Pith. sign in

REVIEW 4 major objections 5 minor 40 references

A classifier on the LLM's mid-layer hidden state measures how much each retrieved context raises the model's confidence, and a reranker trained on that signal improves end-to-end RAG accuracy by up to 4.7 points.

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 · deepseek-v4-flash

2026-08-04 23:29 UTC pith:N7QZNRI6

load-bearing objection A clearly-meant RAG reranking idea whose headline screening claim is undercut by an unvalidated distribution shift in the confidence classifier; the end-to-end gains are real but narrow. the 4 major comments →

arxiv 2509.06472 v2 pith:N7QZNRI6 submitted 2025-09-08 cs.IR

Rethinking LLM Parametric Knowledge as Post-retrieval Confidence for Dynamic Retrieval and Reranking

Haoxiang Jin , Ronghan Li , Zixiang Lu , Qiguang Miao This is my paper
classification cs.IR
keywords internal hidden statesconfidence estimationretrieval-augmented generationrerankingdynamic retrievalknowledge boundarypreference alignmentNQ_Rerank
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 paper attempts to establish that the change in an LLM's internal confidence when a retrieved context is appended to a query — measured from the model's middle-layer hidden state before it produces its first token, in a single forward pass — is a trustworthy signal of whether that context actually helps the model answer correctly. From this confidence-shift signal the authors build a preference dataset, NQ_Rerank, fine-tune a reranker to place 'confidence-increasing' contexts at the top, and add Confidence-Based Dynamic Retrieval (CBDR), which skips retrieval entirely when the LLM's bare-query confidence exceeds a threshold. If correct, RAG systems get a principled way to filter misleading contexts and reduce retrieval cost simultaneously. Reported gains: +5.19 percentage points on context-screening Precision@1, up to +4.7 points on end-to-end RAG accuracy with Llama3-8B-Instruct, and roughly 7–17% of retrieval skipped at minimal accuracy cost. The alignment's value is model-specific — with Qwen2.5-7B-Instruct the same fine-tuned reranker shows no significant accuracy gain, which the paper attributes to architectural and training-method compatibility.

Core claim

On the paper's own terms, the central discovery is that the confidence shift Inc(Q,C_i) = Conf(H_M,Q+C_i) − Conf(H_M,Q) — the difference in the softmax probability that the confidence detection model E assigns to the 'will answer correctly' class when applied to the LLM's mid-layer hidden state just before the first answer token — encodes the target LLM's intrinsic preference among retrieved contexts. A context with positive Inc is a positive preference sample; negative Inc makes it a negative sample. The paper's claim is that this continuous, hidden-state-based signal captures knowledge-boundary interactions better than discrete relevance labels or multi-round sampling, and that it can be o

What carries the argument

The confidence detection model E is a binary classifier over the LLM's hidden state at the middle transformer layer (Layer/2), captured right before the first generated token. Its softmax output Conf(H) = P(Label=1 | E(H)) estimates the probability that the model will answer correctly; the confidence shift Inc(Q,C_i) for context C_i is the difference between Conf(H_M,Q+C_i) and Conf(H_M,Q). This shift is the paper's core mechanism: it generates the preference labels for the fine-tuning dataset NQ_Rerank, provides the supervision signal for the InfoNCE reranker loss, and drives the threshold-based retrieval trigger for CBDR.

Load-bearing premise

The confidence detector E is trained only on hidden states produced when the LLM receives the bare query, but the method then applies it to hidden states produced from query-plus-context inputs, and the paper gives no experiment showing that E's probability remains a valid confidence estimate under that distribution shift.

What would settle it

Take the NQ test set with gold answers, run the LLM on query+context inputs exactly as in Section 3.2.2, and evaluate whether Conf(H_M,Q+C) is higher for contexts whose presence makes the final answer correct than for contexts whose presence makes it incorrect. If the area under the ROC curve of E's probability on this query+context distribution is near 0.5 (or the ordering is reversed), then Inc does not measure genuine confidence enhancement, and the preference labels used to fine-tune the reranker are not grounded in actual answerability. A direct check: if a context scored Inc<0 by E consi

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Context ranking in RAG should follow the downstream LLM's own confidence rather than semantic similarity to the query; the fine-tuned reranker beats four baselines on Precision@1, Recall@1, and MRR@1 across K=1, 3, 5.
  • The confidence signal transfers across QA datasets: the same reranker tuned on NQ-derived preferences improves end-to-end accuracy on both NQ and HotpotQA when paired with Llama3-8B-Instruct.
  • CBDR exposes a tunable accuracy-cost trade-off: setting beta=0.98 skips 7.1% of retrievals while gaining 0.9 points on Top-3 accuracy, and setting beta=0.95 skips 16.7% while losing 0.2 points on Top-1.
  • A single forward pass of the LLM suffices for the preference signal, in contrast to multi-round sampling approaches such as SEAKR, making the method suitable for low-latency settings.
  • Preference alignment is tied to the target LLM: the same fine-tuned reranker gives near-zero gains with Qwen2.5-7B-Instruct, showing the signal is not LLM-agnostic.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If Inc is a faithful preference signal, it could be used as a reward for reinforcement learning over retrieval actions, needing only correct/incorrect answer labels to train E — no human preference judgments.
  • The same confidence-shift measurement could localize which part of a retrieved context caused the gain by computing Inc over progressively truncated spans, offering a cheap attribution tool for debugging RAG failures.
  • Since the reranker's value collapses when the downstream model changes, a direct extension is to train E on each target LLM's own hidden states; the paper's own results predict that doing so should restore the gains for Qwen2.5-7B-Instruct.
  • The threshold beta is tuned globally; adaptive per-query or per-domain thresholds would likely squeeze more efficiency from CBDR, since query difficulty varies widely within a dataset.

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 rethinking LLM parametric knowledge as post-retrieval confidence to improve RAG. The authors train a binary hidden-state classifier E on query-only states (H_{M,Q}) to estimate confidence, define the confidence shift Inc(Q,C_i)=Conf(H_{M,Q+C_i})-Conf(H_{M,Q}) (Eq. 3), use Inc to construct a preference dataset NQ_Rerank, fine-tune a reranker (bge-reranker-v2-m3-ft) with InfoNCE loss, and introduce Confidence-Based Dynamic Retrieval (CBDR) that skips retrieval when the query-only confidence exceeds a threshold beta. Experiments report a 5.19pp improvement in context screening (Table 1), up to 4.7pp end-to-end RAG accuracy gain (Table 2), and reduced retrieval cost with CBDR (Table 3).

Significance. If the confidence-shift signal is valid, the paper proposes a practical and efficient way to align reranking with a target LLM's internal preferences: a single forward pass replaces the multi-round sampling used by methods like SEAKR. The idea of using hidden-state confidence changes as a preference signal is plausible and worth investigating. However, the current evidence does not establish validity: the central screening-accuracy claim is measured against labels produced by the same untested classifier E, and the end-to-end gains are limited to one backbone LLM without statistical grounding. The paper would be significant if the transfer of E to query+context inputs were validated and Table 1 were re-evaluated with external answer-accuracy labels.

major comments (4)
  1. [Section 3.1.1 and Section 3.2.2, Eq. (3)] E is trained only on query-only hidden states H_{M,Q} with labels derived from whether the LLM answers correctly (Section 3.1.1). It is then applied to query+context hidden states H_{M,Q+C_i} to compute Inc in Eq. (3). No experiment reports E's accuracy, AUC, or calibration, even on its training distribution, and no test checks whether the softmax probability of E remains a valid confidence estimate when contexts are appended. Since every preference label in NQ_Rerank, the InfoNCE training in Eq. (5), and the CBDR threshold decisions in Section 3.4 depend on this transfer, the central mechanism is unvalidated. Please report E's discrimination and calibration on query-only data, and directly measure whether Inc correlates with answer correctness under external labels for query+context inputs.
  2. [Table 1] The NQ_Rerank test-set labels are generated by the same classifier E through Inc (Eq. 3). Thus the reported +5.19pp Precision@1 gain of bge-reranker-v2-m3-ft over bge-reranker-v2-m3 may reflect the reranker learning to reproduce E's inductive biases rather than learning to select contexts that genuinely help the LLM answer. The only way to escape this circularity is to evaluate the reranker against externally grounded labels, e.g., whether the top-ranked context actually improves the factual correctness (EM/F1) of the LLM's answer, or a human-annotated helpfulness label. The current Table 1 cannot support the screening-accuracy claim.
  3. [Table 2] The end-to-end accuracy results are the main non-circular evidence, but they are too weak as presented. The improvement is reported only for Llama3-8B-Instruct; with Qwen2.5-7B-Instruct the differences are negligible, which the paper itself acknowledges. No error bars, confidence intervals, or significance tests are given, and the number of evaluation examples is not stated. The table's formatting also conflates values (e.g., '47.2053.30' in the HotpotQA columns), making the numbers hard to verify. Please report per-example counts, repeated-run variance, and significance tests, and discuss why the effect disappears for Qwen.
  4. [Section 3.4 and Table 3] The CBDR efficiency claim is not precisely supported. The column header 'RR↓' is never defined, and the abstract's '7.10% reduction in retrieval costs' only matches one row (beta=0.98 with RR↓=92.90) while the Top-3 accuracy gain in that row is +0.9pp, not the '5.60% accuracy gains' stated in the abstract. The relationship between the reported β values, the retrieval-rate metric, and the claimed cost reduction needs clarification. Also, since the confidence threshold is applied using the same unvalidated E, the CBDR decision rule inherits the transfer risk noted above.
minor comments (5)
  1. [General] The manuscript retains ACM template placeholders ('Conference acronym ’XX', 'Do Not Use This Code', 'Generate the Correct Terms for Your Paper') and the CCS concepts section is not filled in. This is inappropriate for a journal submission.
  2. [Figure 2] The label 'Lable' should be 'Label'.
  3. [Eq. (1)] The notation C_{M,Q} for confidence conflicts with the use of C for retrieved contexts. Please use a distinct notation, e.g., Conf(M,Q).
  4. [Section 5.1] The claim that the observed correlation between reranker scale and performance establishes 'intrinsic validity' of NQ_Rerank does not follow. Scale-dependent ranking ability does not validate that the preference labels reflect helpfulness to the target LLM.
  5. [Section 4.1.2] The text says 'The experiments uniformly employed Llama3-8B-Instruct as the base LLM for downstream task reasoning' but Table 2 also reports Qwen2.5-7B-Instruct results. Rephrase to clarify which role each model plays.

Circularity Check

1 steps flagged

Table 1's screening-accuracy gain is measured against preference labels generated by the paper's own classifier E; the end-to-end Table 2 is independent.

specific steps
  1. self definitional [Section 3.2.1, Eq. 2; Section 3.2.2, Eq. 3; Section 4.2.1, Table 1]
    "this paper defines the following preference criterion: a context 𝐶 is considered to exhibit a positive preference for the target LLM 𝑀 in answering question 𝑄 if and only if it provides effective informational enhancement, satisfying the condition 𝐶𝑜𝑛𝑓(𝐻 𝑀,𝑄+𝐶)>𝐶𝑜𝑛𝑓(𝐻 𝑀,𝑄). ... If 𝐼𝑛𝑐(𝑄,𝐶 𝑖)> 0, the sample is labeled as a positive preference sample. ... this process constructs the final preference dataset, denoted as NQ_Rerank. ... To evaluate whether the fine-tuned Reranker ... comparative experiments were conducted on the NQ_Rerank test set."

    The 'positive preference' label is defined by Conf, which Eq. 2 defines as the softmax output of the authors' classifier E. Thus NQ_Rerank's positive/negative examples are generated by E itself. Table 1 then evaluates rerankers against these same E-generated labels. The reported 5.19pp Precision@1 gain therefore measures how well bge-reranker-v2-m3-ft reproduces E's preferences on a held-out split, not whether Inc corresponds to actual answer correctness. The screening-accuracy claim reduces by construction to agreement with the method's own label generator: the ground truth is defined as the method's own output.

full rationale

The paper's central end-to-end claims are not circular: Table 2 scores RAG accuracy against external ground-truth answers on NQ and HotpotQA, and Table 3's CBDR accuracy/cost trade-off is also externally measured. No load-bearing self-citation or author-uniqueness argument is present; the cited prior work [18] is not by the current authors. The circular component is confined to Table 1's 'contexts screening accuracy' claim, whose positive/negative labels are generated by the same fitted classifier E that defines the preference signal (Eqs. 2-3). The paper also never reports E's own test accuracy, AUC, or calibration (Section 4.1.1 mentions a 'test set (Test) includes 500 positive and 500 negative samples for the final performance evaluation of the mode', but no results are given), and it provides no transfer experiment showing that E's softmax output on query+context states remains a valid confidence estimate. That is a missing-validation concern, not a definitional circularity. Overall score 6: one headline 'prediction' reduces to self-prediction, while the main end-to-end RAG result retains independent empirical content.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The central claim rests on the hidden-state confidence model E and on the postulate that a single forward pass confidence shift is a valid preference signal. E's transfer from query-only to query+context states is the largest unvalidated step. The CBDR threshold beta is the main tuned free parameter; the reported cost savings depend on it.

free parameters (1)
  • CBDR confidence threshold beta = beta = 0.95 and 0.98, varied in Table 3
    Chosen by hand to trade retrieval cost against accuracy; the reported cost savings (RR 83.3% and 92.9%) depend on these values and would change with beta.
axioms (4)
  • domain assumption Mid_Layer (Layer/2) pre-token hidden state is a faithful continuous proxy for the LLM's confidence in answering a question.
    Invoked in Section 3.1.1 to select H_M,Q; justified only by citations [18,36], not re-derived in this paper.
  • domain assumption The classifier E trained on query-only hidden states generalizes to query+context hidden states.
    E's training samples are (H_M,Q, Label_Q) in Section 3.1.1, but inference in Section 3.2.2 uses H_M,Q+C_i; no validation of transfer is reported.
  • domain assumption Answering correctly (by ground-truth match) is equivalent to being confident, so E's P(Label=1) can be called confidence.
    Training labels are correctness labels (Section 3.1.1); the paper then interprets the softmax probability as confidence in Eq. 2.
  • domain assumption Contexts that raise E's predicted confidence are genuinely beneficial contexts for the LLM.
    The preference definition in Section 3.2.1 equates preference with positive confidence shift; this is the core postulate that makes NQ_Rerank labels meaningful.

pith-pipeline@v1.3.0-alltime-deepseek · 13425 in / 13254 out tokens · 134945 ms · 2026-08-04T23:29:25.402619+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Rethinking LLM Parametric Knowledge as Post-retrieval Confidence for Dynamic Retrieval and Reranking." pith.science (2026). https://pith.science/paper/N7QZNRI6

@misc{pith2026250906472,
  author       = {Pith},
  title        = {Pith review of: Rethinking LLM Parametric Knowledge as Post-retrieval Confidence for Dynamic Retrieval and Reranking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N7QZNRI6}},
  note         = {Machine review of arXiv:2509.06472}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large Language Models (LLMs) often generate inaccurate responses (hallucinations) when faced with questions beyond their knowledge scope. Retrieval-Augmented Generation (RAG) addresses this by leveraging external knowledge, but a critical challenge remains: determining whether retrieved contexts effectively enhance the model`s ability to answer specific queries. This challenge underscores the importance of knowledge boundary awareness, which current methods-relying on discrete labels or limited signals-fail to address adequately, as they overlook the rich information in LLMs` continuous internal hidden states. To tackle this, we propose a novel post-retrieval knowledge filtering approach. First, we construct a confidence detection model based on LLMs` internal hidden states to quantify how retrieved contexts enhance the model`s confidence. Using this model, we build a preference dataset (NQ_Rerank) to fine-tune a reranker, enabling it to prioritize contexts preferred by the downstream LLM during reranking. Additionally, we introduce Confidence-Based Dynamic Retrieval (CBDR), which adaptively triggers retrieval based on the LLM`s initial confidence in the original question, reducing knowledge conflicts and improving efficiency. Experimental results demonstrate significant improvements in accuracy for context screening and end-to-end RAG performance, along with a notable reduction in retrieval costs while maintaining competitive accuracy.

Figures

Figures reproduced from arXiv: 2509.06472 by Haoxiang Jin, Qiguang Miao, Ronghan Li, Zixiang Lu.

Figure 1
Figure 1. Figure 1: The left part of the diagram (Figure 1a) contrasts two distinct reranking strategies for RAG systems. One employs a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The complete process of aligning the Reranker with the target LLM involves constructing a preference dataset, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The workflow of Confidence-Based Dynamic Retrieval (CBDR). By varying the confidence threshold [PITH_FULL_IMAGE:figures/full_fig_p004_3.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

40 extracted references · 10 canonical work pages · 3 internal anchors

  1. [1]

    A survey of large language models.arXiv preprint arXiv:2303.18223, 1(2), 2023

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models.arXiv preprint arXiv:2303.18223, 1(2), 2023

    Pith/arXiv arXiv 2023

  2. [2]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yas- mine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhos- ale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Pith/arXiv arXiv 2023

  3. [3]

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, et al. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    To- wards mitigating llm hallucination via self reflection

    Ziwei Ji, Tiezheng Yu, Yan Xu, Nayeon Lee, Etsuko Ishii, and Pascale Fung. To- wards mitigating llm hallucination via self reflection. InFindings of the Association for Computational Linguistics: EMNLP 2023, pages 1827–1843, 2023

    work page 2023

  5. [5]

    Knowledge injection to counter large language model (llm) hallucination

    Ariana Martino, Michael Iannelli, and Coleen Truong. Knowledge injection to counter large language model (llm) hallucination. InEuropean Semantic Web Conference, pages 182–185. Springer, 2023

    work page 2023

  6. [6]

    Leveraging passage retrieval with generative models for open domain question answering.arXiv preprint arXiv:2007.01282, 2020

    Gautier Izacard and Edouard Grave. Leveraging passage retrieval with generative models for open domain question answering.arXiv preprint arXiv:2007.01282, 2020

    Pith/arXiv arXiv 2007

  7. [7]

    Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020

    2020

  8. [8]

    Making retrieval- augmented language models robust to irrelevant context.arXiv preprint arXiv:2310.01558, 2023

    Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Berant. Making retrieval- augmented language models robust to irrelevant context.arXiv preprint arXiv:2310.01558, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Enhancing noise robustness of retrieval-augmented language models with adaptive adversarial training.arXiv preprint arXiv:2405.20978, 2024

    Feiteng Fang, Yuelin Bai, Shiwen Ni, Min Yang, Xiaojun Chen, and Ruifeng Xu. Enhancing noise robustness of retrieval-augmented language models with adaptive adversarial training.arXiv preprint arXiv:2405.20978, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    The power of noise: Redefining retrieval for rag systems

    Florin Cuconasu, Giovanni Trappolini, Federico Siciliano, Simone Filice, Cesare Campagnano, Yoelle Maarek, Nicola Tonellotto, and Fabrizio Silvestri. The power of noise: Redefining retrieval for rag systems. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 719–729, 2024

    2024

  11. [11]

    Survey of hallucination in natural language generation.ACM computing surveys, 55(12):1–38, 2023

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation.ACM computing surveys, 55(12):1–38, 2023

    2023

  12. [12]

    Statistical knowledge assessment for large language models.Advances in Neural Information Processing Systems, 36:29812–29830, 2023

    Qingxiu Dong, Jingjing Xu, Lingpeng Kong, Zhifang Sui, and Lei Li. Statistical knowledge assessment for large language models.Advances in Neural Information Processing Systems, 36:29812–29830, 2023

    work page 2023

  13. [13]

    Benchmarking Knowledge Boundary for Large Language Models: A Different Perspective on Model Evaluation

    Xunjian Yin, Xu Zhang, Jie Ruan, and Xiaojun Wan. Benchmarking knowledge boundary for large language models: A different perspective on model evaluation. arXiv preprint arXiv:2402.11493, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  14. [14]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  15. [15]

    Investigating the factual knowledge boundary of large language models with retrieval augmentation.arXiv preprint arXiv:2307.11019, 2023

    Ruiyang Ren, Yuhao Wang, Yingqi Qu, Wayne Xin Zhao, Jing Liu, Hao Tian, Hua Wu, Ji-Rong Wen, and Haifeng Wang. Investigating the factual knowledge boundary of large language models with retrieval augmentation.arXiv preprint arXiv:2307.11019, 2023

    Pith/arXiv arXiv 2023

  16. [16]

    Knowledge boundary of large language models: A survey.arXiv preprint arXiv:2412.12472, 2024

    Moxin Li, Yong Zhao, Wenxuan Zhang, Shuaiyi Li, Wenya Xie, See-Kiong Ng, Tat-Seng Chua, and Yang Deng. Knowledge boundary of large language models: A survey.arXiv preprint arXiv:2412.12472, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    Enhancing llm reliability via explicit knowledge boundary modeling.arXiv preprint arXiv:2503.02233, 2025

    Hang Zheng, Hongshen Xu, Yuncong Liu, Lu Chen, Pascale Fung, and Kai Yu. Enhancing llm reliability via explicit knowledge boundary modeling.arXiv preprint arXiv:2503.02233, 2025

    arXiv 2025

  18. [18]

    Towards fully exploiting llm internal states to enhance knowledge boundary perception

    Shiyu Ni, Keping Bi, Jiafeng Guo, Lulu Yu, Baolong Bi, and Xueqi Cheng. Towards fully exploiting llm internal states to enhance knowledge boundary perception. arXiv preprint arXiv:2502.11677, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    Divide-Then-Align: Honest Alignment based on the Knowledge Boundary of RAG

    Xin Sun, Jianan Xie, Zhongqi Chen, Qiang Liu, Shu Wu, Yuehe Chen, Bowen Song, Weiqiang Wang, Zilei Wang, and Liang Wang. Divide-then-align: Honest align- ment based on the knowledge boundary of rag.arXiv preprint arXiv:2505.20871, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  20. [20]

    Do large language models know what they don’t know?arXiv preprint arXiv:2305.18153, 2023

    Zhangyue Yin, Qiushi Sun, Qipeng Guo, Jiawen Wu, Xipeng Qiu, and Xuanjing Huang. Do large language models know what they don’t know?arXiv preprint arXiv:2305.18153, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    When do llms need retrieval augmentation? mitigating llms’ overconfidence helps retrieval augmentation

    Shiyu Ni, Keping Bi, Jiafeng Guo, and Xueqi Cheng. When do llms need retrieval augmentation? mitigating llms’ overconfidence helps retrieval augmentation. arXiv preprint arXiv:2402.11457, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Are large language models more honest in their probabilistic or verbalized confidence? InChina Conference on Information Retrieval, pages 124–135

    Shiyu Ni, Keping Bi, Lulu Yu, and Jiafeng Guo. Are large language models more honest in their probabilistic or verbalized confidence? InChina Conference on Information Retrieval, pages 124–135. Springer, 2024

    work page 2024

  23. [23]

    Large language monkeys: Scaling inference compute with repeated sampling.arXiv preprint arXiv:2407.21787, 2024

    Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V Le, Christo- pher Ré, and Azalia Mirhoseini. Large language monkeys: Scaling inference compute with repeated sampling.arXiv preprint arXiv:2407.21787, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Statistical uncertainty quan- tification for aggregate performance metrics in machine learning benchmarks

    Rachel Longjohn, Giri Gopalan, and Emily Casleton. Statistical uncertainty quan- tification for aggregate performance metrics in machine learning benchmarks. arXiv preprint arXiv:2501.04234, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Unsupervised real-time hallucination detection based on the internal states of large language models.arXiv preprint arXiv:2403.06448, 2024

    Weihang Su, Changyue Wang, Qingyao Ai, Yiran Hu, Zhijing Wu, Yujia Zhou, and Yiqun Liu. Unsupervised real-time hallucination detection based on the internal states of large language models.arXiv preprint arXiv:2403.06448, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    Inside: Llms’ internal states retain the power of hallucination detection

    Chao Chen, Kai Liu, Ze Chen, Yi Gu, Yue Wu, Mingyuan Tao, Zhihang Fu, and Jieping Ye. Inside: Llms’ internal states retain the power of hallucination detection. arXiv preprint arXiv:2402.03744, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Improving generalization in intent detection: Grpo with reward- based curriculum sampling.arXiv preprint arXiv:2504.13592, 2025

    Zihao Feng, Xiaoxue Wang, Ziwei Bai, Donghang Su, Bowen Wu, Qun Yu, and Baoxun Wang. Improving generalization in intent detection: Grpo with reward- based curriculum sampling.arXiv preprint arXiv:2504.13592, 2025

    Pith/arXiv arXiv 2025

  28. [28]

    LLM-Independent Adaptive RAG: Let the Question Speak for Itself

    Maria Marina, Nikolay Ivanov, Sergey Pletenev, Mikhail Salnikov, Daria Gal- imzianova, Nikita Krayko, Vasily Konovalov, Alexander Panchenko, and Viktor Moskvoretskii. Llm-independent adaptive rag: Let the question speak for itself. arXiv preprint arXiv:2505.04253, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  29. [29]

    Seakr: Self-aware knowledge retrieval for adaptive retrieval augmented generation.arXiv preprint arXiv:2406.19215, 2024

    Zijun Yao, Weijian Qi, Liangming Pan, Shulin Cao, Linmei Hu, Weichuan Liu, Lei Hou, and Juanzi Li. Seakr: Self-aware knowledge retrieval for adaptive retrieval augmented generation.arXiv preprint arXiv:2406.19215, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Replug: Retrieval-augmented black-box language models.arXiv preprint arXiv:2301.12652, 2023

    Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Replug: Retrieval-augmented black-box language models.arXiv preprint arXiv:2301.12652, 2023

    Pith/arXiv arXiv 2023

  31. [31]

    Query optimization for parametric knowledge refinement in retrieval-augmented large language models.arXiv preprint arXiv:2411.07820, 2024

    Youan Cong, Cheng Wang, Pritom Saha Akash, and Kevin Chen-Chuan Chang. Query optimization for parametric knowledge refinement in retrieval-augmented large language models.arXiv preprint arXiv:2411.07820, 2024

    work page arXiv 2024

  32. [32]

    Understand what llm needs: Dual preference alignment for retrieval-augmented generation

    Guanting Dong, Yutao Zhu, Chenghao Zhang, Zechen Wang, Ji-Rong Wen, and Zhicheng Dou. Understand what llm needs: Dual preference alignment for retrieval-augmented generation. InProceedings of the ACM on Web Conference 2025, pages 4206–4225, 2025

    work page 2025

  33. [33]

    Bridging relevance and reasoning: Rationale distillation in retrieval-augmented generation

    Pengyue Jia, Derong Xu, Xiaopeng Li, Zhaocheng Du, Xiangyang Li, Yichao Wang, Yuhao Wang, Qidong Liu, Maolin Wang, Huifeng Guo, et al. Bridging relevance and reasoning: Rationale distillation in retrieval-augmented generation. arXiv preprint arXiv:2412.08519, 2024

    arXiv 2024

  34. [34]

    Layer by layer: Uncovering hidden representations in language models.arXiv preprint arXiv:2502.02013, 2025

    Oscar Skean, Md Rifat Arefin, Dan Zhao, Niket Patel, Jalal Naghiyev, Yann LeCun, and Ravid Shwartz-Ziv. Layer by layer: Uncovering hidden representations in language models.arXiv preprint arXiv:2502.02013, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Reasoning models know when they’re right: Probing hidden states for self-verification.arXiv preprint arXiv:2504.05419, 2025

    Anqi Zhang, Yulin Chen, Jane Pan, Chen Zhao, Aurojit Panda, Jinyang Li, and He He. Reasoning models know when they’re right: Probing hidden states for self-verification.arXiv preprint arXiv:2504.05419, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    The internal state of an llm knows when it’s lying.arXiv preprint arXiv:2304.13734, 2023

    Amos Azaria and Tom Mitchell. The internal state of an llm knows when it’s lying.arXiv preprint arXiv:2304.13734, 2023

    Pith/arXiv arXiv 2023

  37. [37]

    Toutanova, Llion Jones, Ming-Wei Chang, Andrew Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Matthew Kelcey, Jacob Devlin, Kenton Lee, Kristina N. Toutanova, Llion Jones, Ming-Wei Chang, Andrew Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. Natural questions: a benchmark for question answering research.Transact...

    2019

  38. [38]

    Hotpotqa: A dataset for diverse, explainable multi-hop question answering.arXiv preprint arXiv:1809.09600, 2018

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering.arXiv preprint arXiv:1809.09600, 2018

    Pith/arXiv arXiv 2018

  39. [39]

    The llama 3 herd of models.arXiv e-prints, pages arXiv–2407, 2024

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models.arXiv e-prints, pages arXiv–2407, 2024

    2024

  40. [40]

    Qwen2.5: A party of foundation models, September 2024

    Qwen Team. Qwen2.5: A party of foundation models, September 2024. Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009

    work page 2024