Pith. sign in

REVIEW 4 major objections 5 minor 30 references

One forward pass tells LLMs when to retrieve and when to think harder

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 ↗

Regression probes on LLM hidden states estimate knowledge insufficiency and ambiguity to guide adaptive retrieval and reasoning in QA.

T0 review reviewed 2026-07-09 challenge →

load-bearing objection Decomposed uncertainty routing is a good idea, but the decomposition itself is never validated the 4 major comments →

arxiv 2607.07380 v1 pith:Z4N74UJS submitted 2026-07-08 cs.IR

Interpretable Uncertainty for Adaptive Retrieval and Reasoning in Question Answering

Ritajit Dey , Iadh Ounis , Graham McDonald This is my paper
classification cs.IR
keywords uncertainty estimationadaptive retrievalretrieval-augmented generationquestion answeringhidden state probingknowledge insufficiencyknowledge ambiguityLLM interpretability
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.

The reading

This paper proposes that two distinct forms of uncertainty—knowledge insufficiency (the model lacks the fact) and knowledge ambiguity/conflict (the model has competing versions of the fact)—can be estimated from LLM hidden states using lightweight regression probes in a single forward pass. The key claim is that these two signals, once decomposed, can drive a simple threshold-based decision mechanism: trigger retrieval-augmented generation when insufficiency is high, apply additional reasoning (such as chain-of-thought or self-consistency) when ambiguity is high, and answer directly when both are low. The paper reports that this framework outperforms both an LLM-only baseline and an always-on RAG baseline on the Natural Questions dataset using Llama-2-7b-chat, with improvements of 3.3–5.9 percentage points depending on the uncertainty metric used.

Core claim

The central mechanism is a pair of regression probes trained on LLM hidden states that predict two quantities: (1) the number of occurrences of a fact in the pretraining corpus, serving as a proxy for knowledge insufficiency, and (2) the entropy across competing versions of a fact, serving as a proxy for knowledge ambiguity or conflict. Because both probes operate on hidden states extracted during a single forward pass, the system can classify the type of uncertainty the model faces—missing knowledge versus conflicting knowledge—and route the question to the appropriate intervention (retrieval or reasoning) without repeated sampling or auxiliary models.

What carries the argument

Two regression probes (one predicting fact occurrence counts, one predicting fact-version entropy), a threshold-based decision mechanism, Llama-2-7b-chat hidden states, Semantic Entropy and Weighted Entropy Production Rate as evaluation metrics, McNemar's test for significance

Load-bearing premise

The probes are trained to predict corpus occurrence counts and entropy of fact versions from hidden states, but the ground-truth labels come from external corpus statistics whose reliability as proxies for the model's actual internal knowledge state is not validated. If the probes learn dataset-specific correlations rather than genuine uncertainty decomposition, the interpretability claim weakens.

What would settle it

If the probes' uncertainty estimates do not generalize beyond the specific model (Llama-2-7b-chat) and dataset (Natural Questions) used, or if the occurrence-count and entropy labels do not actually correspond to the model's internal knowledge state, then the decomposed signals would reduce to opaque learned classifiers with no interpretability advantage.

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

Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The paper proposes an uncertainty-aware framework for adaptive question answering that distinguishes between two forms of uncertainty—knowledge insufficiency (estimated via regression probes predicting fact occurrence counts in the pretraining corpus) and knowledge ambiguity/conflict (estimated via probes predicting entropy of competing fact versions)—and routes each to a different intervention: RAG for insufficiency, additional reasoning for ambiguity. The framework is evaluated on Natural Questions with Llama-2-7b-chat, reporting accuracy improvements over LLM-only and always-on RAG baselines with McNemar's test significance. The paper is positioned as a workshop submission (WExIR '26) with preliminary results.

Significance. The conceptual framing—decomposing uncertainty into insufficiency vs. ambiguity and mapping each to a distinct, interpretable intervention—is well-motivated and addresses a genuine gap in the adaptive retrieval literature. The proposal to estimate both signals from hidden states in a single forward pass is practically appealing for latency-sensitive settings. However, the significance is substantially undermined by the absence of validation for the core components: no probe accuracy metrics, no signal decorrelation analysis, and no ablation isolating the contribution of the decomposition itself. The reported accuracy improvements, while promising, cannot be attributed to the proposed decomposition without these validations.

major comments (4)
  1. §2, 'Preliminary Results': The paper reports accuracy improvements (+5.9% over LLM-only, +3.3% over always-on RAG for SE) and states significance via McNemar's test, but no table, no exact accuracy figures, no p-values, and no description of the experimental setup (number of test examples, number of trials, confidence intervals) are provided. The results are described only in running text. This must be formalized in a table with complete details.
  2. §2, 'Uncertainty Estimation from Hidden States': No probe performance metrics (R², MSE, or correlation between predicted and ground-truth values) are reported for either the occurrence-count probe or the entropy probe. Without evidence that the probes predict their targets with reasonable accuracy, the claim that the system estimates 'knowledge insufficiency' and 'knowledge ambiguity/conflict' from hidden states is unsupported. The reader and the stress-test note both identify this as the central gap, and I agree: this is load-bearing for the paper's contribution.
  3. §2, 'Adaptive Decision Mechanism' and §2, 'Preliminary Results': No correlation analysis between the two uncertainty signals is reported. If the insufficiency and ambiguity signals are highly correlated, the decomposition into two distinct interventions is not justified, and a single-signal baseline could achieve the same routing. Additionally, no ablation compares the two-signal decomposition against a single-signal baseline that triggers both RAG and reasoning from one combined score. Without these, the reported improvements could be entirely explained by a single uncertainty signal, with the decomposition adding no value.
  4. §2, 'Preliminary Results': The threshold selection procedure for the insufficiency and ambiguity thresholds is not described. It is unclear whether thresholds were tuned on a held-out set, selected heuristically, or set on the test data. If thresholds are selected on the test set, the reported improvements are not reproducible and may be inflated. The probe training methodology (which layer's hidden states, how many training examples, train/validation split) is also unspecified.
minor comments (5)
  1. Figure 1 is referenced but not visible in the reviewed text. Ensure the figure is legible and self-contained, as the pipeline description is otherwise difficult to follow from text alone.
  2. §1: The example 'Who is the head coach of Chelsea?' lists Thomas Tuchel, Graham Potter, Mauricio Pochettino, and Enzo Maresca. Verify these are accurate at publication time, as football management changes frequently.
  3. References [16], [17], [20] are dated 2026. Confirm these are published or have stable identifiers (DOIs/arXiv IDs). Reference [20] is cited as 'arXiv:2511.04418' which appears to be a 2025 arXiv preprint; the title and venue should be consistent.
  4. §2: The term 'knowledge insufficiency' is used to mean 'the model has not seen the fact enough times in pretraining.' This conflates corpus frequency with the model's actual internal knowledge. A brief note acknowledging this proxy relationship would strengthen the framing.
  5. The paper would benefit from a single sentence in §1 or §2 stating the contributions explicitly (e.g., 'Our contributions are: (1) ..., (2) ...'), as is conventional even for workshop papers.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and constructive review. The referee correctly identifies that the current manuscript, as a workshop submission with preliminary results, omits several critical experimental details and validations. We agree with all four major comments and will address each in the revised manuscript. Below we respond point by point.

read point-by-point responses

  1. Referee: §2, 'Preliminary Results': Results described only in running text with no table, exact figures, p-values, or experimental setup details.

    Authors: The referee is correct. The current manuscript reports results only in running text without a formal table or experimental details. In the revision, we will add a complete results table including exact accuracy figures for all conditions (LLM-only, always-on RAG, and our adaptive framework under both SE and WEPR), McNemar's test p-values, the number of test examples, and confidence intervals. We will also describe the experimental setup in full: test set size, number of trials, evaluation metric (exact match for NQ), and retrieval configuration (e.g., retriever, index, number of passages). revision: yes

  2. Referee: §2, 'Uncertainty Estimation from Hidden States': No probe performance metrics (R², MSE, correlation) reported for either probe.

    Authors: We agree this is the most critical gap. The claim that hidden states encode usable estimates of knowledge insufficiency and ambiguity is load-bearing for the paper's contribution, and without probe evaluation metrics it is unsupported. In the revision, we will report R², MSE, and Pearson/Spearman correlation between predicted and ground-truth values for both the occurrence-count probe and the entropy probe. We will also report these metrics across multiple hidden layers to identify which layer yields the strongest signal, and we will describe the probe architecture (linear vs. MLP), training data size, and train/validation/test splits. If probe performance turns out to be weak for one or both signals, we will report this honestly and discuss the implications for the framework's viability. revision: yes

  3. Referee: §2, 'Adaptive Decision Mechanism' and 'Preliminary Results': No correlation analysis between the two uncertainty signals; no ablation comparing two-signal decomposition against a single-signal baseline.

    Authors: This is a fair and important point. If the two signals are highly correlated, the decomposition adds no value over a single combined score. In the revision, we will add: (1) a correlation analysis (Pearson and Spearman) between the insufficiency and ambiguity signals across the test set, and (2) an ablation comparing the full two-signal decomposition against a single-signal baseline that triggers both RAG and reasoning from one combined uncertainty score. We will also include a baseline that uses only the insufficiency signal (triggering RAG only) and only the ambiguity signal (triggering reasoning only), to isolate the contribution of each component. These ablations are essential for attributing the observed improvements to the decomposition rather than to a single signal, and we agree the paper cannot make its claimed contribution without them. revision: yes

  4. Referee: §2, 'Preliminary Results': Threshold selection procedure and probe training methodology unspecified.

    Authors: The referee is correct that these details are missing and that without them the results are not reproducible and may appear inflated. In the revision, we will specify: (1) the threshold selection procedure, including whether thresholds were tuned on a held-out validation set (they will be—we will make this explicit and report the validation set size), (2) the probe training methodology including which hidden layer(s) were used, the number of training examples, the train/validation/test split, and the probe optimization details (learning rate, epochs, early stopping). We confirm that thresholds were not selected on the test set, and we will state this explicitly in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No significant circularity found; the probe training targets (corpus occurrence counts, entropy) are distinct from the evaluation metric (QA accuracy), and all load-bearing citations are to external authors.

full rationale

The paper trains regression probes on hidden states to predict two quantities: (1) fact occurrence counts in the pretraining corpus (following Kandpal et al. [9], Kang et al. [10] — external citations) and (2) entropy of competing fact versions (following Tomov et al. [20] — external citation). These predicted quantities are then used in a threshold-based routing mechanism (RAG vs. reasoning vs. direct answer). The evaluation metric is downstream QA accuracy on NQ, which is a different quantity from the probe training labels. Therefore, the 'prediction' (routing decision) is not equivalent to the training input by construction. No self-citation chain exists: the methodological foundations (probing [3–5], occurrence counts [9,10], entropy [20]) all cite external authors. The paper does not claim that the probes' predicted values are themselves novel findings — it explicitly frames them as learned estimates of externally-defined quantities. The absence of reported probe accuracy (R²/MSE), signal decorrelation analysis, and single-signal ablation are validity and completeness concerns (correctness risk), not circularity: they mean the decomposition claim is under-supported, not that it reduces to its inputs by definition. The one point is assigned for the minor observation that the paper's framing of 'interpretable uncertainty signals' is somewhat self-referential in that the interpretability labels (insufficiency/ambiguity) are assigned by the authors to the probe outputs without external validation that the probes actually capture those semantics — but this is a labeling/validity issue, not a construction-level circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper introduces no new entities (particles, forces, dimensions, etc.). It introduces a framework combining existing concepts. The free parameters are the two decision thresholds and the probe weights, none of which have stated values. The axioms are domain assumptions about the validity of occurrence counts and entropy as proxies for internal model knowledge state, all inherited from cited prior work but not independently validated in this paper.

free parameters (3)
  • knowledge insufficiency threshold
    Threshold above which RAG is triggered; value not stated in the paper, presumably tuned on NQ.
  • knowledge ambiguity/conflict threshold
    Threshold above which additional reasoning is triggered; value not stated in the paper, presumably tuned on NQ.
  • probe regression parameters
    Weights of the regression probes mapping hidden states to occurrence counts and entropy values; trained on external corpus statistics, architecture and training details not specified.
axioms (4)
  • domain assumption Fact occurrence counts in the pretraining corpus are a reliable proxy for the LLM's internal knowledge sufficiency.
    Section 2 states knowledge insufficiency is estimated as occurrence counts following Kandpal et al. [9] and Kang et al. [10]. This assumes the proxy is valid, which is not independently verified in this paper.
  • domain assumption Entropy of competing fact versions in the pretraining corpus is a reliable proxy for knowledge ambiguity/conflict in the LLM.
    Section 2 states knowledge ambiguity/conflict is conceptualized as entropy of occurrences following Tomov et al. [20]. The validity of this proxy for the model's internal state is assumed.
  • domain assumption Hidden states from a single forward pass contain sufficient information to predict both occurrence counts and entropy values.
    The entire framework depends on probes being able to extract these signals from hidden states. No probe accuracy metrics are reported to validate this assumption.
  • domain assumption The two uncertainty signals are sufficiently decomposable that routing to different interventions (RAG vs. reasoning) improves outcomes over undifferentiated policies.
    The decision mechanism assumes that insufficiency and ambiguity are orthogonal enough that separate thresholds make sense. This is supported by preliminary results but not systematically tested.

reviewed 2026-07-09 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Interpretable Uncertainty for Adaptive Retrieval and Reasoning in Question Answering." pith.science (2026). https://pith.science/paper/Z4N74UJS

@misc{pith2026260707380,
  author       = {Pith},
  title        = {Pith review of: Interpretable Uncertainty for Adaptive Retrieval and Reasoning in Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z4N74UJS}},
  note         = {Machine review of arXiv:2607.07380}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large language models (LLMs) achieve a strong performance in question answering (QA), but remain prone to hallucinations and suffer from limited transparency. Retrieval-augmented generation (RAG) can improve factuality, yet decisions about when and how to retrieve from external resources are typically based on opaque policies or computationally inefficient multi-step prompting procedures. We propose an uncertainty-aware framework for adaptive QA based on explicit signals derived from LLM internal representations. We distinguish between knowledge insufficiency and knowledge ambiguity or conflict, and efficiently estimate these from hidden states in a single forward pass. These signals guide system behaviour: RAG is triggered when knowledge is insufficient, while additional reasoning is applied when ambiguity or conflict is high. By grounding adaptive decisions in decomposed and efficiently estimable uncertainty signals, this approach provides a transparent and practical alternative to existing retrieval and reasoning strategies supporting the design of interpretable user-facing tools.

Figures

Figures reproduced from arXiv: 2607.07380 by Graham McDonald, Iadh Ounis, Ritajit Dey.

Figure 1
Figure 1. Figure 1: The proposed uncertainty-aware QA framework. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

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

30 extracted references · 30 canonical work pages · 2 internal anchors

  1. [1]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2024. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024

    work page 2024

  2. [2]

    Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell

    Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?. InProceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT 2021, Toronto, Canada, March 3-10, 2021

    work page 2021

  3. [3]

    Alexis Conneau, Douwe Kiela, Holger Schwenk, Loïc Barrault, and Antoine Bordes. 2018. What You Can Cram into a Single Vector: Probing Sentence Interpretable Uncertainty for Adaptive Retrieval and Reasoning in Question Answering WExIR ’26, July 24, 2026, Melbourne, Australia Embeddings for Linguistic Properties. InProceedings of the 56th Annual Meeting of ...

    work page 2018

  4. [4]

    John Hewitt and Christopher D. Manning. 2019. A Structural Probe for Finding Syntax in Word Representations. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers)

    work page 2019

  5. [5]

    Nora Hollenstein, Federico Pirovano, Ce Zhang, Lena Jäger, and Lisa Beinborn

    work page

  6. [6]

    Multilingual Language Models Predict Human Reading Behavior. InPro- ceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2021, Online, June 6-11, 2021

    work page 2021

  7. [7]

    Eyke Hüllermeier and Willem Waegeman. 2021. Aleatoric and Epistemic Uncer- tainty in Machine Learning: An Introduction to Concepts and Methods.Machine Learning110, 3 (2021), 457–506

    work page 2021

  8. [8]

    Soyeong Jeong, Jinheon Baek, Sukmin Cho, Sung Ju Hwang, and Jong Park

    work page

  9. [9]

    Adaptive-RAG: Learning to Adapt Retrieval-Augmented Large Language Models. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), NAACL 2024, Mexico City, Mexico, June 16-21, 2024

    work page 2024

  10. [10]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of Hallucination in Natural Language Generation.ACM Comput. Surv.55, 12, Article 248 (2023)

    work page 2023

  11. [11]

    Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel

    work page

  12. [12]

    InInterna- tional Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA

    Large language models struggle to learn long-tail knowledge. InInterna- tional Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA

    work page 2023

  13. [13]

    Cheongwoong Kang and Jaesik Choi. 2023. Impact of Co-occurrence on Fac- tual Knowledge of Large Language Models. InFindings of the Association for Computational Linguistics: EMNLP 2023, Singapore, December 6-10, 2023

    work page 2023

  14. [14]

    Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Genera- tion. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023

    work page 2023

  15. [15]

    Tom Kwiatkowski et al. 2019. Natural Questions: A Benchmark for Question Answering Research.Transactions of the Association for Computational Linguistics 7 (2019), 452–466

    work page 2019

  16. [16]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. InAdvances in Neural Information Processing Systems 33: Annual Conference on Neural Information...

    work page 2020

  17. [17]

    Alex Mallen, Akari Asai, Zexuan Zhong, et al . 2023. When Not to Trust Lan- guage Models: Investigating Effectiveness of Retrieval-Augmented Generation. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023

    work page 2023

  18. [18]

    Sara Vera Marjanovic, Haeun Yu, Pepa Atanasova, Maria Maistro, Christina Lioma, and Isabelle Augenstein. 2024. DYNAMICQA: Tracing Internal Knowledge Conflicts in Language Models. InFindings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-16, 2024

    work page 2024

  19. [19]

    Charles Moslonka, Hicham Randrianarivo, Arthur Garnier, and Emmanuel Mal- herbe. 2026. Learned Hallucination Detection in Black-Box Llms Using Token- Level Entropy Production Rate. InAdvances in Information Retrieval - 48th Eu- ropean Conference on Information Retrieval, ECIR 2026, Delft, The Netherlands, March 29 - April 2, 2026, Proceedings, Part I. Spr...

    work page 2026

  20. [20]

    Vy Nguyen, Ziqi Xu, Jeffrey Chan, Estrid He, Feng Xia, and Xiuzhen Zhang

    work page

  21. [21]

    Hallucinate Less by Thinking More: Aspect-Based Causal Abstention for Large Language Models. InFortieth AAAI Conference on Artificial Intelligence, Thirty-Eighth Conference on Innovative Applications of Artificial Intelligence, Six- teenth Symposium on Educational Advances in Artificial Intelligence, AAAI 2026, Singapore, January 20-27, 2026

    work page 2026

  22. [22]

    Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language Models as Knowledge Bases?. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Pro- cessing, EMNLP-IJCNLP 2019, Hong Kong, Chin...

    work page 2019

  23. [23]

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling LLM Test- Time Compute Optimally can be More Effective than Scaling Model Parameters. arXiv:2408.03314

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  24. [24]

    Tim Tomov, Dominik Fuchsgruber, Tom Wollschläger, and Stephan Günnemann

    work page

  25. [25]

    arXiv:2511.04418

    The Illusion of Certainty: Uncertainty Quantification for LLMs Fails Under Ambiguity. arXiv:2511.04418

    work page arXiv

  26. [26]

    Hugo Touvron et al . 2023. Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv:2307.09288

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  27. [27]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-Consistency Improves Chain of Thought Reasoning in Language Models. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023

    work page 2023

  28. [28]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. InAdvances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28...

    work page 2022

  29. [29]

    Sarah Wiegreffe and Yuval Pinter. 2019. Attention is not not Explanation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Pro- cessing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019

    work page 2019

  30. [30]

    Di Wu et al. 2024. REPOFORMER: selective retrieval for repository-level code completion. InForty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024

    work page 2024

This paper was first reviewed by glm-5.2 on July 9, 2026.