Criteria-Grounded Verification

• Criteria-grounded verification is a framework that uses explicit, external criteria to validate claims with transparent reasoning traces, moving beyond simple execution success.
• It structures evidence through decomposed, modular verification chains and symbolic checks, enabling rigorous, multi-level evaluation across diverse applications.
• Empirical findings indicate that criteria-grounded methods improve accuracy and auditability, yielding enhanced performance in multimodal, clinical, and safety-critical environments.

Criteria-grounded verification is a family of verification procedures in which a model’s judgment is tied to explicit acceptance criteria and externally inspectable evidence rather than to free-form plausibility, execution success, or model self-report. Across recent work, the criteria may be encoded as support relations between a claim and evidence, structured rubrics, expert guidelines, symbolic rules, reconstructed physics contracts, or deterministic safety predicates; the grounding may come from documents, images, trajectories, knowledge bases, retrieved guidelines, or environment state. The common objective is to make verification inspectable, auditable, and intervention-sensitive: a decision should change when the evidence or criterion changes, and the basis for that change should be visible (Rao et al., 2 Apr 2026, Arasteh et al., 10 Apr 2026, Feng et al., 2 Jul 2026).

1. Conceptual definition and scope

At its most explicit, criteria-grounded verification reformulates a task from bare classification into verification with an exposed trace. ThinknCheck extends standard grounded claim verification from a classifier $$\mathcal{M}: \mathcal{C}\times\mathcal{D}\rightarrow\{0,1\}$$ to a reasoning-aware verifier $$\mathcal{R}: \mathcal{C}\times\mathcal{D}\rightarrow\mathcal{T}\times\{0,1\}$$, where $\mathcal{T}$ is a reasoning trace. Case-grounded evidence verification defines a binary support variable $\ell(c,e,y)$ over local case context, external evidence, and a structured claim. Vera encodes a safety case as $\sigma=\langle g,s_0,V_g\rangle$, where verification is a deterministic predicate over execution artifacts (Rao et al., 2 Apr 2026, Arasteh et al., 10 Apr 2026, Feng et al., 2 Jul 2026).

The criteria themselves vary by domain, but they are always externalized. In caption verification, the criterion is grounded semantic consistency between a sentence and an image, together with explanation and localization of the mismatch. In prescription auditing, the criteria are explicit pharmaceutical rules such as dosage thresholds, contraindication conditions, interactions, and special-population restrictions. In scientific simulation, the criterion is whether generated code encodes the intended PDE contract rather than merely whether it runs. In education, the criterion is alignment with authorised syllabus outcomes, prescribed verbs, glossary definitions, performance band descriptors, and marking-guideline principles (Gao et al., 25 Jun 2026, Zhu et al., 11 Mar 2026, Song et al., 10 May 2026, Xu et al., 16 Jun 2026).

This scope makes the notion broader than claim checking alone. The same design pattern appears in multimodal preference judgment, web-agent trajectory verification, action selection for robots, legal and safety rule synthesis, clinical consultation, formal verification, and named entity recognition. A plausible implication is that the phrase identifies a verification architecture rather than a single task family.

2. Core mechanisms for grounding and inspection

A recurrent mechanism is decomposition into smaller verifiable units before final judgment. Bonsai recursively decomposes a root claim into an “incomplete entailment tree,” retrieves top-$k$ evidence for leaf sub-claims, assigns likelihood scores through anchoring and adjustment, and propagates probabilities compositionally through the tree. MJ1 enforces a five-stage chain $$(p,R_A,R_B)\xrightarrow{g_O}O\xrightarrow{g_C}C\xrightarrow{g_V}V\xrightarrow{g_E}E\xrightarrow{g_s}s$$, separating observations, claim extraction, verification, evaluation, and scoring. VerifiNER verifies span factuality, then type factuality, and only then contextual relevance, explicitly requiring span verification to precede type verification (Sanders et al., 4 Apr 2025, Kumar et al., 9 Mar 2026, Kim et al., 2024).

Another mechanism is evidence structuring rather than raw evidence concatenation. Case-grounded evidence verification organizes concept-specific evidence pools into supportive evidence, hard non-support, and easy non-support, making the support relation itself label-defining. MIND compresses dialogue history into a fact-only clinical retrieval state and retrieves criteria-grounded supports from a Psychiatric Reasoning Bank, with each support note recording known facts, critical gaps or exclusion checks, and the rationale for the next inquiry. GLEAN retrieves disease-relevant clinical guidelines, produces step-wise guideline-alignment signals, aggregates them into surrogate features, and accumulates them across the trajectory before calibration (Arasteh et al., 10 Apr 2026, Li et al., 4 Mar 2026, Zhang et al., 3 Mar 2026).

A third mechanism is symbolic or formal admissibility checking. The neuro-symbolic causal rule-synthesis framework routes candidate rules through a Rule Verification Engine that performs syntax and schema validation, logical consistency analysis, and safety and invariant verification before commitment to the knowledge base. Knowledge-graph-driven assertion synthesis in formal verification links requirements, RTL structure, properties, counterexamples, and coverage reports in a verification-centric KG, so that each generated property is traceable to both design criteria and tool outcomes (Rehan et al., 30 Apr 2026, Viswambharan et al., 7 May 2026).

These mechanisms share an operational objective: the verifier should expose not only a verdict but also the criterion path, evidence path, and failure mode. This is why many systems emit tags, structured traces, query plans, or localized evidence rather than a scalar score alone.

3. Representative forms across domains

The same verification logic is instantiated with different criterion forms and evidence substrates.

System	Criterion form	Grounding source
ThinknCheck	YES/NO entailment after a short structured rationale	Document for a claim
Bonsai	Probabilistic sub-claim tree	Retrieved evidence bank
MJ1	Criterion-based multimodal score after verification chain	Prompt and response images
Universal Verifier	Process and outcome rubric	Full screenshot trajectory
PDE-grounded intent verification	Intent Fidelity Score over physics contract	Reconstructed MOOSE semantics
PharmGraph-Auditor	Chain of verification over pharmaceutical rules	Hybrid pharmaceutical KB
Vera	Deterministic safety predicate	Environment state and tool traces

In text-centric verification, the verifier often checks whether evidence supports a proposition for a specific instance. ThinknCheck uses short structured rationales before binary entailment labels; Bonsai turns a claim into a tree of verifiable sub-claims; case-grounded verification makes the support relation itself the supervision target; VerifiNER starts from an existing NER prediction and revises it using KB-grounded span and type checks (Rao et al., 2 Apr 2026, Sanders et al., 4 Apr 2025, Arasteh et al., 10 Apr 2026, Kim et al., 2024).

In multimodal settings, the criterion is usually tied to grounding failure. MJ1 is designed around the claim that multimodal judges can rely on language priors or positional bias unless forced to verify textual claims against visual observations. GAVEL requires three simultaneous outputs for inconsistent image-text pairs: verification, natural-language explanation, and bounding-box localization. VeriSpace verifies candidate robot actions by spatial validity and goal progress using 3D-aware scene encoding and spatially grounded action reasoning (Kumar et al., 9 Mar 2026, Gao et al., 25 Jun 2026, Zhao et al., 9 Jun 2026).

In agentic and high-stakes environments, the criterion is often formalized as an outcome predicate or rubric. The Universal Verifier separates process reward from outcome reward for computer-use agents and uses a divide-and-conquer relevance pipeline over all screenshots. Vera defines executable safety cases whose predicates are grounded first in environment state, then in tool-call evidence, and only lastly in agent responses. Agentic Rubrics for SWE agents derive repository-grounded checklists before candidate patches are scored execution-free against explicit criteria (Rosset et al., 5 Apr 2026, Feng et al., 2 Jul 2026, Raghavendra et al., 7 Jan 2026).

In clinical, scientific, and formal domains, the grounding object is often a structured external knowledge source. CXReasonAgent reasons only from tool-produced diagnostic and visual evidence extracted from chest X-rays. PharmGraph-Auditor decomposes auditing into SQL and Cypher sub-queries over a Hybrid Pharmaceutical Knowledge Base. PDE-grounded intent verification reconstructs the encoded PDE from MOOSE objects and compares it against an intended contract with the Intent Fidelity Score. KG-based formal verification grounds assertion synthesis in structured IRs and tool feedback (Lee et al., 26 Feb 2026, Zhu et al., 11 Mar 2026, Song et al., 10 May 2026, Viswambharan et al., 7 May 2026).

4. Empirical findings

Several papers report that explicit grounding and structured verification materially improve performance over direct judging. On LLMAggreFact, ThinknCheck attains 78.1 balanced accuracy, surpassing MiniCheck-7B at 77.4 with 7x fewer parameters; removing the reasoning step reduces balanced accuracy to 57.5. On SciFact, ThinknCheck reaches 64.7 balanced accuracy, a +14.7 absolute gain over MiniCheck-7B, and the domain-specialized ThinknCheck-Science reaches 61.0% accuracy on GSMClaims while also improving LLMAggreFact and SciFact (Rao et al., 2 Apr 2026).

In multimodal judgment, grounding constraints also help before and after training. MJ1 reports that its grounded verification mechanism alone improves an untrained base model on MMRB2 by +3.8 points on Image Editing and +1.7 on Multimodal Reasoning, and after training the judge reaches 77.0% accuracy on MMRB2 with only 3B active parameters. GAVEL shows that strong closed-source models still struggle, with GPT-5 obtaining the best [email protected] of 34.1 and the best explanation score of 70, while a supervised VisionLLM-based baseline improves localization from 15.1/12.1/9.1 to 42.6/30.2/18.4 at [email protected]/[email protected]/[email protected] (Kumar et al., 9 Mar 2026, Gao et al., 25 Jun 2026).

Evidence-sensitive supervision produces unusually strong intervention tests. In radiology, the full case+evidence+claim verifier reaches AUROC 97.43, AUPRC 87.78, and F1 78.72 under correct evidence, but performance collapses under swapped evidence to AUROC 55.62, AUPRC 21.73, and F1 29.77; similar collapse appears on the external CheXpert-Plus case distribution. In agentic clinical diagnosis, GLEAN surpasses the best baseline by 12% in AUROC and 50% in Brier score reduction, and active verification further improves AUROC and Brier across the three diseases studied (Arasteh et al., 10 Apr 2026, Zhang et al., 3 Mar 2026).

For long-horizon agent verification, rubric- and evidence-grounded designs substantially reduce verifier failure. The Universal Verifier reports false positive rates of 1% on internal outcome labels and 8% on Browserbase OM2W outcome labels, compared with 22% and 40% for WebJudge and 45% and 60% for WebVoyager, while achieving agreement with humans about as high as human-human agreement. Vera, evaluated on four production agent frameworks, reports average attack execution success rates of 90.6% under single-channel attacks and 93.9% under multi-channel attacks, indicating that its executable safety cases are broadly realizable across heterogeneous systems (Rosset et al., 5 Apr 2026, Feng et al., 2 Jul 2026).

5. Distinctions from adjacent approaches and recurrent misconceptions

Criteria-grounded verification is not equivalent to adding a chain-of-thought prompt. ThinknCheck reports that zero-shot chain-of-thought on the base Gemma3-1B harms accuracy relative to direct answers, and that preference optimization with a simple format+accuracy reward underperforms supervised reasoning. The paper’s rationale-length study further reports an inverted-U relation in which mid-length rationales work best, while very short and very long chains hurt accuracy (Rao et al., 2 Apr 2026).

It is also not equivalent to merely appending retrieved evidence. Case-grounded evidence verification explicitly argues that evidence grounding is meaningful only if the model’s decision changes when the evidence changes, formalized by the intervention principle $f_\phi(c,e^\star,y)>f_\phi(c,\tilde e,y)$ for perturbed evidence. This is why wrong-state and topic-related negatives are central to the supervision design rather than incidental augmentations (Arasteh et al., 10 Apr 2026).

A further misconception is that direct context-conditioned judging is sufficient if the model is strong enough. The SGV paper identifies agreement bias in multimodal LLM verifiers: they tend to agree with the shown trajectory and rationalize flawed behavior. Self-Grounded Verification addresses this by separating prior retrieval from evaluation, yielding gains of up to 20 points in accuracy and failure detection rates and improving online supervision across web navigation, computer use, and robotic manipulation (Andrade et al., 15 Jul 2025).

Execution success is likewise not a substitute for criteria-grounded correctness. In MOOSE code generation, execution-only repair improves runtime success while leaving 39.1% to 40.0% of all 220 cases runnable but still solving the wrong physics across the three main deployment-audit models; “FalseExec” is defined as passing execution with $\mathrm{IFS}<0.85$. Vera makes an analogous distinction in safety evaluation by prioritizing environment state over tool-call evidence and responses, so that a textual refusal after an unsafe action does not erase a realized violation (Song et al., 10 May 2026, Feng et al., 2 Jul 2026).

6. Limitations, failure modes, and open problems

The surveyed systems are explicit that grounding is not a complete guarantee of correctness. Case-grounded evidence verification uses oracle-style evidence from curated pools and does not solve full retrieval; performance also degrades under evidence-source shift, especially in AUPRC and calibration, and remains sensitive to backbone choice. GAVEL establishes a benchmark and learnable supervision signal, but leaves open data scaling, data mixture design, better model architectures, and more systematic study of training for grounded verification (Arasteh et al., 10 Apr 2026, Gao et al., 25 Jun 2026).

In scientific and formal domains, structural verification has bounded scope. PDE-grounded intent verification states that IFS is a structural verifier for PDE intent, not a physical validity certificate: it does not verify mesh quality, solver convergence, discretization error, agreement with experimental measurements, or all coefficient/material subtleties. KG-based assertion synthesis improves local correctness, syntax stabilization, and provenance, but complex temporal and arithmetic reasoning, global invariants, and semantic debugging remain challenging and design-dependent (Song et al., 10 May 2026, Viswambharan et al., 7 May 2026).

High-stakes guideline- or rule-grounded systems inherit the limitations of their external criteria. GLEAN notes that guidelines may be incomplete, outdated, or flawed, and that verification should complement rather than replace human judgment. The neuro-symbolic causal rule-synthesis framework describes iterative refinement of candidate rules but does not formalize an automatic repair algorithm. PharmGraph-Auditor improves the safety-efficiency trade-off in prescription auditing, yet indication checking can still generate false positives when clinical procedural knowledge is missing (Zhang et al., 3 Mar 2026, Rehan et al., 30 Apr 2026, Zhu et al., 11 Mar 2026).

Agentic safety evaluation also has observability limits. Vera only tests inference-time risks that can be exercised through the runtime interface and only retains cases whose outcomes can be decided from observable evidence. Harms that leave no durable trace in environment state, logs, or outputs may therefore be undercounted. This suggests an unresolved boundary condition for criteria-grounded verification: the criterion may be explicit, but verification still depends on whether the environment exposes adequate evidence for that criterion (Feng et al., 2 Jul 2026).