Claim-Level Grounding Approach

• Claim-Level Grounding is a method that decomposes complex outputs into atomic claims, enabling independent verification of each assertion.
• It employs logical operators like immediate, mediate, and grounding tree to capture derivational structure and detailed inferential provenance.
• Empirical studies, especially in clinical and biomedical contexts, show improved precision, recall, and reduced hallucinations through fine-grained evaluation and reinforcement learning.

A claim-level grounding approach is a formal and computational methodology in which complex outputs—such as long-form text generated by LLMs or logical derivations—are decomposed into atomic claims whose correctness, support, and provenance can be assessed independently. The goal is to provide higher-fidelity factual grounding, increased transparency, and more granular diagnostics compared to sequence-level or holistic evaluation, particularly in domains where factual rigor and interpretability are critical (e.g., clinical documentation, biomedical question answering, proof theory).

1. Formal Foundations: Claim-Level Grounding and Logical Operators

The claim-level perspective promotes the decomposition of arguments or model outputs into minimal units of assertion, often referred to as "atomic claims." In formal logic, this perspective is operationalized via dedicated grounding operators that encode the provenance and inferential structure of claims. Notably, (Genco, 2023) introduces a language with three operators:

• $G_i$ (Immediate grounding): $\Gamma[\Delta] G_i B$ denotes that the set of immediate grounds $\Gamma$ (under conditions $\Delta$) suffice to establish $B$ in a single inferential step.
• $G_m$ (Mediate grounding): $\Gamma[\Delta] G_m B$ encodes $\Gamma$ as a mediate (transitive) ground of $B$—the transitive closure over immediate grounds.
• $G_t$ (Grounding tree): $\Gamma[\Delta] G_t B$ internalizes full derivation trees, encapsulating the entire chain of immediate grounding steps within a single sentential object.

The calculus supports modular construction, immediate-to-mediate chaining, and explicit recovery of inferential structure ("detour-elimination"), with precise inference rules governing introduction and elimination for each operator. This makes the claim-level approach highly amenable to proof-theoretic analysis, transitive closure operations, and harmony conditions, albeit at the expense of full logicality due to dependence on domain-specific grounding-rule schemata.

2. Generative Model Evaluation via Claim-Level Metrics

In natural language generation, claim-level grounding replaces coarse sequence metrics (e.g., BLEU, ROUGE) with fine-grained evaluations tracking the presence, omission, or hallucination of atomic claims relative to source facts. In long-form clinical note generation, (Jhaveri et al., 26 Sep 2025) frames the generation task as optimizing a policy $\pi_\theta$ to map dialogue $x$ to output $\tau$ with maximal completeness and factuality at the claim level.

The core innovation is DocLens—a deterministic evaluator extracting two sets of atomic claims:

• $R_x$: Reference claims derived from source $x$
• $O$: Claims output by the model in $\tau$

For each $r\in R_x$, $e(r)=1$ if $\tau$ entails $r$; for $o\in O$, $e'(o)=1$ if $x$ entails $o$. Precision and recall metrics are then computed:

$$\mathrm{Recall} = \frac{1}{|R_x|}\sum_{r\in R_x}e(r) \qquad \mathrm{Precision} = \frac{1}{|O|}\sum_{o\in O}e'(o)$$

The single-claim reward is a scaled F$_1$:

$$R_{\text{claim}} = 10 \times \frac{2\,\mathrm{Precision}\,\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}+\epsilon}$$

This signal penalizes both omissions and hallucinations at atomic granularity, aligning directly with clinical priorities and overcoming annotation bottlenecks and incompleteness associated with reference-based metrics.

3. Automated Verification and Fusion of Claim-Level Evidence

Claim-level grounding in retrieval-augmented text generation entails both extraction and verification of atomic claims. (Ji et al., 10 Jan 2026) presents the MedRAGChecker framework, which operates as follows on biomedical QA tasks:

1. Decomposition: Given a question $q$, context $D$, and an answer $a$, a trainable extractor $f_{\mathrm{ex}}$ generates a set of atomic claims $C = \{c_1,\ldots,c_n\}$.
2. Textual NLI Verification: Each claim $c_i$ is passed, together with evidence $D$, to an ensemble of student natural language inference (NLI) checkers, yielding $p_m(y|c_i,D)$ for $$y\in\{\text{Entail},\text{Neutral},\text{Contradict}\}$$.
3. KG Consistency: Claims are aligned (via string matching) to a biomedical knowledge graph (KG). Triples are scored using TransE embedding distance and passed through a sigmoid for probabilistic interpretation.
4. Soft Fusion: The final calibrated support probability for each claim fuses NLI and KG signals in logit space:

$$P^*(c) = \sigma \left( \beta\,\text{logit}(p_{\text{NLI}}(c)) + (1-\beta)\,\text{logit}(s_{\text{KG}}(c)) \right)$$

where $\beta$ is a tunable weight, and $s_{\text{KG}}$ is the weighted sum of KG and text-alignment scores.

1. Diagnostics & Aggregation: Verdicts per claim are aggregated into compositional answer-level metrics (e.g., Faith, Halluc, SafetyErr), enabling systematic identification of retrieval, inference, and safety-critical errors.

4. Optimization Algorithms and Training Protocols

For learning generative models robust to claim-level errors, reinforcement learning with claim-level rewards is essential. (Jhaveri et al., 26 Sep 2025) employs the Group Relative Policy Optimization (GRPO) algorithm:

• For a dialogue $x$, $k$ candidate outputs $\{\tau_1,\ldots,\tau_k\}$ are sampled.
• Claim-level reward $r_j$ (via DocLens) is computed for each $\tau_j$.
• The group mean $\bar r$ defines a baseline; the GRPO objective is

$$\mathcal{L}(\theta) = \frac{1}{k}\sum_{j=1}^k (r_j - \bar r)\log \pi_\theta(\tau_j|x)$$

• Gradients reinforce above-average candidates; no separate value network or reference note is needed.
• A reward-gating strategy ($\tau=0.6$) zeros out updates from candidates with low relative F$_1$, reducing variance and accelerating convergence.

Stepwise training includes precomputing reference claims, sampling rollouts, evaluating with DocLens, and optimizing via GRPO, all with high memory efficiency (single A100-80GB GPU).

5. Empirical Performance and Diagnostic Capabilities

Claim-level methods provide both quantitative and qualitative improvements in factuality and completeness:

• In clinical note generation (Jhaveri et al., 26 Sep 2025):

| Model & Epochs | Precision | Recall | F1 | |-----------------------|-----------|---------|--------| | Base (no RL) | 0.8436 | 0.6460 | 0.7317 | | GRPO (3 epochs) | 0.8987 | 0.6919 | 0.7819 | | GRPO + gating (2 ep.) | 0.8992 | 0.6887 | 0.7800 |

Out-of-domain on ACI-Bench: similar F1 gain ($\approx$4.6 pts). Subjective GPT-5 ratings indicate fewer omissions and hallucinations in GRPO-tuned models.

• In biomedical QA (Ji et al., 10 Jan 2026):
  • Ensemble claim checkers achieve $\approx$87\% accuracy and Macro-F1 $\approx$60\%.
  • KG-NLI fusion increases safety-critical claim Macro-F1 from 59.2\% (NLI-only) to 64.7\%.
  • From fused vs NLI-only: Faith increases 5–7 points, Halluc decreases 3–8 points, SafetyErr decreases 4–8 points.
  • Claim-level signal correlates with expert judgments ($\rho_\text{Faith,Correctness}=0.47$).

These results demonstrate reproducible gains in factual completeness and error detection over surface-level or aggregate metrics.

6. Significance, Limitations, and Theoretical Perspective

Claim-level grounding frameworks enable rigorous, reproducible, and scalable factuality evaluation and optimization in both symbolic and neural settings. Their modularity supports adaptations to domain-specific priorities (e.g., guideline adherence, billing in clinic), knowledge fusion, and fine-grained error analysis. Proof-theoretic grounding (Genco, 2023) highlights modular separation between immediate, mediate, and tree-structured derivations, offering opportunities for balance analysis and normalization within non-logical grounding calculi.

However, certain operators ($G_m$) entail informational loss regarding grounding steps, and practical claim extraction/verification must contend with imperfect extraction and matching (student extractor F1 $\approx$23-24\%, (Ji et al., 10 Jan 2026)). The gating and deterministic reward procedures mitigate, but do not eliminate, possible training noise or reference incompleteness. This suggests continued refinement of claim extraction, ontology alignment, and model calibration are essential for broad deployment.

Claim-level grounding is thus an essential methodological advance for high-stakes, factual text generation and for the formal study of inferential provenance in logic and AI.