ProsQA: Evaluating LLM Hallucinations

• ProsQA is a methodology that uses synthetic QA pairs with controlled factuality to distinguish grounded outputs from hallucinated ones.
• It employs a multi-stage pipeline involving prompt-based rewriting and curated document pools to generate and label claims systematically.
• The approach underpins evaluation protocols and preference-based alignment, driving safer, more interpretable reasoning in LLMs.

ProsQA (Prompted Reasoning over Synthetic Question–Answer Pairs) refers to a methodological and benchmark-building paradigm wherein synthetic datasets of question–answer (QA) pairs are used to expose, quantify, and remediate hallucination and grounding failures in LLMs, particularly in complex retrieval-augmented or open-domain reasoning settings. ProsQA methodologies often incorporate explicitly constructed, controllable claims or questions linked to curated document corpora. These facilitate fine-grained evaluation of LLM reasoning, grounding, and hallucination detection, as well as enabling preference-based distillation and alignment objectives.

1. Definition and Conceptual Motivation

ProsQA encompasses synthetic question–answer or claim–support pairs, typically generated (or filtered) using prompted LLMs, and constructed such that their factual status with respect to input context or retrieved documents is precisely known. This approach allows rigorous distinction between grounded and hallucinated model outputs, overcoming the ambiguity and artifact prevalence endemic to web-mined QA corpora. In hallucination detection, ProsQA provides a controlled substrate for investigating both data-driven and reasoning-driven error modes, supporting development of theoretically grounded, reference-free detectors (Zeng et al., 26 Jan 2026), as well as evidence-grounded small reasoning models (Bergeron et al., 1 Oct 2025).

A ProsQA corpus enables researchers to:

• Induce synthetic claims/questions (and labels) over a curated context set, ensuring full label control.
• Evaluate models’ capacity to distinguish grounded answers from hallucinated or speculative completions.
• Serve as a foundation for alignment data (preferences between candidate completions) and chain-of-thought (CoT) ablation.

2. Dataset Construction and Pipeline Architecture

ProsQA corpus creation typically follows multi-stage pipelines to ensure diversity, realism, and controllable truth labels. For instance, the HalluClaim corpus (Bergeron et al., 1 Oct 2025) is constructed by:

• Extracting a domain-agnostic raw document pool from large-scale crawls (e.g., FineWeb), with preprocessing filters for language, de-duplication, and minimal length.
• Applying multistage curation, including safety, quality, and stylistic re-formatting via prompt-based LLM rewriting (e.g., using Llama-3.3 DR with 18 distinct styles).
• Generating synthetic claims or questions for each document with a large claim generator (e.g., Llama-3.3 CG), labeling them as "grounded", "intrinsic hallucination", or "extrinsic hallucination".
• Optionally balancing the dataset by sampling equal proportions of grounded and hallucinated claims.

This methodology ensures high coverage, label reliability, and stylistic variety, which are critical for stress-testing model grounding and facilitating nuanced evaluation of hallucination behaviors.

3. Role in Hallucination Detection and Mitigation

ProsQA benchmarks have catalyzed advances in hallucination risk modeling and mitigation. The distinction between "data-driven" (training-time) and "reasoning-driven" (decoding-time) hallucinations (Zeng et al., 26 Jan 2026) leverages ProsQA settings by providing an environment where both sources of error can be independently manipulated and measured.

• Data-driven hallucinations arise when a model’s induced RKHS subspace (as captured by its NTK features) cannot approximate the true answer; ProsQA enables explicit probe construction and NTK-based geometric analysis.
• Reasoning-driven hallucinations emerge from inference-time instability—small perturbations amplify step-by-step, producing speculative answers; ProsQA claims facilitate rollout-based detection and reasoning-centric benchmarks (e.g., GSM8K, MATH-500).

The ProsQA paradigm is thus integral to both theory-driven approaches (deriving hallucination risk bounds) and empirical frameworks (e.g., NTK-based HalluGuard detectors (Zeng et al., 26 Jan 2026)) for systematic hallucination quantification.

4. ProsQA in Preference-based Alignment and Distillation

ProsQA-style preference data plays a central role in methods such as Odds Ratio Preference Optimization (ORPO) fine-tuning. For example, in HalluGuard-SRM (Bergeron et al., 1 Oct 2025), synthetic document–claim pairs feed into preference pipelines where responses from large and small generators are paired, agreement-verified, and filtered via LLM consensus. This process yields datasets of (prompt, preferred response, rejected response) tuples with high alignment fidelity, suitable for direct preference optimization objectives.

Such ProsQA-derived preference datasets support:

• Efficient distillation of large-model reasoning and alignment into smaller backbones (e.g., Qwen3-4B SRM).
• Simultaneous chain-of-thought supervision and model selection under transparent, evidence-grounded criteria.
• Robust evaluation and ablation of reasoning and justification quality.

5. Evaluation Protocols and Empirical Results

ProsQA underlies leading evaluation suites and enables statistically robust benchmarking against hallucination and grounding failures. In "HalluGuard: Evidence-Grounded Small Reasoning Models to Mitigate Hallucinations in Retrieval-Augmented Generation" (Bergeron et al., 1 Oct 2025), the HalluClaim-based procedure supports fine-grained claim-level judgment with chain-of-thought transparency and strict JSON-format justification outputs. On the RAGTruth benchmark (a ProsQA-style subset of LLM-AggreFact), HalluGuard-4B achieves 84.0% balanced accuracy, matching 7–8B baselines while using half the parameters. On the full benchmark, it attains 75.7% BAcc, equal to GPT-4o (75.9%).

Ablations reveal that removal of reasoning traces (/no_think) reduces BAcc by 8.1%, and SFT-only baselines drop >27%, emphasizing explicit ProsQA-driven reasoning as critical for hallucination discrimination. Justification quality, scored via G-Eval with GPT-4o, shows near parity between HalluGuard-4B and a 32B model across relevance, coherence, and consistency dimensions.

6. Limitations and Future Directions

Despite high empirical utility, ProsQA frameworks exhibit inherent constraints:

• Synthetic claims, while controllable, may not capture the spectrum of real-world hallucination patterns or adversarial linguistic artifacts (Bergeron et al., 1 Oct 2025).
• Unimodal and monolingual focus (typically English text) restricts coverage of multimodal, cross-lingual, or highly specialized reasoning domains.
• Labeling granularity is sometimes limited: intrinsic vs. extrinsic hallucinations are often collapsed, and strict JSON enforcement may underestimate model accuracy where outputs deviate in non-contentful ways.

Planned extensions include cross-modal ProsQA (e.g., tabular or diagram-referenced claims), expansion to non-English and domain-specialized corpora, and more granular labeling for hallucination subtype classification.

7. Significance in LLM Safety and Agent Design

The ProsQA paradigm has established itself as a foundational instrument for LLM safety research, reasoning traceability, and hallucination mitigation. It underpins advances in spectral anomaly detection for tool use hallucinations (Noël, 8 Feb 2026), theoretical risk quantification (Zeng et al., 26 Jan 2026), and evidence-grounded self-justifying reasoning models (Bergeron et al., 1 Oct 2025). ProsQA enables reproducible, scalable, and interpretable hallucination detection frameworks, promoting both experimental rigor and alignment transparency for deployment in high-stakes, safety-critical AI agent workflows.