: lightman-2023-verify-step-by-step

ObjectConceptProcess Reward Models: Step-Level Verifiers for Reasoning
PredictBefore revealProcess Reward Models: Step-Level Verifiers for Reasoning prediction
WitnessCompare codeProcess Reward Models: Step-Level Verifiers for Reasoning code witness 1
RoomAsk groundedChecking local snapshot

ConceptProcess Reward Models: Step-Level Verifiers for ReasoningAlignment

Code witness comparisonProcess Reward Models: Step-Level Verifiers for Reasoning code witness 1assert outcome_winner == "lucky_final"Prediction before revealProcess Reward Models: Step-Level Verifiers for Reasoning predictionManipulate one control and predict the visible change.

Grounded room questionWhat is the smallest example that makes Process Reward Models: Step-Level Verifiers for Reasoning click without losing the math?Checking local snapshot

Research Room

Attach the question to an exact object

Pick the concept, equation, source, code witness, claim, misconception, or demo state before asking for help. The handoff stays grounded to that object.

Next local actionNo local draft saved yet

Open the draft below to save one note and next action in this browser.

conceptAlignment

Process Reward Models: Step-Level Verifiers for Reasoning

Anchored question

What is the smallest example that makes Process Reward Models: Step-Level Verifiers for Reasoning click without losing the math?

Source boundaryInspect source ids: lightman-2023-verify-step-by-stepStable content-object key attached

Role lenses for this object

These are fixed, deterministic perspectives derived from the selected object. They do not represent people, community contributions, or independent review.

Learner evidence requestAsk what would make "Process Reward Models: Step-Level Verifiers for Reasoning" feel predictable rather than familiar.

Assumption

Source ids lightman-2023-verify-step-by-step must support the exact object, not just the surrounding topic.

Source-checking summary

Treat this as a mechanism object: connect the definition to one equation, code witness, or demo before broadening the discussion.

Proposed experiment

Ask the learner to perturb one representation, then check whether the same invariant survives in math, code, and demo.

Next action

The learner can state the mechanism in their own words

Evidence4 checks

PredictionChecking carried observation

ActionReady for one action

AILearner handoff ready

Open source object

01PredictionChecking browser-local route memory

02EvidenceChecking for a carried observation

03BoundaryInspect source ids: lightman-2023-verify-step-by-step

04Next moveSave one next action

Local action draftNo local draft saved yetExpand only when ready to capture one local next action

Local action draft

This draft stays locally in this browser for concept:alignment/process-reward-models.

Draft noteNext action

No local draft saved.

Evidence to inspect

Source ids to inspect: lightman-2023-verify-step-by-step
Definition, prerequisite, and contrast concept links
The equation or code witness that makes the concept operational
One demo state that shows the invariant instead of a slogan

What would resolve this

The learner can state the mechanism in their own words
The learner can name the prerequisite that would repair confusion
The learner can predict how the mechanism changes under one perturbation

Object-attached AI handoff

I am working in Continuous Function's research reading room. Object: concept - Process Reward Models: Step-Level Verifiers for Reasoning Object key: concept:alignment/process-reward-models Context: Alignment Anchor id: concept/concept-notebook/alignment/process-reward-models Open question: What is the smallest example that makes Process Reward Models: Step-Level Verifiers for Reasoning click without losing the math? Evidence to inspect: - Source ids to inspect: lightman-2023-verify-step-by-step - Definition, prerequisite, and contrast concept links - The equation or code witness that makes the concept operational - One demo state that shows the invariant instead of a slogan Deterministic role lenses for this object: - Boundary: fixed perspectives, not people, community contributions, or independent review - Source-checking summary: Treat this as a mechanism object: connect the definition to one equation, code witness, or demo before broadening the discussion. - Proposed experiment: Ask the learner to perturb one representation, then check whether the same invariant survives in math, code, and demo. - Teach/transfer move: Turn the mechanism into one sentence that predicts a neighboring concept. - Assumptions: - Source ids lightman-2023-verify-step-by-step must support the exact object, not just the surrounding topic. - The stable content-object key lets local drafts, prompts, and route memory attach without changing the source page. - The concept explanation is local atlas prose until checked against its math, code, and source support. - Prerequisite gaps should become a repair route, not a reason to leave the object vague. - Role-lens requests: - Learner: ask for "Ask what would make "Process Reward Models: Step-Level Verifiers for Reasoning" feel predictable rather than familiar." | assumption: Source ids lightman-2023-verify-step-by-step must support the exact object, not just the surrounding topic. | next action: The learner can state the mechanism in their own words - Researcher: ask for "Source ids to inspect: lightman-2023-verify-step-by-step" | assumption: The stable content-object key lets local drafts, prompts, and route memory attach without changing the source page. | next action: The learner can name the prerequisite that would repair confusion - Experimenter: ask for "Choose one variable or condition to perturb before asking for an explanation." | assumption: The concept explanation is local atlas prose until checked against its math, code, and source support. | next action: The learner can predict how the mechanism changes under one perturbation - Professor: ask for "Find the smallest transferable rule a learner could reuse without the AI." | assumption: Prerequisite gaps should become a repair route, not a reason to leave the object vague. | next action: Teach or transfer: Turn the mechanism into one sentence that predicts a neighboring concept. What would resolve this: - The learner can state the mechanism in their own words - The learner can name the prerequisite that would repair confusion - The learner can predict how the mechanism changes under one perturbation Answer as a careful research tutor: stay source-grounded, separate verified evidence from assumptions, name the relevant math objects, and end with one next action. Current deterministic role lens for this object: - Role lens: Learner - Evidence request: Ask what would make "Process Reward Models: Step-Level Verifiers for Reasoning" feel predictable rather than familiar. - Assumption to keep visible: Source ids lightman-2023-verify-step-by-step must support the exact object, not just the surrounding topic. - Proposed experiment: Ask the learner to perturb one representation, then check whether the same invariant survives in math, code, and demo. - Next action: The learner can state the mechanism in their own words

concept/concept-notebook/alignment/process-reward-models
concept:alignment/process-reward-models

Learning Map

Before / Now / Try / Next

BeforeRLHF: Reward Modeling + KL-Regularized Policy Optimization

NowIntuition → Math → Code → Demo

TryManipulate one control and predict the visible change.

NextTest-Time Compute: Spending Inference Budget on Search

Learnerpicture first Researchersource check Experimentertest the demo Professorname invariant

01Predict 02Ground 03Code 04Carry

01Intuitionready
02Mathready
03Codeready
04Interactive Demoready

Object Companion

Ask beside the selected object

A process reward model scores intermediate reasoning steps instead of only terminal answers, giving denser verifier feedback for reranking and search while remaining a learned proxy.

Your question

GoalComfortStyleStuck on

Context prompt

You are my AI learning companion for Continuous Function. Current context: Alignment concept. Learning surface: Process Reward Models: Step-Level Verifiers for Reasoning. What this page says: A process reward model scores intermediate reasoning steps instead of only terminal answers, giving denser verifier feedback for reranking and search while remaining a learned proxy. Current section: Intuition, math, code, and interactive demo. Suggested next step: Manipulate one control and predict the visible change.. Learner goal: Understand the idea. Learner comfort level: New to this. Preferred explanation style: Visual first. Task: Explain the central idea in plain language, then restate it with the exact math objects from the page. Answer in a way that helps me learn: ask one clarifying question only if needed, use intuition before notation, and end with one thing I should try on the page.

Copy-only tool: this panel composes a prompt for an AI tool you already use. Nothing you enter here is sent from this page.

Domain

Alignment

alignmentprocess-supervisionreward-modelingreasoningverifiers

Prerequisites

RLHF: Reward Modeling + KL-Regularized Policy Optimization Reward Hacking: Overoptimizing Preference Proxies Cross-Entropy KL Divergence (Relative Entropy)

Leads To

Test-Time Compute: Spending Inference Budget on Search Tree Search Reasoning: Allocating Inference Budget Across Prefixes

Direct Preference Optimization Kahneman-Tversky Optimization instruction-tuning

Within this domain