- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3(+1)
Read review
Paper summary
The paper studies fine-grained attribution in RAG (attributing a unit of context document that support an answer span). Previous methods tend to come as a post-hoc intervention, such as by reading attention scores of a frozen LLM or by training a small estimator on top of it. This paper fine-tunes the LLM itself, such that a fixed set of preselected attention heads attributes better. The paper uses two terms: the generation model writes the answer and the attribution model scores the evidence. The paper separates the two model and only the attribution model is trained.
The proposed method, Training Internal Representations (TIR), has two parts:
- a pairwise hinge loss whose negatives are weighted by a softmax over their scores (to focus on hard negatives);
- a stop-gradient on the value-norm factor of the MV score (to solve "value-norm cheating")
TIR updates every layer up to the deepest selected head (17 layers of Llama-3.1-8B), using LongCite-45K truncated to 10K tokens. The experiments cover Llama-3.1-8B and Qwen3-4B, four scoring signals, and evaluation on LongCite, HotpotQA and VERI-GRAN under a Recall@GoldK protocol.
Critical strengths
The paper provides interesting findings:
- S1: The learning to attribute is an interesting approach.
- Current single-pass attribution methods treat a model’s attention as a fixed signal (post-hoc on a frozen LLM). Progress in this post-hoc method comes from choosing a better head or a better post-processing rule. This paper shows that we can train attributability via supervision.
- The gains seem to be large enough to change the practical default. On Llama, training lifts the four untrained scoring signals by 3.2 to 18.3 points R@GoldK (Table 4). The MV score gains 6 to 7 points on both Llama-3.1-8B and Qwen3-4B (Table 1). All numbers come from the final optimizer step, with no checkpoint selection on evaluation data (A.2).
- BlockRank and QRRanker (see W3 below) train attention heads in a similar way for passage reranking. As far as I know, this is the first demonstration for sentence-level attribution of an answer span (correct me if i am wrong).
- However, how much this specific strength matters depends on whether scoring with attention is what helps. If a fine-tuned scorer that does not use attention does equally well, the contribution reduces to “supervised fine-tuning helps”, and the probing-to-learning shift loses most of its force. See W1 for more details.
- S2: The paper shows that attribution training data carries a structural prior.
- LongCite’s citations are generated automatically and are local and contiguous: the standard deviation of evidence positions is 1.61 sentences, against 7.21 on HotpotQA (Table 2, B.3). LongCite-trained Attribution models get worse on HotpotQA, while joint training improves both (Table 2).
- This could potentially inform research that aim to build attribution datasets or trains on them. The spread of evidence in the training data becomes a design variable that needs reporting and control. Silver-label datasets made with LLMs, which the paper proposes for scaling (Section 1), are likely to inherit whatever biases the LLM that labels them has.
- The paper treats this finding as a limitation to work around with more data. I think it is the paper’s most transferable insight, but it might also be the simplest alternative hypothesis for the in-domain gains. See W2 for more details.
Critical weaknesses
- W1: The trained attribution model is in effect a long-context reranker, but the paper misses reranker baselines that are trained on the same data.
- TIR fine-tunes a separate model to score the units of a (document, question, answer, span) input (Section 3), which makes it a long-context “reranker” trained on LongCite labels. Table 1 compares it only with untrained attention scores (MV, Value Flow) and with AT2, a small estimator on frozen attention features that trails untrained MV (70.76 vs. 73.52).
- Table 4 suggests the choice of scoring signal does not matter much once the model has been fine-tuned. The four Llama signals span 59.40–73.75 R@K before training and 76.99–80.31 after. HSC, which is not an attention score, rises from 59.40 to 77.67. This indicates that supervised fine-tuning causes the gain, but not necessarily the attention.
- Missing baselines: the same layers, data and loss with a score that does not use attention (e.g. or a linear scoring head); a small reranker (~1B parameters); and citation models which are trained on the same data (LongCite-8B, SelfCite), all evaluated under GoldK. If these match TIR, the result reduces to “supervised fine-tuning helps in-distribution”.
- W2: The out-of-distribution results do not support the generalization claims.
- Training on one dataset hurts the other (Table 2, Section 4.3). LongCite-only training lowers HotpotQA R@K from 76.62 to 72.36, and HotpotQA-only training lowers LongCite from 73.52 to 70.85 or 67.50.
- VERI-GRAN is the only test set never trained on, and all three single dataset training lower the score in VERI-GRAN (, and of 320 units, since Table 2 values are multiples of ). The “0.63 points” out-of-distribution gain is units from one joint-training run. This does not support “a degree of cross-domain generalization” (abstract) nor contribution 2.
- Alternative hypothesis that the authors can test: TIR learns the “location” prior of LongCite’s auto-generated citations ( vs. on HotpotQA) rather than better evidence matching. That would explain the drop on dispersed HotpotQA evidence and the lack of gain on single-unit VERI-GRAN. It also predicts that a locality heuristic (untrained MV’s top unit plus its neighbours, up to ) recovers much of the LongCite gain. Unfortunately, the paper did not test this.
- W3: The novelty claim is too broad.
- The paper claims that “no existing method improves attribution performance by training the model’s internal representations themselves.” Recent work trains selected heads so that their attention ranks context: BlockRank (Gupta et al., 2025), QRRanker (Li et al., 2026) and RetMask (Ma & Okazaki, 2026), which trains through retrieval heads to improve citation. HeadRank (Wang et al., 2026) is close to concurrent and I do not weigh it. Earlier work supervises attention as well: LISA (Strubell et al., 2018) trains a designated head, and Stacey et al. (2022), Bao et al. (2018) and Mi et al. (2016) supervise attention with explanations, rationales and alignments. The novelty is the sentence-level, span-conditioned setting. Also, using the same literal claim of this paper, fine-tuned citation models such as LongCite and SelfCite also train internal representations; what distinguishes TIR is that its loss acts on an internal attention score rather than on generated output.
- The paper motivates its score with an interpretability argument (Kobayashi et al., 2020) but does not cite the debate on attention as explanation (Jain & Wallace, 2019; Wiegreffe & Pinter, 2019). Trained attention can diverge from what a model relies on (Pruthi et al., 2020), and TIR’s attention sits in a separate model. Its scores are therefore "corroborative", saying which sentences support the span, rather than "contributive", saying what the generation model used (Worledge et al., 2024). The paper should state this, because “audit” in the introduction suggests the latter.
- The value-norm obstacle is specific to MV. Raw attention trained with TIR reaches 80.09, against 80.31 for MV (Table 4), so dropping the value norm avoids the problem entirely. The mechanism also rests on a single curve of (Figure 2).
References
- Bao et al. (2018), Deriving Machine Attention from Human Rationales, EMNLP.
- Chuang et al. (2025), SelfCite: Self-Supervised Alignment for Context Attribution in Large Language Models, ICML.
- Cohen-Wang et al. (2025), Learning to Attribute with Attention, arXiv.
- Gupta et al. (2025), Scalable In-context Ranking with Generative Models, NeurIPS.
- Jain & Wallace (2019), Attention is not Explanation, NAACL.
- Li et al. (2026), Query-focused and Memory-aware Reranker for Long Context Processing, arXiv.
- Ma & Okazaki (2026), From Interpretability to Performance: Optimizing Retrieval Heads for Long-Context Language Models, Findings of ACL.
- Mi et al. (2016), Supervised Attentions for Neural Machine Translation, EMNLP.
- Pruthi et al. (2020), Learning to Deceive with Attention-Based Explanations, ACL.
- Stacey et al. (2022), Supervising Model Attention with Human Explanations for Robust Natural Language Inference, AAAI.
- Strubell et al. (2018), Linguistically-Informed Self-Attention for Semantic Role Labeling, EMNLP.
- Wang et al. (2026), HeadRank: Decoding-Free Passage Reranking via Preference-Aligned Attention Heads, arXiv.
- Wiegreffe & Pinter (2019), Attention is not not Explanation, EMNLP.
- Worledge et al. (2024), Unifying Corroborative and Contributive Attributions in Large Language Models, SaTML.
- Zhang et al. (2025), LongCite: Enabling LLMs to Generate Fine-grained Citations in Long-context QA, Findings of ACL.
Questions for authors
-
Q1: Were the LongCite-Test documents truncated to 10K tokens the same way as the training documents, by deleting random non-evidence passages? If so, how do untrained MV and TIR compare on the untruncated documents?
-
Q2: How much does TIR depend on head preselection? Results with six random heads from layer 16 (which keeps the trained stack fixed) or with the six lowest-ranked heads would show whether training creates attribution ability or sharpens existing retrieval heads. Table 9 randomizes the selected heads' weights but keeps their positions. Also, are heads chosen on a validation set (Section 3.2) or on training samples (A.3), and are they reselected for the HotpotQA and joint runs?
-
Q3 (minor clarification): Trying to run the math: With the Section 4.4 scores and
I get at and at , not the reported and . For TIR, Eq. 4 with gives , not . Is there a normalization that the text does not state?
4 comments
-
W2 — generalization. We added three human-annotated benchmarks, using the same checkpoint with no further training:
Benchmark Base TIR TIR vs Base Attributed-QA: 23 real RAG systems, 9,483 answers with human-verified evidence 77.89 79.61 1.72 QASPER: 300 questions, human evidence paragraphs 76.72 77.03 0.31 GaRAGe: 1,939 questions, human grounding annotations 86.90 87.13 0.23 Attributed-QA is the real-world RAG audit setting the review asks for. QASPER and GaRAGe show no significant change. We also evaluated TIR trained on a LongCite + HotpotQA mixture: it is significantly above the untrained baseline on both test datasets of LongCite and HotpotQA (+3.99 / +11.21pp) and shows no drop on any of the four held-out datasets (VERI-GRAN +0.63pp, Attributed-QA +0.72pp, QASPER +1.52pp, GaRAGe +0.26pp).
A locality heuristic—computing the final score as the current sentence’s score plus the average score of its neighboring candidates, with tuned on the train set—improves performance on LongCite by 1.84 pp. By contrast, training with TIR on LongCite yields an improvement of about 7 pp, indicating that the information learned during training goes far beyond such local features. Applying the locality heuristic to the TIR-trained model yields only a 0.15 pp improvement, suggesting that this heuristic may already have been captured during training.
-
W3 — novelty claim.
Earlier work supervises attention using alignments, human rationales, or syntactic structure (Mi et al., 2016; Bao et al., 2018; Strubell et al., 2018). These studies use external guidance to shape attention, whereas we optimize attention scores end to end for sentence-level evidence ranking. We will expand the related work discussion to acknowledge these contributions.
Regarding prior work that trains attention for ranking or attribution, BlockRank (Gupta et al., 2025) jointly introduces structured sparse attention for efficient in-context ranking and an auxiliary contrastive objective to optimize query-document relevance in attention. Its primary setting is passage-level document ranking. QRRanker (Li et al., 2026) uses InfoNCE to train intermediate-layer attention signals in Qwen3-4B and outperforms trained generative rerankers; it focuses on passage-level, query-conditioned ranking; RetMask (Ma & Okazaki, 2026) uses retrieval heads to construct preference data for DPO. In contrast, our work focuses on sentence-level evidence attribution conditioned on an answer span. Our contribution is distinguished by identifying the problems of InfoNCE when training attention-based fine-grained attribution models and by developing corresponding solutions.
The MV signal used in our training builds on the vector-norm analysis of Kobayashi et al. (2020), as noted by the reviewer. We will also discuss the debate on attention as explanation (Jain & Wallace, 2019; Wiegreffe & Pinter, 2019) to better situate our work within this literature. Our attribution model is separate from the answer-generation model and therefore identifies corroborative evidence rather than passages that contributed to answer generation, following the distinction in Worledge et al. (2024). We will clarify this distinction in the motivation.
Regarding the role of value-norm stop-gradient, we do not claim that MV provides a substantial performance advantage over raw attention: both achieve comparable results after TIR training (80.31 vs. 80.09). Instead, the stop-gradient addresses a training-specific failure mode of MV, where the value-norm factor can become an optimization shortcut and undermine attention-based evidence ranking. By preventing gradients from exploiting this factor, the mechanism allows the more interpretable MV signal to remain effective during training. Thus, its contribution lies in making MV trainable for attribution, rather than in improving its final performance over raw attention.
We will narrow our novelty claim to the design of a dedicated mechanism for training internal representations for fine-grained evidence attribution conditioned on an answer span.
- Q1 — Truncation. Yes. We applied the same 10K truncation procedure independently to examples in both the training and test splits, removing randomly sampled non-evidence passages. We will explain the specific preprocessing strategy in the paper. Results on the untruncated data are shown below:
| Untruncated | Untrained MV | TIR | Δ |
|---|---|---|---|
| BLEU@GoldK | 61.47 | 66.81 | +5.34 |
| Recall@GoldK | 60.87 | 65.45 | +4.58 |
-
Q2 — Head selection. The attention heads were selected using 1,000 examples randomly sampled from the LongCite training set, and they were not re-selected on other datasets. We will revise and expand the description in Section 3.2 accordingly.
The experimental results show that head preselection has some effect on training outcomes, primarily depending on whether the best attention heads are selected. Taking Recall@GoldK on the LongCite dataset as an example, training with six attention heads randomly selected from layer 16 (0, 6, 25, 27, 28, 29) or with the six worst attention heads (12, 14, 15, 28, 30, 31) yields results 1.25 pp and 10.09 pp lower than the best TIR training result, respectively. This indicates that attention heads with poorer retrieval scores have limited improvement during training. In addition, we find that training with the best 1/16/32 heads is very close to the best result (−0.22 / −0.07 / −0.27 pp), suggesting that the layer-16 hidden_state contains relatively rich features that can be used for fine-grained attribution. These results show that head preselection affects training outcomes, particularly when using heads with poor initial retrieval scores. Meanwhile, selecting only a small number of high-ranking heads achieves performance close to the best result, suggesting that a compact set of attention heads can expose sufficient information for fine-grained attribution through internal representation training.
-
Q3 — Mathematical corrections. Thank you for identifying these errors. We rechecked the gradient calculations and confirmed that the reported InfoNCE gradients contained numerical errors. We will replace them with the correctly computed values provided in the review. We also omitted the factor in Eq. (4), which accounts for the discrepancy in the TIR gradients. After including this factor, our TIR gradients match the calculations in the review. These corrections do not affect our qualitative conclusion regarding competition among positive units, and we will correct the corresponding equations and numerical examples in the paper.
Thank you for the new analyses. I think they have mostly answered my concerns. From my understanding, it seems that the results confirm my thinking: that the gain is mainly from the fine-tuning, not from the attention-based design. Also, the out-of-distribution gains are small.
Given these, I decide to raise my soundness score but keep my overall rating. I would still encourage the authors to revise the paper to state that the gain comes from the training process rather than the attention readout itself. The authors should also narrow down the generalization claim.
We thank the reviewer for the constructive assessment.
W1: Matched non-attention scorer. We trained a linear interaction scorer with the same layer-16 prefix, inputs, labels, loss, data order, optimizer and update budget as TIR. The token feature h is the hidden state at the output of the layer-16 input_layernorm; u is the mean of h over all tokens of the candidate sentence, and t is the mean of h over all tokens of the target answer span. Linear(16384→1) maps [u,t,u⊙t,∣u−t∣] to the score of that sentence (16,385 readout parameters). LongCite R@GoldK:
The results show that this method (Linear readout + trained backbone) performs almost identically to TIR, indicating that the choice between these two readout designs is not critical. When we freeze the internal representations h and train only the linear readout, Recall@GoldK is only 70.49, with even larger gaps relative to TIR on the three additional test datasets (QASPER −21.0, Attributed-QA −13.2, GaRAGe −11.2). These results demonstrate clear gains from training internal representations. We will supplement this experiment and conclusion.
Both SelfCite and LongCite-8B are generative fine-grained attribution models, making it difficult to control their citation placement at predefined positions in an answer. This makes it difficult to evaluate Recall@GoldK using the existing human annotations. Generative fine-grained attribution models typically use LLM judges to evaluate their citations. For comparison with our non-generative method, we evaluated LongCite-8B and SelfCite-8B on Attributed-QA and HotpotQA, where the attribution target is the entire answer. We extracted the cited evidence units from each generated answer in their order of appearance and retained the first GoldK unique units as the predicted evidence set. It trails our method by approximately 11 pp and 17 pp, respectively. These gaps persist even when the generated answer contains the gold answer. On Attributed-QA, 87.1% of generated answers contain the gold short answer; on this subset, LongCite-8B achieves 71.14% Recall@GoldK, 8.2 pp below TIR. On HotpotQA, 77.7% of generated answers contain the gold answer; on this subset, LongCite-8B achieves 55.63% Recall@GoldK, 17.19 pp below TIR. On the subset where SelfCite-8B generates correct answers, the gap remains 4.2 pp (Attributed-QA) and 17 pp (HotpotQA).
Results from training reranker-based models: