acceptodds
Under review as a conference paper at ICLR 2027

Trust or Abstain? A Self-Aware RAG Approach under Knowledge Conflicts

Abstract

Retrieval-augmented generation (RAG) improves large language models (LLM) by incorporating external context, but it also introduces knowledge conflicts when retrieved contextual knowledge (CK) and parametric knowledge (PK) disagree, or both fail. Existing approaches mainly coordinate which source to use, without explicitly asking whether each answer path is correct. We argue that faithful RAG requires LLM self-awareness, namely the ability to recognize the limits of its own knowledge and reasoning. To ground this problem, we construct a model-specific, ground-truth-aligned knowledge-conflict benchmark by evaluating four LLM backbones on PK-only and CK-conditioned answer paths over approximately 69K query-context instances per backbone, drawn from five conflict-QA datasets. We then introduce SABER, a lightweight Self-Aware Belief Estimator for RAG. SABER extracts and combines a self-prior and source-conditioned reasoning representations from frozen LLM hidden states, estimates PK-side and CK-side correctness beliefs with two lightweight predictors, and maps them into a 4-cell decision space covering trust PK, trust CK, trust both, or abstain. Across four backbones and five datasets, SABER improves end-to-end accuracy and conflict-specific faithfulness over ten baselines, with the largest gains on conflict-heavy datasets. When configured to abstain, SABER's risk-coverage curve Pareto-dominates every prompt-based abstainer, providing a tunable balance between coverage and risk on answered.

Primary area:
interpretability and explainable AI
Keywords:
retrieval-augmented generation, knowledge conflicts

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Z2YLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how RAG systems can choose between PK-only and CK-conditioned answers, or abstain when neither is reliable. It constructs a model-specific benchmark using the correctness of both answer paths. SABER combines question-only hidden states with representations from multiple reasoning traces and self-evaluation passes, and trains two lightweight correctness predictors while keeping the backbone frozen. The predictors select an existing candidate answer or optionally trigger abstention. Experiments on four 3B–8B models report improvements in conflict handling and selective answering over the evaluated inference-time baselines.

Critical strengths
  • The model-specific benchmark distinguishes PK-only and CK-conditioned answer correctness and explicitly includes cases where both answers are wrong. This enables a more informative evaluation of answer selection and abstention than overall accuracy alone.

  • SABER requires training only small prediction heads while keeping the backbone frozen. Across the four evaluated backbones, the results show improved conflict handling and more effective abstention than other baselines.

Critical weaknesses
  • The self-evaluation prompt explicitly relies on the model's own knowledge. Could this lead the model to reject factually correct contextual information beyond its PK? How does the CK evaluation distinguish between the required knowledge being absent from the context and the model being misled by its own PK?
  • The contribution beyond supervised correctness probing remains unclear. How does SABER compare with a simple classifier trained on the same labels using candidate-answer hidden states, without additional reasoning or self-evaluation? Besides, how does performance change when self-evaluation or either the PK or CK feature branch is removed?
  • The evaluation is restricted to small (3B-8B) models, with no larger open-weight models, limiting the breadth of the empirical findings and making the framework's broader conclusions less convincing.
  • The statement in lines 403–406 that reduced variants outperform or match SABER in both metrics is inconsistent with Figure 4: on Qwen-2.5-3B, C3 has slightly higher Acc (56.2% vs. 56.1%).
Questions for authors

Please refer to Weaknesses.

6 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/5)

We sincerely thank you for your valuable comments and insightful suggestions, and for recognizing our model-specific benchmark, the lightweight design of SABER on a frozen backbone, and its improved conflict handling and abstention. Below we address each of your questions in turn.

W1. Whether the own-knowledge self-evaluation prompt leads SABER to reject correct context beyond its PK, and how SABER distinguishes knowledge missing from the context from being misled by its own PK

We thank you for this insightful question. Below we first explain how SABER's design addresses each of your two questions, and then verify both with experiments.

SABER does not reject correct context beyond its PK, because its CK belief is learned from whether the CK answer is actually correct. SABER does not act on the model's verbal verdict. It reads hidden states over the reasoning span, and the CK-side trace is generated under a prompt that instructs the model to assume the document is reliable and reason only from it. These states therefore reflect how the model reasons with the document. The CK predictor maps them to a belief trained on yCKy_{CK}, the observed correctness of the CK answer. Correct contexts beyond the model's PK are common in our benchmark, where the instances with only the CK answer correct make up 35.4–37.3% of the test set on each backbone. A predictor that rejected such contexts would be penalized on all of them. Therefore, the supervision itself teaches SABER to trust them.

SABER separates missing knowledge in the context from being misled by its own PK, because the CK predictor also sees how reliable the model's own knowledge is. The CK predictor takes the self-prior together with the CK-side representation, and the self-prior reflects how much the model knows about the question, and thus how much it needs external evidence. The CK predictor therefore knows when the model's own knowledge is weak, and it can discount a conflict between the document and such knowledge. Together with the PK belief, the 4-cell decision separates the situations you describe.

(1) The context supplies knowledge beyond the model's PK (C01C_{01}). The PK answer is wrong and the CK answer is right, so SABER should answer with the context. This is where being misled by its own PK would cause harm.

(2) The context misleads while the model's PK is right (C10C_{10}). SABER should keep the PK answer.

(3) The context lacks the knowledge and the model's PK is also wrong (C00C_{00}). SABER should abstain.

In both (1) and (3), the model's PK is wrong, so a judgment driven by the model's own knowledge cannot tell them apart. The only difference is whether the context contains the needed knowledge, which is exactly what the CK belief estimates.

AuthorsOpenReview ↗

Rebuttal by Authors (2/5)

W1 (contined). Whether the own-knowledge self-evaluation prompt leads SABER to reject correct context beyond its PK, and how SABER distinguishes knowledge missing from the context from being misled by its own PK

To test whether SABER rejects correct context beyond its PK, as you ask, we compare it with the two baselines in our paper that rely on the model's verbal self-judgment. Context-Eval asks the model whether the retrieved passage is reliable before answering, which is the context-side evaluation you describe. Internal-Eval asks the model whether its own PK answer is correct and chooses between this answer and the context accordingly.

Results. We report CF, the accuracy on the C01C_{01} instances, averaged over the five data sources in Tables 1 and 2. The results are shown below.

BackboneContext-Eval (%) ↑Internal-Eval (%) ↑SABER (%) ↑
Llama-3.1-8B49.557.585.3
Qwen-2.5-7B69.475.393.3
Llama-3.2-3B65.867.091.7
Qwen-2.5-3B26.867.594.4

From the table, we observe that SABER's CF is 18.0–67.6% higher than that of both verbal baselines on every backbone. This is because a verbal judgment that consults the model's own knowledge tends to reject a context that contradicts this knowledge even when the context is correct. In contrast, SABER's CK belief is supervised by the actual correctness of the CK answer. Therefore, the risk you raise affects verbal self-evaluation, and SABER avoids it.

To test whether SABER distinguishes the three situations, as you ask, we examine SABER's decisions in each cell on the test set, under the abstention setting of Table 3 (τ=0.5\tau=0.5). In 94.1–97.6% of the C01C_{01} instances, the context contains the gold answer, so these instances are correct contexts beyond the model's PK.

Results. We report the share of each decision in each cell, as ranges over the four backbones. The results are shown below.

CellAnswer with CK (%)Answer with PK (%)Abstain (%)
C01C_{01}81.8–86.91.5–5.511.1–12.7
C10C_{10}9.8–14.446.1–62.028.2–40.5
C00C_{00}16.6–20.95.2–12.968.7–78.3

From the table, we have the following observations.

(1) SABER follows correct context beyond its PK. On C01C_{01}, it answers with the context in 81.8–86.9% of cases and falls back to its own knowledge in only 1.5–5.5%. The rest are abstentions, which produce no wrong answer. This is because the CK belief is supervised by the correctness of the CK answer. Therefore, the self-evaluation does not lead SABER to reject such context.

(2) SABER separates a context that lacks the knowledge from a correct context that conflicts with wrong PK. The PK answer is wrong in both C01C_{01} and C00C_{00}, yet SABER answers with the context in 81.8–86.9% of the C01C_{01} instances and in only 16.6–20.9% of the C00C_{00} instances, where it mostly abstains (68.7–78.3%). On C10C_{10}, it keeps the PK answer in 46.1–62.0% of cases and follows the misleading context in only 9.8–14.4%. This is because the CK belief estimates whether the context supports a correct answer, while the PK belief separately estimates whether the model's own knowledge does. Therefore, the three situations you describe receive three different decisions.

AuthorsOpenReview ↗

Rebuttal by Authors (3/5)

W2. SABER's contribution beyond supervised correctness probing, a comparison with a simple candidate-answer classifier, and ablations of self-evaluation and the feature branches

We thank you for these insightful questions. Below we first clarify SABER's contribution beyond supervised correctness probing, then compare SABER with the simple classifier you describe, and finally report the ablations you request.

Supervised probing is only the tool SABER uses to read hidden states. SABER's contribution lies in what it estimates and how it decides. Standard correctness probing estimates whether one given answer is correct. SABER instead formulates knowledge conflict as two separate correctness questions, one for the PK path and one for the CK path, and labels both paths for each backbone in our model-specific benchmark. It reads the reliability of each path from the states after the model reasons under that source, and it combines the two beliefs in a 4-cell decision that answers with PK, answers with CK, or abstains. We verify this contribution below.

To test whether the two-path design adds to probing a single answer, as you ask, we keep SABER's own predictions behind Table 3 and change only the decision rule. A single-path rule corresponds to standard correctness probing. It answers from one path when its belief reaches τ=0.5\tau=0.5 and abstains otherwise.

Results. We report the Score and abstention F1 of Table 3 on the test set. The results are shown below.

BackboneDecision ruleScore (%) ↑F1 (%) ↑
Llama-3.1-8BPK path only13.059.9
CK path only39.371.5
Two paths (SABER)43.972.5
Qwen-2.5-7BPK path only10.862.7
CK path only38.375.3
Two paths (SABER)43.478.0
Llama-3.2-3BPK path only7.963.8
CK path only34.975.9
Two paths (SABER)38.576.8
Qwen-2.5-3BPK path only5.764.9
CK path only37.575.9
Two paths (SABER)39.676.6

From the table, we observe that with identical predictions, the two-path rule outperforms the better single-path rule on every backbone, by 2.1–5.1% in Score and 0.7–2.7% in F1. This is because a single-path rule can only accept or reject one answer. When that answer is wrong, it cannot fall back to the other path, and it cannot recognize the cases where both paths fail. Therefore, the gain comes from the two-path formulation rather than from probing itself.

SABER outperforms a simple classifier trained on our labels, because such a classifier judges only one answer. To compare SABER with the simple classifier you describe, we train an MLP on the hidden state of the model as it generates its CK answer, without reasoning or self-evaluation, to predict whether this answer is correct. As in standard correctness probing, it keeps the CK answer when the answer is judged correct and otherwise falls back to the PK answer, or abstains in the setting of Table 3.

Results. We report Acc, MFS, and the Score of Table 3 on the test set. The results are shown below.

BackboneMethodAcc (%) ↑MFS (%) ↑Score (%) ↑
Llama-3.1-8BSimple classifier63.181.639.4
SABER65.983.643.9
Qwen-2.5-7BSimple classifier57.285.636.7
SABER61.088.143.4
Llama-3.2-3BSimple classifier57.283.035.2
SABER60.885.938.5
Qwen-2.5-3BSimple classifier53.884.332.7
SABER57.082.939.6

From the table, we observe that SABER achieves higher Acc and Score than the simple classifier on every backbone, by 2.8–3.8% and 3.3–6.9% respectively, as well as a higher average MFS. This is because the classifier estimates only whether the CK answer is correct. When it rejects this answer, it cannot tell whether the PK answer is reliable, so it either switches to the PK answer blindly or abstains. SABER instead estimates both paths and chooses between them. Therefore, SABER's contribution beyond correctness probing lies in its two-path estimation and decision.

Reviewer_BnkFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work investigates how retrieval-augmented LLMs select between parametric and contextual knowledge, or decline to answer when neither source is trustworthy. The authors build a benchmark for judging answer-path validity and introduce SABER. The method trains reliability predictors using query-only self-prior, reasoning and self-evaluation representations, while keeping the base LLM frozen. Experiments examine conflict resolution, abstention behaviour, individual module contributions, and cross-model generalisation.

Critical strengths
  1. The paper addresses a practically important problem: determining when the model should rely on parametric knowledge, retrieved context, or refrain from answering entirely.
  2. It unifies answer selection and abstention within one framework, without requiring fine-tuning of the underlying language model.
  3. Empirical results demonstrate improved performance on conflict handling and abstention across multiple backbone models.
Critical weaknesses
  1. The model can still use parametric knowledge when contextual knowledge is provided. It may therefore answer correctly even when the context is wrong. A correct answer does not establish reliance on contextual knowledge. Claims about contextual faithfulness need evidence that the supplied context supports these answers.
  2. The query-only hidden state may simply reflect question difficulty or familiarity. Predicting correctness from this state does not establish that the model is assessing its own knowledge. The paper also needs to show what the later parametric knowledge representation adds. An ablation should retain the self-prior and contextual knowledge features and remove only the later representation.
  3. Different backbones use different extraction layers, so a new backbone may require another layer search. The layer analysis shows that overall accuracy changes little across nearby layers. It does not show how the same layer changes affect conflict handling or abstention. The paper should report the search cost and test whether one selection rule works across backbones.
  4. The benchmark labels answer correctness through alias matching and GPT-4o-mini verification. Only answers accepted by the matcher are sent to GPT-4o-mini. Any correct answer missed by matching remains incorrectly labeled. These labels train the predictors and define the evaluation categories. Additional checks are needed to ensure the accuracy of these labels.
Questions for authors

Please refer to the Weaknesses section.

6 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/4)

We sincerely thank you for your valuable comments and insightful suggestions, and for recognizing the practical importance of our problem, our unified framework for answer selection and abstention without fine-tuning the LLM, and its consistent improvements across backbones. Below we address each of your questions in turn.

W1. Evidence that the supplied context supports the answers counted in context faithfulness (CF)

We thank you for this insightful point. Below we first show how our model-specific labels keep answers from parametric knowledge out of CF, and then give direct evidence that the context supports the answers counted in CF.

Because the labels reflect each model's own behavior, a correct answer that comes from parametric knowledge never enters CF. For each backbone, every instance carries two labels, namely whether the model's answer is correct without the context (yPKy_{PK}) and with the context (yCKy_{CK}). These labels separate the cases you describe into three cells.

(1) C01C_{01}, counted in CF. The model answers wrongly without the context and correctly with it. Its own knowledge is insufficient, so the correct answer depends on the context.

(2) C11C_{11}, not counted in CF. The model answers correctly on both paths, so the CK answer may come from parametric knowledge, as you point out. However, both answers are correct here, so the source of the CK answer does not affect any metric of conflict handling.

(3) C10C_{10}, measured by KF. When the context is wrong and the model knows the answer but is misled by the context, the instance belongs to C10C_{10}, and KF measures whether a method keeps the correct parametric answer.

Therefore, CF measures context faithfulness only where the context is the source of the correct answer. In contrast, existing conflict-QA datasets label a context as supportive or conflicting only at the dataset level, so they cannot tell whether a correct answer comes from the context or from the model's own knowledge.

The supplied context supports the answers counted in CF. To test whether the supplied context supports these answers, as your question asks, we measure how often the context contains the gold answer, with any of its aliases as a whole-word match.

Results. We report this rate in C01C_{01}, in C10C_{10}, and over all test instances for each backbone. The results are shown below.

BackboneC01C_{01} (%)C10C_{10} (%)All (%)
Llama-3.1-8B94.129.459.1
Qwen-2.5-7B97.623.159.1
Llama-3.2-3B97.027.759.1
Qwen-2.5-3B97.427.259.1

From the table, we observe that the context contains the gold answer in 94.1–97.6% of C01C_{01} instances but only in 23.1–29.4% of C10C_{10} instances, compared with 59.1% over all test instances. This is because the correct answers in C01C_{01} come from the context, whereas in C10C_{10} the context mostly does not support the gold answer and leads the model away from its correct knowledge. Therefore, the answers counted in CF are supported by the supplied context, as your comment requires, and in the cases counted in KF, the context mostly does not support the gold answer.

Revision. In Appendix D of the revised PDF, we state this definition of CF explicitly and add the table above.

AuthorsOpenReview ↗

Rebuttal by Authors (2/4)

W2. Whether the self-prior only reflects question difficulty, and what the later PK representation adds

We thank you for these two insightful points. Below we first explain why the query-only state reflects the model's estimate of its own knowledge, and then report the ablation you requested, which shows what the later PK representation adds.

The self-prior estimates whether the model knows the answer, and question difficulty and familiarity are part of this estimate. We read the self-prior from a forward pass on the question alone, before the model generates any answer, in the spirit of P(IK) [1]. For a given model, a question is difficult or unfamiliar precisely when the model does not know the answer. Predicting correctness from this state thus reflects the model's assessment of its own knowledge, of which the difficulty and familiarity of the question to the model are a natural part. However, the self-prior is only the part of the judgment made before answering, and whether a question is answered correctly is not a property of the question alone. Our benchmark labels each instance by the observed behavior of each model, and on the same test questions, the four backbones do not all receive the same yPKy_{PK} for 30.6% of questions. Therefore, SABER does not rely on the self-prior alone, and it adds the PK-side conditional representation, which is read after the model reasons from its own knowledge.

The later PK representation improves both SABER's judgment of the PK answer and its final accuracy on every backbone. To show what the later PK representation adds, as you suggest, we keep the self-prior and the CK-side features and remove only the PK-side conditional representation, so that the PK predictor uses the self-prior alone. We retrain the full model and this variant in the same run under the same protocol.

Results. We report the change caused by the removal (Δ, variant minus full model) in three metrics, namely the AUROC of the PK predictor for yPKy_{PK}, which measures how accurately it judges whether the PK answer is correct, the answer accuracy (Acc) of Tables 1 and 2, and the abstention Score of Table 3. The results are shown below.

BackboneΔAUROC (%)ΔAcc (%)ΔScore (%)
Llama-3.1-8B−1.5−1.0−0.7
Qwen-2.5-7B−2.4−0.4−0.9
Llama-3.2-3B−1.5−0.1−0.3
Qwen-2.5-3B−0.5−0.50.0

From the table, we have the following observations.

(1) The later PK representation adds information about whether the PK answer is correct. Without it, the AUROC of the PK predictor drops by 0.5–2.4% on every backbone. This is because the self-prior is read before the model answers, whereas the PK-side representation is read after the model reasons from its own knowledge and reflects whether this reasoning holds up. Therefore, the later representation provides information that the query-only state does not contain.

(2) This information improves SABER's final decisions. Without it, Acc drops by 0.1–1.0% on every backbone, and the abstention Score decreases or stays the same on every backbone. This is because a less accurate PK belief makes SABER more often choose a wrong PK answer or discard a correct one. Therefore, the self-prior and the later PK representation are complementary, and SABER needs both.

Revision. We add this ablation to Appendix E.6.

References

[1] Kadavath et al. Language Models (Mostly) Know What They Know. arXiv:2207.05221, 2022.

AuthorsOpenReview ↗

Rebuttal by Authors (3/4)

W3. Layer selection across backbones, its effect on conflict handling and abstention, and its search cost

We thank you for these practical questions. Below we first show that the same layer changes affect conflict handling and abstention very little, then show that a single depth rule matches the searched layers, so that a new backbone needs no layer search, and finally report the cost of the search when it is run.

The same layer changes affect conflict handling and abstention very little. Figure 5 and Figure 7 (Appendix) show that on every backbone the selected layer pair lies inside a broad high-accuracy plateau rather than at an isolated peak. As you point out, however, accuracy alone does not show how conflict handling and abstention respond. To test this, we retrained SABER at the selected layer pair and at eight neighboring pairs, obtained by moving ℓp\ell_p or ℓc\ell_c by ±2 and ±4 layers.

Results. We measure on the test set the range (maximum minus minimum) of MFS and of the Table 3 Score and F1 across these nine layer pairs. The results are shown below.

BackboneMFS range (%) ↓Score range (%) ↓F1 range (%) ↓
Llama-3.1-8B3.51.21.2
Qwen-2.5-7B5.42.41.7
Llama-3.2-3B4.51.90.7
Qwen-2.5-3B3.11.21.0

From the table, we observe that across the nine layer pairs, the Score varies by at most 2.4% and F1 by at most 1.7% on every backbone, and MFS stays within a range of 3.1–5.4%. This is because neighboring layers encode similar information about whether each answer is correct, so SABER's decisions change little. Therefore, SABER's conflict handling and abstention do not depend on a finely tuned layer.

A single 50%-depth rule works across backbones, so a new backbone needs no layer search. To test the single selection rule you asked about, we set both ℓp\ell_p and ℓc\ell_c to 50% of the model depth without any search, and retrained SABER in the same run as at the searched layers.

Results. We report the difference on the test set (Δ, rule minus searched layers). The results are shown below.

BackboneΔAcc (%)ΔMFS (%)ΔScore (%)ΔF1 (%)
Llama-3.1-8B−0.4−1.3−0.3−0.7
Qwen-2.5-7B−1.0−1.5−0.4−0.1
Llama-3.2-3B−0.5−0.7−0.1+0.3
Qwen-2.5-3B0.0−1.8−0.6+0.4

From the table, we observe that the rule stays within 1.0% in Acc, 1.8% in MFS, 0.6% in Score, and 0.7% in F1 of the searched layers on all four backbones. The differences in MFS, Score, and F1 are all smaller than their variation among neighboring layers in the previous table. This is because factual and truthfulness information is most linearly accessible in intermediate layers [1, 2, 3], and SeaKR [4] likewise reads its uncertainty signal from the middle layer. Therefore, a new backbone can use SABER with this rule and without a layer search.

Even when the search is run, it costs little, because it never reruns the LLM. The hidden states of all layers are cached in the single extraction pass that SABER needs anyway, and evaluating a layer pair only retrains the two small MLP predictors, which takes about half a minute. Therefore, the full stride-2 grid costs 2.1–3.1 GPU hours per backbone, and it parallelizes trivially.

Revision. We add the search cost, the layer-robustness table with conflict and abstention metrics, and the results of the 50%-depth rule to Appendix E.5.

References

[1] Azaria and Mitchell. The Internal State of an LLM Knows When It's Lying. arXiv:2304.13734, 2023.

[2] Burns et al. Discovering Latent Knowledge in Language Models Without Supervision. arXiv:2212.03827, 2022.

[3] Chen et al. INSIDE: LLMs' Internal States Retain the Power of Hallucination Detection. ICLR, 2024.

[4] Yao et al. SeaKR: Self-aware Knowledge Retrieval for Adaptive Retrieval Augmented Generation. arXiv:2406.19215, 2025.

Reviewer_XJiXOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses knowledge conflicts between parametric knowledge (PK) and retrieved contextual knowledge (CK) in retrieval-augmented generation (RAG). It formulates the problem as independently determining the correctness of answers derived from PK and CK, and establishes a benchmark across five conflict-QA datasets and four LLMs. The proposed method, SABER, extracts query-only representations (self-priors) and self-evaluation representations from pathway-specific reasoning traces using the hidden states of a frozen LLM. A lightweight predictor then selects between PK, CK, and abstention. SABER outperforms baseline methods, with particularly substantial improvements on datasets exhibiting frequent knowledge conflicts.

Critical strengths

S1. Clear Problem Formulation

The paper explicitly incorporates cases where both knowledge pathways produce incorrect answers into the decision space, naturally introducing abstention rather than forcing the model to select one of the two knowledge sources. This provides a distinct perspective from existing source-arbitration approaches.

S2. Construction of Model-Specific Ground-Truth Benchmarks

Existing conflict-QA datasets define whether retrieved contexts are supportive or contradictory only at the dataset level. However, what a model knows or does not know varies across models. To address this, the authors separately label the correctness of PK and CK pathways for each of four backbone models with different architectures and sizes. This enables training and evaluation based on actual model behavior. Furthermore, the proposed metrics, CF, KF, and MFS, reveal limitations of accuracy alone, which may make methods that consistently favor a single knowledge source appear effective.

Critical weaknesses

W1. Lack of comparison with methods based on LLM internal representations

SABER trains a lightweight predictor on the LLM’s hidden states, yet most baselines are prompt-based, and none of them leverages internal representations. To show that SABER’s gains stem from its specific design rather than from the use of hidden states per se, a comparison with internal-representation-based methods is needed. Specifically, adding comparisons with the original TACS [1], which actually trains a classifier on hidden states, and with SeakR [2] would help clarify the effectiveness of SABER’s design.

W2. Missing baseline in the abstention experiment (Table 3)

R-Tuning [3], which is included in Tables 1 and 2, is not included in Table 3. Since R-Tuning is explicitly trained to respond with “I don’t know,” it is a natural baseline for the abstention setting. It would be helpful if the authors could clarify why it was not included, and including it in Table 3 would provide a more comprehensive evaluation of SABER’s abstention performance.

W3. Effectiveness of the Core Contribution

The authors present the decomposition of self-awareness into two components, the self-prior and conditional representations, as a key contribution. However, the ablation results show that adding the self-prior improves MFS by only 0.5–1.2 points, and even degrades performance for two of the four backbones in the single-trace setting (e.g., −3.0 points for Qwen-2.5-3B). This may be because the self-evaluation input already includes the question, making the information captured by the self-prior redundant with the conditional representations. Furthermore, the fact that a very shallow layer (lp=2l_p=2) is selected for the self-prior in Qwen-2.5-3B suggests that the prior signal itself may be weak.

[1] Yu, Tian, Shaolei Zhang, and Yang Feng. "Truth-aware context selection: Mitigating hallucinations of large language models being misled by untruthful contexts." Findings of the Association for Computational Linguistics: ACL 2024. 2024.

[2] Yao, Zijun, et al. "Seakr: Self-aware knowledge retrieval for adaptive retrieval augmented generation." Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

[3] Zhang, Hanning, et al. "R-tuning: Instructing large language models to say ‘i don’t know’." Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Questions for authors
  1. Prior work has already addressed abstention, including in the RAG setting [3, 4]. How does SABER differ from these approaches, and why does this difference make it more effective for abstention?

  2. Could the predictor be learning artifacts from the construction of counterfactual contexts in the datasets when predicting the correctness of the CK pathway? How accurately can a context-only classifier predict yCKy_{CK} without access to the question or reasoning traces?

  3. Why does adding the self-prior degrade performance in the single-trace setting?

  4. SABER generates multiple reasoning traces and requires additional forward passes per query, which may increase its online inference cost. Could the authors compare the online inference cost and performance of SABER with those of the other methods?

  5. The predictor is trained on the correctness of short answers generated without reasoning, whereas its inputs are extracted from reasoning traces generated separately with different prompts. Since the correctness of answers with and without reasoning may differ, I wonder whether SABER remains reliable when such discrepancies occur. Could the authors also report how often the two agree in correctness?

[3] Zhang, Hanning, et al. "R-tuning: Instructing large language models to say ‘i don’t know’." Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

[4] Joren, Hailey, et al. "Sufficient context: A new lens on retrieval augmented generation systems." International Conference on Learning Representations. Vol. 2025. 2025.

8 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/6)

We sincerely thank you for your valuable comments and insightful suggestions, and for recognizing our clear problem formulation with abstention, our model-specific benchmark, and the CF, KF, and MFS metrics. Below we address each of your questions in turn.

W1. Comparison with internal-representation baselines (original TACS and SeaKR)

We thank you for this insightful suggestion. Below we first compare SABER with the original TACS and SeaKR under the same protocol, and then report a controlled experiment that keeps the hidden states fixed and removes only the two-path design.

SABER's gains come from its design, not from the use of hidden states per se. We added the original TACS [1] and SeaKR [2], two methods that also use internal representations, and ran them under exactly the same protocol as SABER. SABER achieves the best MFS in all 20 backbone–dataset settings, ahead of the strongest of them by 12.8% MFS on average.

Setup. All methods are trained on the union of the seven training splits, with the same data split and hyperparameters selected on the validation set. (a) Original TACS. As in the original paper, classifiers on hidden states judge whether the context is truthful, and the untruthful parts are masked before generation, at the token level (TACS-T) or the sentence level (TACS-S). (b) SeaKR. SeaKR measures uncertainty by the consistency of hidden states across sampled answers. We apply its rule for using retrieved knowledge, which returns the PK answer when the uncertainty of the closed-book answer is below a threshold tuned on the validation set, and the CK answer otherwise.

Results. We report the metrics of Tables 1 and 2 on the test set, averaged over the four backbones and five data sources. The results are shown below.

MethodAccCFKFMFS
TACS-P47.583.014.448.7
TACS-T53.495.423.559.4
TACS-S55.794.149.371.7
SeaKR54.875.469.372.3
SABER61.291.279.185.1

From the table, we observe that SABER achieves the highest Acc, KF, and MFS, and it is the only method that keeps both CF and KF high. This is because TACS judges only the context and has no estimate for the PK answer, so it often fails to keep the model's correct answer when the context misleads (KF 49.3 for TACS-S). SeaKR decides from the uncertainty of the model's own answer alone and has no estimate for the context, so a single threshold trades CF for KF (CF 75.4). SABER instead predicts the correctness of both answers. Therefore, using hidden states alone does not explain SABER's gains.

Removing the two-path design while keeping the hidden states loses much of SABER's gain. To separate the design from the hidden states, as your question asks, we keep SABER's predictions behind Table 3 and change only the decision rule to a single path, which answers from that path when its belief reaches τ=0.5\tau=0.5 and abstains otherwise. We report the Score / abstention F1 of Table 3. The results are shown below.

Decision ruleLlama-3.1-8BQwen-2.5-7BLlama-3.2-3BQwen-2.5-3B
PK path only13.0 / 59.910.8 / 62.77.9 / 63.85.7 / 64.9
CK path only39.3 / 71.538.3 / 75.334.9 / 75.937.5 / 75.9
Two paths (SABER, Table 3)43.9 / 72.543.4 / 78.038.5 / 76.839.6 / 76.6

From the table, we observe that with identical predictions, the two-path rule outperforms the better single-path rule on every backbone, by 2.1–5.1% in Score and 0.7–2.7% in F1. This is because a single-path rule can only accept or reject one answer and cannot fall back to the other path when that answer is wrong. Therefore, SABER's gain comes from its two-path design, not from the hidden states.

Revision. We add the original TACS and SeaKR to Tables 1 and 2 as a new category of internal-representation methods (Section 4), with their descriptions in Appendix C.1, and add the single-path control to Section 4.2 (Table 4).

References

[1] Yu et al. Truth-Aware Context Selection: Mitigating Hallucinations of Large Language Models Being Misled by Untruthful Contexts. arXiv:2403.07556, 2024.

[2] Yao et al. SeaKR: Self-aware Knowledge Retrieval for Adaptive Retrieval Augmented Generation. arXiv:2406.19215, 2025.

AuthorsOpenReview ↗

Rebuttal by Authors (2/6)

W2. R-Tuning as a baseline in the abstention experiment (Table 3)

We thank you for this suggestion. Below we add R-Tuning to Table 3 on all four backbones, report the results, and explain the gap.

Setup. Following R-Tuning's original definition, the model's own answer is labeled sure if correct and unsure otherwise. R-Tuning is trained with LoRA on our RAG prompt to append this tag, using SABER's training split, and abstains when P(sure) < 0.5, the same default threshold as SABER. Both methods are scored on the same answers and labels, so the comparison isolates the abstention decision.

Results. We report the Table 3 metrics on the test set at τ = 0.5. The results are shown below.

BackboneMethodScore ↑CovR_C ↓F1 ↑
Llama-3.1-8BR-Tuning36.453.616.172.6
SABER43.965.016.272.5
Qwen-2.5-7BR-Tuning35.450.615.074.8
SABER43.458.512.978.0
Llama-3.2-3BR-Tuning30.347.418.074.7
SABER38.558.617.276.8
Qwen-2.5-3BR-Tuning28.647.119.772.9
SABER39.660.017.076.6

SABER answers more questions and still achieves a higher Score on every backbone. Its coverage is 7.9–12.9% higher than R-Tuning's and its Score 7.5–11.0% higher. Answering more usually raises the risk on answered questions. However, SABER's risk is 0.8–2.7% lower on three backbones and on par on Llama-3.1-8B, and its abstention F1 follows the same pattern. Therefore, SABER answers more questions without a higher risk, and it abstains more precisely.

The gap comes from the two-path decision. R-Tuning can only accept or reject one answer, so its Score is capped by that answer's accuracy. SABER instead estimates the correctness of both the PK and the CK answer, so it can fall back to the model's own knowledge when the context is wrong. Where only the PK answer is correct, SABER answers with it 46.1–62.0% of the time, whereas R-Tuning can only abstain or output the wrong CK answer. Furthermore, SABER keeps the backbone frozen and trains only two small MLPs, whereas R-Tuning fine-tunes the LLM.

Revision. We add R-Tuning to Table 3 and its setup to Appendix C.1.

W3 & Q3. Contribution of the self-prior beyond the conditional representations, including the single-trace setting and the layer choice

We thank you for these insightful questions. Below we first explain what the self-prior captures beyond the conditional representations, and then show how this accounts for the size of the gain, the single-trace result (Q3), and the shallow layer.

The self-prior and the conditional representations capture different information, even though both contain the question. The self-prior is the model's pre-answer estimate of whether it knows the answer, read in the spirit of P(IK) [1] from a forward pass on the question alone, with no trace and no context. The conditional representations are instead read after a trace under a given source, so they reflect how the model reasons with that source, and the CK-side state is further shaped by the context. This matters most under knowledge conflict, because a misleading context can make the model look confident. Our SeaKR experiment (W1) shows this directly. Whether or not the context is correct, answers sampled with it are much more consistent than those without it, and on Llama-3.1-8B the average number of distinct answers drops from 13.0 to 6.3. Since the self-prior never sees the context, it gives both predictors an independent anchor.

Because this information is complementary, the self-prior consistently improves SABER. In the full configuration (K=3K=3), it improves MFS by 0.5–1.2% on all four backbones.

This complementarity relies on a reliable conditional state, which explains the single-trace setting (Q3). A single trace is one stochastic sample that may hit a dead end or react to surface features of the context (Section 3.2), so its conditional state is noisy, and adding the self-prior on top of it can hurt. For this reason, SABER aggregates multiple traces. Averaging reduces the noise and benefits all backbones, most on the smallest model, raising MFS by 3.2% on Qwen-2.5-3B (from P+C1P{+}C_1 to SABER). Once the conditional state is reliable, the self-prior improves all four backbones.

Finally, the shallow layer ℓp=2\ell_p=2 on Qwen-2.5-3B reflects robustness to the layer choice, not a weak signal. It lies inside a broad high-accuracy plateau of the layer grid (Appendix E.4). Therefore, the self-prior signal is available across many depths, and SABER does not depend on fine-grained layer tuning.

Revision. We add the ablation of the side-specific conditional representations to Appendix E.6.

References

[1] Kadavath et al. Language Models (Mostly) Know What They Know. arXiv:2207.05221, 2022.

AuthorsOpenReview ↗

Rebuttal by Authors (3/6)

Q1. How SABER differs from prior abstention work (R-Tuning, Sufficient Context) and why this helps abstention

We thank you for this insightful question. Below we first explain how SABER differs from these approaches, then why this difference makes its abstention more effective, and finally show the evidence from our experiments.

R-Tuning and Sufficient Context judge a single answer, whereas SABER judges both. Under knowledge conflict, a RAG system should abstain only when neither path, PK or CK, gives a correct answer. When one path fails, the system therefore needs to know whether the other path is reliable, in order to decide between switching to it and abstaining. R-Tuning [1] learns to express uncertainty on the questions whose answer is wrong, and in its original setting this answer is the closed-book PK answer. It never assesses the CK candidate, so it cannot tell when the retrieved context would correct the error. Sufficient Context [2] decides whether to keep the RAG answer from context sufficiency and the model's confidence in that answer. It never assesses the PK candidate, so it cannot tell when the closed-book answer is still correct.

Judging both answers lets SABER switch when one path fails and abstain only when both fail. SABER predicts the correctness of the PK answer and of the CK answer separately. The higher belief decides which answer to give, and whether it passes the threshold decides whether to answer at all. This reduces both types of error under knowledge conflict. First, SABER makes fewer false abstentions, because when one path fails but the other passes the threshold, it still answers with the latter. Second, it makes fewer false answers, because it abstains whenever both paths fall below the threshold.

To test both effects, we report SABER's decisions on the test set in the abstention setting of Table 3. For each backbone, we give the share of C01C_{01} instances (only the context is right) that SABER answers with the context, the share of C10C_{10} instances (only the model's own answer is right) where it keeps its own answer, and the share of C00C_{00} instances (neither is right) where it abstains. The results are shown below.

BackboneContext in C01C_{01} (%) ↑Own answer in C10C_{10} (%) ↑Abstain in C00C_{00} (%) ↑
Llama-3.1-8B81.862.068.7
Qwen-2.5-7B86.150.278.3
Llama-3.2-3B86.946.174.3
Qwen-2.5-3B86.752.771.8

From the table, we observe that SABER follows the context when only the context is right, keeps its own answer when only that answer is right, and abstains when neither is right. This is because every decision uses the correctness estimates of both paths, whereas a method that judges only one answer cannot recover both C01C_{01} and C10C_{10}. Therefore, judging both answers makes SABER's abstention more effective, and its risk–coverage curve dominates every prompt-based abstainer on all four backbones (Figure 3).

Revision. We add this comparison to the Related Work section (Section 5) of the revised PDF.

References

[1] Zhang et al. R-Tuning: Instructing Large Language Models to Say 'I Don't Know'. NAACL, 2024.

[2] Joren et al. Sufficient Context: A New Lens on Retrieval Augmented Generation Systems. ICLR, 2025.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Response by Authors

Dear Area Chairs and Reviewers,

We sincerely thank the reviewers for their time, insightful comments, and constructive suggestions. The rebuttal and discussion phase has been highly productive, allowing us to clarify our work and conduct new experiments. With these constructive suggestions, we believe the paper has been significantly strengthened in both empirical rigor and presentation.

Strengths Acknowledged by the Reviewers

  1. Novel and practically important problem formulation. By explicitly including the cases where both knowledge paths fail and thereby naturally introducing abstention, SABER offers a distinct perspective beyond existing source-arbitration approaches, addressing the practically important question of which source to trust or whether to abstain (Reviewers XJiX, BnkF, Z2YL).
  2. Principled model-specific benchmark and insightful metrics. Labeling the correctness of both knowledge paths for each backbone, including the cases where both answers are wrong, enables training and evaluation grounded in actual model behavior of LLMs, and the proposed CF, KF, and MFS reveal limitations that accuracy alone hides (Reviewers XJiX, Z2YL).
  3. Lightweight, unified, and practical framework. SABER decides among answering with parametric knowledge, answering with the context knowledge, and abstaining in a single framework, by training only two small heads on the frozen backbone's hidden states, so it applies to open-weight models without fine-tuning the LLM (Reviewers BnkF, Z2YL).
  4. Strong and consistent empirical improvements. SABER achieves better conflict handling and more effective abstention than the baselines across four backbones from two model families and sizes from 3B to 8B, demonstrating the effectiveness of its design (Reviewers BnkF, Z2YL).

Main Concerns Raised by the Reviewers

  1. Additional comparisons. Comparisons with methods that also use internal representations, such as TACS and SeaKR, and with a simple classifier on the candidate answer were suggested to further clarify the source of SABER's gains, along with R-Tuning as an additional abstention baseline (Reviewers XJiX, Z2YL).
  2. Component ablations. Further ablations of the self-prior and the side-specific conditional representations were requested (Reviewers XJiX, BnkF).
  3. Further validation of labels and beliefs. Further support was requested for the label accuracy, for the reliance of correct context-conditioned answers on the context, and for the self-assessment interpretation of the self-prior (Reviewer BnkF).
  4. Scalability to larger models. Further evidence was requested that the findings generalize to larger models (Reviewer Z2YL).

Revisions in the Rebuttal and Discussion Phase

  1. Comparisons with internal-representation methods. We compared SABER with TACS, SeaKR, a simple candidate-answer classifier, and single-path decision rules under the same protocol, showing that SABER outperforms them across backbones and that its gains come from the two-path design rather than from the use of hidden states.
  2. Component ablations. We ablated each component in turn, showing that the side-specific, reasoning-based representations are essential to conflict handling, and clarified the self-prior as a complementary auxiliary signal.
  3. Label and interpretation analyses. We measured the answers missed by alias matching and audited the labels with a stronger judge and, in the discussion phase, conducted an answer-masking test and a cross-model self-prior analysis, showing that the labels are highly accurate, that contextual faithfulness reflects reliance on the context, and that the self-prior reflects each model's own knowledge boundary.
  4. Abstention comparison. We added R-Tuning for abstention comparison, showing that SABER answers more questions with a higher Score while keeping the backbone frozen and making the appropriate decision in each case.
  5. Efficiency and scalability. We showed that SABER is efficient online, robust to the layer choice without layer search, and remains the best among all methods on a new 14B backbone, with gains that grow with model size.

Final Remarks

We believe the additional experiments and clarifications address the main concerns, while the acknowledged core contributions remain strong. We are grateful that Reviewer XJiX confirmed that our response addressed most concerns and raised the score, that Reviewer BnkF explicitly considered our empirical contribution sufficient for acceptance to the conference, and that Reviewer Z2YL acknowledged that our thorough explanations and additional experiments addressed the concerns. We are committed to continuously improving this paper. We once again sincerely appreciate the reviewers' recognition, suggestions, and encouragement.

Sincerely,

Authors of Submission 3454

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.