Pith. sign in

REVIEW 3 major objections 7 minor 22 references

Hallucinated RAG sentences barely react when you remove the evidence; grounded ones collapse.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-11 20:49 UTC pith:ADXPARZH

load-bearing objection Solid span-level RAG hallucination detector with honest scope: training-free ablation features beat intrinsic baselines on context-built answers, match chunk NLI without a second model, and fail cleanly on short-answer QA. the 3 major comments →

arxiv 2607.04223 v1 pith:ADXPARZH submitted 2026-07-05 cs.CL cs.AIcs.LG

Detecting Hallucinations in Retrieval-Augmented Generation through Grounding-Aware Sensitivity by Perturbation (GASP)

Mohamed Aly Bouke This is my paper
classification cs.CL cs.AIcs.LG
keywords hallucination detectionretrieval-augmented generationcontext perturbationgrounding sensitivityspan-level detectionexplainable AIJensen-Shannon divergencelarge language models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Retrieval-augmented systems still invent unsupported sentences, and most detectors only say whether a whole answer looks bad. This paper argues that the right test is relational: hold the answer fixed and re-score it after you strip out the retrieved context, then score each sentence by how much its likelihood falls. A sentence built from a passage should lose probability when that passage disappears; a free-running hallucination should barely notice. The method, GASP, turns those likelihood drops and distribution shifts into a per-sentence grounding-sensitivity score, returns the chunk that mattered most as an explanation, and needs no separate verifier or labeled training data for its default threshold. On long, context-constructed answers it ranks unsupported content above grounded content more reliably than perplexity, length, whole-context entailment, or self-consistency, and it matches a strong chunk-level NLI checker without importing that second model. The same signal weakens when short answers can come from the model’s own knowledge, which the authors treat as a scope boundary rather than a universal failure.

Core claim

Grounding in RAG is best detected as dependence on evidence: re-score a fixed answer under the full context, under no context, and with each retrieved chunk removed; grounded sentences show large log-likelihood drops and Jensen-Shannon shifts when their support is gone, while hallucinated sentences stay almost unchanged. That grounding sensitivity, aggregated per sentence, is a usable span-level hallucination score that needs only a probabilistic scorer and that can flag sentences and name the supporting passage without a separate verifier.

What carries the argument

Grounding sensitivity: the change in a fixed span’s token log-likelihood and predictive distribution when the retrieved context is perturbed (full vs empty vs leave-one-chunk-out). The authors interpret decoding under a context as a context-conditioned random nonlinear iterated function system whose attractor of grounded continuations moves when evidence is removed; the detector reads that movement through likelihood drops and JSD.

Load-bearing premise

A grounded sentence’s probability under a competent scorer really does hinge on a specific retrieved passage rather than on what the model already knows or on leftover topic and style, so removing that passage produces a large local drop.

What would settle it

On a span-annotated summarization or data-to-text RAG set with answer-grouped folds, check whether grounded sentences still show systematically larger full-versus-removed likelihood drops and JSD than hallucinated ones; if the ranking falls to chance or fails to beat perplexity under the same protocol, the central claim does not hold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper proposes GASP, a span-level detector of RAG hallucinations that scores each answer sentence by grounding sensitivity: how much its token likelihood and predictive distribution change when the retrieved context is removed entirely or leave-one-chunk-out. Features are log-likelihood gaps/drops and Jensen–Shannon divergences under a fixed answer; the default detector is a training-free threshold on the standardized features, with an optional LightGBM variant. Decoding is interpreted as a context-conditioned RNIFS whose attractor moves under evidence edits (Assumption 1, Prop. 1), with empirical checks on hidden-state trajectories. Under a leakage-clean, response-grouped protocol on RAGTruth (three small instruction-tuned scorers from two families), GASP-threshold reaches response AUC ≈0.73 and span AUC ≈0.67, significantly above perplexity and above length, whole-context NLI, and SelfCheckGPT-style self-consistency, while matching a well-configured chunk-level NLI verifier at span level without a separate model. The signal transfers to TofuEval meeting summarization but not to short-answer RAGBench, which the authors treat as a scope boundary for context-constructed outputs.

Significance. If the ranking results hold, GASP is a practically useful, training-free, black-box guardrail for RAG systems that need sentence-level flags and a candidate supporting passage without maintaining a separate verifier or resampling answers. Strengths that raise the contribution above a routine baseline paper include: (i) a leakage-clean grouped protocol with bootstrap CIs and paired significance; (ii) a default zero-label threshold that matches the trained classifier; (iii) an honest negative transfer result on RAGBench that delimits the method; (iv) ablations, K/aggregation sensitivity, truncation audit, attribution proxies, and a qualitative case study; and (v) open-source scoring and evaluation code. The RNIFS framing is interpretive rather than a transformer guarantee, but the empirical core is self-contained and reproducible.

major comments (3)
  1. §8.4 reports that ~42–48% of contexts and ~25–28% of answers are truncated (700/200 tokens) and that truncation can remove the supporting passage of a grounded sentence, so absolute AUCs may be depressed or the effective grounding contract altered. Relative ranking vs baselines under the same truncation remains valid, but the load-bearing absolute numbers in the abstract and Tables 7–8 are not stress-tested with a longer-window re-run on a subset. A short ablation (e.g., 100 responses at 2k context on a larger GPU, or a no-truncation subset) would show whether the reported AUCs are conservative, optimistic, or stable, and would make the central performance claim less provisional.
  2. Table 12 and §8.2 show that a well-configured max-chunk NLI verifier (0.677) matches GASP-threshold (0.673) at span level within CIs. The paper correctly emphasizes no separate model, response-level scoring, and free attribution, but the abstract and §12 still lead with “improvements” and “clear margins” in a way that can be read as span-level superiority over entailment. Please rephrase the span-level claim to “competitive with chunk-level NLI while requiring no separate verifier and providing response-level scores and attribution,” so the central contribution is not overstated relative to the only competitive baseline.
  3. §8.5 evaluates the supporting-chunk explanation only against an NLI proxy (mean entailment 0.50 vs 0.05/0.07 controls; 69% agreement with best NLI chunk). Explainability is listed as a main contribution (Intro item 3; §5.5; §9.5). Without any human check of attribution quality, that claim rests on a circular proxy (NLI is also a baseline). A small labeled sample (even 50–100 grounded sentences scored correct/partial/topical/incorrect) would make the explanation contribution load-bearing rather than suggestive; if that is left to future work, the contribution list and deployment discussion should be toned down accordingly.
minor comments (7)
  1. §5.1 / Prop. 1: The RNIFS development is long relative to its empirical payoff (§8.7 shows trajectory movement correlates with the likelihood feature but does not itself separate grounded vs hallucinated spans). Consider shortening the proof sketch and moving more of the dynamical-systems material to an appendix, keeping the claim hierarchy of §9.1 prominent earlier.
  2. Table 9 (task split) reports point estimates with wide intervals and notes ~200 responses per task; the 1.5B summarization span AUC of 0.535 includes chance. Flag more clearly in the main text that per-task claims are underpowered relative to the pooled result.
  3. SelfCheckGPT baseline (§6.3, Table 12): N=4 regenerations from the small scorer at T=1.0 may understate sampling-consistency methods. A brief note that a stronger generator or larger N could raise this baseline would avoid over-reading the near-chance 0.486.
  4. Figures 3–5: ensure error bars and the chance line are legible in grayscale; the shared whole-context NLI curve in Fig. 5 is helpful but the competitive chunk-level NLI from Table 12 is only mentioned in text—consider adding it as a dashed reference or a table callout in the caption.
  5. Notation Table 2 lists W1, TV, L-bar, D used mainly in §5.1; a forward pointer from the table to Prop. 1 would help readers who skip the theory.
  6. Minor consistency: abstract says “about 0.73” / “about 0.67” while tables give scorer-specific means; either report the three-scorer range or the default-threshold means explicitly in the abstract.
  7. §13 Limitations: the exclusion of white-box mechanistic detectors and LettuceDetect is well motivated in §6.3; a one-sentence pointer there from the related-work positioning table would reduce the sense that strong recent baselines were simply omitted.

Circularity Check

0 steps flagged

No significant circularity: grounding features are counterfactual likelihood measurements independent of labels; RNIFS framing is interpretive, not load-bearing for the AUC claims.

full rationale

GASP’s features (gap, jsd∅, drop, jsdloo) are defined by re-scoring a fixed answer under full, empty, and leave-one-out contexts (Eqs. 8–11) and do not encode hallucination labels. The default detector is a training-free threshold on the standardized negated feature sum; the optional LightGBM is evaluated under response-grouped folds against external human annotations (RAGTruth, TofuEval, RAGBench). Baselines (perplexity, length, NLI, SelfCheckGPT-style) are external. Proposition 1 bounds an idealized RNIFS under an average-contraction assumption the authors explicitly do not claim holds for transformers (Sections 5.1, 9.1); Propositions 2–3 are elementary consequences of independence and JSD boundedness, not fits. The self-citation to the author’s RNIFS paper [16] supplies vocabulary for the attractor reading and is not used to force the empirical ranking results, which rest on measured likelihood/JSD contrasts versus held-out labels. No step reduces a claimed prediction to a fitted input or to a self-definitional identity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The empirical detector rests mainly on standard LM scoring and external benchmarks. Load-bearing modeling choices are the average-contraction RNIFS assumption (interpretive), fixed design knobs (K, token caps, aggregation), and the domain premise that grounded content is context-dependent for the scorer. No new physical entity is postulated; 'grounding sensitivity' is an operational definition.

free parameters (4)
  • K (number of context chunks)
    Fixed at 5 before experiments; sensitivity checked for {3,5,10}. Controls localization vs cost and enters every leave-one-out feature.
  • Context/answer truncation (700 / 200 tokens)
    Hardware-driven caps; ~42–48% of contexts and ~25–28% of answers truncated. Affects measured sensitivity and is not learned but is a free design choice that can remove supporting evidence.
  • GASP-threshold operating point / feature standardization
    Training-free detector standardizes four features and thresholds their negated sum; the decision threshold is deployment-chosen even though AUC is threshold-free.
  • LightGBM hyperparameters (300 trees, lr 0.05, 31 leaves)
    Fixed conventional settings for the optional trained variant; not tuned on eval data per the text, but still free choices for that variant.
axioms (4)
  • ad hoc to paper Average contraction and regular selection of token maps (Assumption 1): combined modulus L-bar < 1 so the transfer operator contracts in W1.
    Used to derive Proposition 1 bounding attractor movement by JSD. Authors state it does not hold exactly for transformers and treat it as a guide (Section 5.1).
  • domain assumption A competent probabilistic scorer that follows retrieved context will assign lower likelihood to grounded spans when supporting units are removed.
    Core causal premise of the detector; fails when parametric knowledge substitutes (RAGBench, Section 7.7).
  • domain assumption Human span annotations on RAGTruth/TofuEval/RAGBench are treated as ground truth for faithfulness/grounding.
    Construct validity of all AUC claims (Section 10).
  • standard math Autoregressive factorization and JSD as a bounded symmetric divergence between predictive distributions.
    Standard LM and information-theory background (Sections 3.2–3.4).
invented entities (2)
  • Grounding sensitivity (operational quantity) independent evidence
    purpose: Scalar per-span measure of dependence on evidence via likelihood drops and JSD under context edits.
    Defined in Definition 1 from re-scoring; not an external physical object. Independent handle is the empirical ranking and attribution proxies.
  • Context-conditioned RNIFS attractor of grounded continuations no independent evidence
    purpose: Interpretive dynamical-systems object whose movement under chunk removal is said to underlie the detector.
    Casting of decoding as RNIFS (Section 5.1); correlation dimension and trajectory movement are partial checks, not independent discovery of a new dynamical system in transformers.

pith-pipeline@v1.1.0-grok45 · 35843 in / 3551 out tokens · 38981 ms · 2026-07-11T20:49:43.918398+00:00 · methodology

0 comments
read the original abstract

Retrieval-augmented generation (RAG) reduces but does not eliminate hallucination, and existing detectors return a single answer-level score that does not indicate which sentence is unsupported, or why. To close this gap, we introduce Grounding-Aware Sensitivity by Perturbation (GASP), a span-level detector that scores each answer sentence by how strongly its likelihood depends on the retrieved evidence, a quantity we term grounding sensitivity. GASP holds the answer fixed and re-scores it under the full context, under no context, and with each chunk removed, then measures the log-likelihood drops and Jensen-Shannon divergences (JSD). The likelihood of a grounded sentence collapses once its supporting passage is removed, whereas a hallucinated sentence is almost unaffected, a contrast we interpret by casting decoding as a random nonlinear iterated function system (RNIFS). We evaluate GASP on three benchmarks (RAGTruth, TofuEval, RAGBench) with three instruction-tuned scorers from two model families (Qwen2.5-0.5B, Qwen2.5-1.5B, and SmolLM2-1.7B) under a leakage-clean protocol. On RAGTruth it reaches a response-level area under the ROC curve (AUC) of about 0.73 and a span-level AUC of about 0.67, improving significantly over perplexity and by clear margins over length, whole-context natural language inference (NLI), and self-consistency baselines. The only baseline competitive at the span level is a well-configured chunk-level entailment verifier, which requires a separate model, whereas a training-free threshold on the grounding features matches the trained classifier without labeled data and serves as the default detector. Beyond RAGTruth, the signal transfers to TofuEval but not to short-answer question answering in RAGBench, showing GASP is best suited to outputs constructed from the retrieved context rather than answers recoverable from parametric knowledge.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 3 linked inside Pith

  1. [1]

    Retrieval- augmented generation for knowledge-intensive NLP tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, et al., “Retrieval- augmented generation for knowledge-intensive NLP tasks,” inAdvances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  2. [2]

    Survey of halluci- nation in natural language generation,

    Z. Ji, N. Lee, R. Frieske, T. Yu, et al., “Survey of halluci- nation in natural language generation,”ACM Computing Surveys, vol. 55, no. 12, pp. 1–38, 2023

    2023

  3. [3]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,

    L. Huang, W. Yu, W. Ma, W. Zhong, et al., “A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,”ACM Trans- actions on Information Systems, 2025

    2025

  4. [4]

    Detecting hallucinations in large language models using semantic entropy,

    S. Farquhar, J. Kossen, L. Kuhn, and Y . Gal, “Detecting hallucinations in large language models using semantic entropy,”Nature, vol. 630, pp. 625–630, 2024

    2024

  5. [5]

    INSIDE: LLMs’ internal states retain the power of hallucination detection,

    C. Chen et al., “INSIDE: LLMs’ internal states retain the power of hallucination detection,” inInternational Conference on Learning Representations (ICLR), 2024

    2024

  6. [6]

    LLM-Check: Investigating detection of hallucinations in large language models,

    G. Sriramanan, S. Bharti, V . S. Sadasivan, S. Saha, P. Kattakinda, and S. Feizi, “LLM-Check: Investigating detection of hallucinations in large language models,” inAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  7. [7]

    SelfCheck- GPT: Zero-resource black-box hallucination detection for generative large language models,

    P. Manakul, A. Liusie, and M. J. F. Gales, “SelfCheck- GPT: Zero-resource black-box hallucination detection for generative large language models,” inProceedings of the Conference on Empirical Methods in Natural Lan- guage Processing (EMNLP), 2023

    2023

  8. [8]

    FActScore: Fine-grained atomic evaluation of factual precision in long form text generation,

    S. Min, K. Krishna, X. Lyu, M. Lewis, et al., “FActScore: Fine-grained atomic evaluation of factual precision in long form text generation,” inProceedings of the Con- ference on Empirical Methods in Natural Language Pro- cessing (EMNLP), 2023

    2023

  9. [9]

    ContextCite: Attributing model generation to context,

    B. Cohen-Wang, H. Shah, K. Georgiev, and A. Madry, “ContextCite: Attributing model generation to context,” inAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  10. [10]

    Lookback lens: Detecting and mitigating contextual hallucinations in large language models using only attention maps,

    Y .-S. Chuang, L. Qiu, C.-Y . Hsieh, R. Krishna, Y . Kim, and J. Glass, “Lookback lens: Detecting and mitigating contextual hallucinations in large language models using only attention maps,” inProceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2024

    2024

  11. [11]

    ReDeEP: Detecting hallucination in retrieval-augmented generation via mechanistic inter- pretability,

    Z. Sun et al., “ReDeEP: Detecting hallucination in retrieval-augmented generation via mechanistic inter- pretability,” inInternational Conference on Learning Representations (ICLR), 2025

    2025

  12. [12]

    LUMINA: Detecting hallucinations in RAG systems with context-knowledge signals,

    S. Yeh, Y . Li, and T. Mallick, “LUMINA: Detecting hallucinations in RAG systems with context-knowledge signals,” inInternational Conference on Learning Repre- sentations (ICLR), 2026

    2026

  13. [13]

    RAGTruth: A hallucination corpus for developing trustworthy retrieval- augmented language models,

    C. Niu, Y . Wu, J. Zhu, S. Xu, et al., “RAGTruth: A hallucination corpus for developing trustworthy retrieval- augmented language models,” inProceedings of the An- nual Meeting of the Association for Computational Lin- guistics (ACL), 2024

    2024

  14. [14]

    RAGAs: Automated evaluation of retrieval augmented generation,

    S. Es, J. James, L. Espinosa-Anke, and S. Schockaert, “RAGAs: Automated evaluation of retrieval augmented generation,” inProceedings of the European Chapter of the Association for Computational Linguistics (EACL): System Demonstrations, 2024

    2024

  15. [15]

    Kovács and G

    Á. Kovács and G. Recski,LettuceDetect: A hallucina- tion detection framework for RAG applications, arXiv preprint arXiv:2502.17125, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Fractal attractors in random nonlinear it- erated function systems: Existence, stability, and dimen- sional properties,

    M. A. Bouke, “Fractal attractors in random nonlinear it- erated function systems: Existence, stability, and dimen- sional properties,”Chaos, Solitons & Fractals, vol. 207, p. 117 996, 2026

    2026

  17. [17]

    Light- GBM: A highly efficient gradient boosting decision tree,

    G. Ke, Q. Meng, T. Finley, T. Wang, et al., “Light- GBM: A highly efficient gradient boosting decision tree,” Advances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  18. [18]

    TofuEval: Evaluating hallucinations of large language models on topic-focused dialogue summarization,

    L. Tang, I. Shalyminov, A. W.-m. Wang, J. Fang, et al., “TofuEval: Evaluating hallucinations of large language models on topic-focused dialogue summarization,” in Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2024

    2024

  19. [19]

    MeetingBank: A benchmark dataset for meeting summarization,

    Y . Hu, T. Ganter, H. Deilamsalehy, F. Dernoncourt, H. Foroosh, and F. Liu, “MeetingBank: A benchmark dataset for meeting summarization,” inProceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2023

    2023

  20. [20]

    RAGBench: Explain- able benchmark for retrieval-augmented generation sys- tems,

    R. Friel, M. Belyi, and A. Sanyal, “RAGBench: Explain- able benchmark for retrieval-augmented generation sys- tems,” inarXiv preprint arXiv:2407.11005, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Qwen2.5 technical report,

    Qwen Team, “Qwen2.5 technical report,” Alibaba Group, Tech. Rep., 2025, arXiv preprint

    2025

  22. [22]

    SmolLM2: When smol goes big – data-centric training of a small lan- guage model,

    L. B. Allal, A. Lozhkov, E. Bakouch, et al., “SmolLM2: When smol goes big – data-centric training of a small lan- guage model,”arXiv preprint arXiv:2502.02737, 2025. Page 23 of 23

    Pith/arXiv arXiv 2025