acceptodds
Under review as a conference paper at ICLR 2027

Learning to Attribute by Training Internal Representations

Abstract

For any specified factual span in an answer generated by a large language model (LLM) with retrieval-augmented generation, fine-grained attribution aims to locate the fine-grained evidence in the context that best supports that span. This enables humans to efficiently audit LLM outputs and ensures the reliability of the results. Among approaches to this task, methods based on the model's internal representations offer fast inference, high accuracy, and pluggability. However, existing methods either apply a pretrained model directly for attribution or train only a small number of auxiliary parameters; no prior work has attempted to improve attribution by training the model's internal representations themselves. In this paper, we identify two obstacles during training for attribution: (1) the attribution-relevant training signal is sparse and correlated, while standard contrastive objectives can not focus on this sparse ranking signal, and (2) the value-vector cheating problem. To address these obstacles, we propose TIR (Training Internal Representations), which uses a margin ranking loss weighted by hard negatives and applies a stop-gradient to the value-norm component. Experiments show that TIR consistently improves attribution across multiple backbone models, multiple datasets, and multiple internal signals, and even exhibits a degree of cross-domain generalization, demonstrating the feasibility of obtaining more accurate attribution models through training.

Primary area:
interpretability and explainable AI
Keywords:
Fine-grained Attribution, Large Language Models, Attention Heads, Long-context Question Answering

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_KLNROpenReview ↗
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. hT⊤hdi\mathbf{h}T^\top \mathbf{h}{d_i} 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 (−11-11, −10-10 and −39-39 of 320 units, since Table 2 values are multiples of 1/3201/320). The “0.63 points” out-of-distribution gain is +2+2 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 (σ=1.61\sigma = 1.61 vs. 7.217.21 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 KK) 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 max⁡u∣vu∣2\max_u |\mathbf{v}_u|_2 (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 (2.0,1.4,1.9,0.1)(2.0, 1.4, 1.9, 0.1) and

    ∂LNCE∂sk=pk−1/Mτ,\frac{\partial \mathcal{L}_{\text{NCE}}}{\partial s_k} = \frac{p_k - 1/M}{\tau},

    I get (−0.12,−0.29,+0.35,+0.06)(-0.12, -0.29, +0.35, +0.06) at τ=1\tau = 1 and (+7.62,−10.0,+2.38,0)(+7.62, -10.0, +2.38, 0) at τ=0.05\tau = 0.05, not the reported (−0.18,−0.26,+0.31,+0.13)(-0.18, -0.26, +0.31, +0.13) and (+4.38,−9.95,+5.58,≈0)(+4.38, -9.95, +5.58, \approx 0). For TIR, Eq. 4 with m=1m = 1 gives (−0.25,−0.25,+0.5,0)(-0.25, -0.25, +0.5, 0), not (−0.50,−0.50,+1.00,0)(-0.50, -0.50, +1.00, 0). Is there a normalization that the text does not state?

4 comments

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\mathbf h is the hidden state at the output of the layer-16 input_layernorm; uu is the mean of h\mathbf h over all tokens of the candidate sentence, and tt is the mean of h\mathbf h over all tokens of the target answer span. Linear(16384→1) maps [u,t,u⊙t,∣u−t∣][u,t,u\odot t,|u-t|] to the score of that sentence (16,385 readout parameters). LongCite R@GoldK:

    ReadoutBackbone frozenBackbone trained
    MV73.5180.31
    Linear interaction70.4980.39

    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\mathbf 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).

    DatasetLongCite-8B Recall@GoldKSelfCite-8B Recall@GoldKUntrained MVTIRΔ TIR−LongCite8BΔ TIR−SelfCite-8B
    Attributed-QA68.37%72.50%77.89%79.61%+11.24pp+7.11pp
    HotpotQA54.55%54.99%76.62%72.36%+17.81pp+17.37pp

    Results from training reranker-based models:

    ModelSizeRecall@GoldKBLEU@GoldK
    BGE-v2-m3 fine-tuned560M54.1054.70
    Qwen3-Reranker-0.6B w/o fine-tuned0.6B53.1953.53
    Qwen3-Reranker-0.6B fine-tuned0.6B65.6665.88
    TIR4B80.3180.26
  • W2 — generalization. We added three human-annotated benchmarks, using the same checkpoint with no further training:

    BenchmarkBaseTIRTIR vs Base
    Attributed-QA: 23 real RAG systems, 9,483 answers with human-verified evidence77.8979.611.72
    QASPER: 300 questions, human evidence paragraphs76.7277.030.31
    GaRAGe: 1,939 questions, human grounding annotations86.9087.130.23

    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 nn neighboring candidates, with nn 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:
UntruncatedUntrained MVTIRΔ
BLEU@GoldK61.4766.81+5.34
Recall@GoldK60.8765.45+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 1∣N∣\frac{1}{|N|} 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.

Reviewer_zrfpOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the task of fine-grained evidence attribution in Retrieval-Augmented Generation (RAG) systems. While prior internal-representation-based attribution methods rely on untrained models or auxiliary parameters, this work explores directly training the language model's internal representations into dedicated attributors. And, this paper proposes TIR (Training Internal Representations), which combines a hard-negative-weighted margin ranking loss with a stop-gradient operation applied to the value norm. Experiments on LongCite, HotpotQA, and VERI-GRAN datasets using Llama-3.1-8B and Qwen3-4B backbones show that TIR substantially improves attribution metrics (e.g., Recall@GoldK) across multiple internal signals.

Critical strengths

Experiments on LongCite, HotpotQA, and VERI-GRAN datasets using Llama-3.1-8B and Qwen3-4B backbones show that TIR substantially improves attribution metrics (e.g., Recall@GoldK) across multiple internal signals.

Critical weaknesses
  1. Methodological incrementalism and heuristic fixes. The core methodological contribution relies on combining two well-established techniques: a pairwise margin ranking loss weighted by softmax-based negative difficulty (reminiscent of standard learning-to-rank and hard-negative mining) and a stop-gradient patch. While stopping gradients through the value norm effectively prevents norm explosion, it acts as an ad-hoc empirical fix rather than a principled solution to representation collapse or scale instability.
  2. Over-reliance on static pre-selected attention heads. The method restricts internal signal extraction and fine-tuning to a fixed, pre-selected set of top-6 attention heads chosen offline on validation data. Freezing channel selection beforehand limits the model's capacity to dynamically adapt or discover complementary attribution heads across different layers during joint optimization.
  3. Limited scope of evaluation benchmarks. The experiments are largely evaluated on two synthetic/automatically constructed long-context QA datasets (LongCite and HotpotQA). The out-of-distribution evaluation is confined to VERI-GRAN, which contains only 197 test instances with a single annotated evidence unit per instance (K=1K=1). Testing on more challenging, human-annotated real-world benchmark datasets (such as AttnQA or real-world RAG audit datasets) would be necessary to thoroughly establish practical efficacy.
Questions for authors

Please refer to the weaknesses.

1 comment

We thank the reviewer. We ran the test each weakness calls for; short answers below.

  • W1: Informativeness. We agree neither component is new on its own. Our contribution is the diagnosis of why the standard choices fail when internal representations are trained for attribution, and each diagnosis is backed by a measurement.

    • Loss. With multiple positives, InfoNCE introduces competition among positive units: any positive whose probability exceeds the uniform target of 1/1/ receives a gradient that decreases its score ($\partial \mathcal{L}/\partial s_k=(p_k-1/M)/\tau $). It also continues to allocate gradient to negatives according to their softmax probabilities, whereas our margin objective eliminates the contribution of negatives that already satisfy the margin. Our objective therefore avoids both probability competition among positives and unnecessary optimization of sufficiently separated negatives.
    • Stop-gradient. Q/K determine the attention allocation, while the value norm multiplicatively scales each token's contribution to the evidence score. Of these two learnable components, the latter provides a shortcut that should not be optimized in this task. For positive tokens whose margins are violated, the loss has a negative gradient with respect to their value norms so the model inflates it instead of learning the match. Without the stop-gradient, training collapses to 50.21, below the untrained 73.52, while the max value norm keeps growing (Fig. 2). The fix follows from this gradient analysis, not from trial and error.
  • W2: fixed top-6 heads. Preselecting a fixed, sparse head set has precedents beyond attribution, including document re-ranking (QRHead; Zhang et al., 2025), hallucination detection (TOHA; Bazarova et al., 2026), and truthful generation (ITI; Li et al., 2023). Selected head sets are also used for targeted fine-tuning to mitigate visual hallucinations (TF-HH; Yang et al., 2025). To test whether this caps TIR, we trained nested top-K sets under one protocol. Relative to K=6, K=1 / 16 / 32 change Recall@GoldK by −0.22 / −0.07 / −0.27pp. These results show that TIR does not rely on a large collection of heads: even a single selected head captures nearly all the performance gain, and expanding the fixed head set provides negligible benefit. We acknowledge that the current method uses a globally selected head set rather than per-instance dynamic routing; investigating query-dependent head selection is an orthogonal direction beyond the scope of this work.

  • W3: benchmark scope. We added three additional benchmarks with human-provided answers, evidence annotations, or attribution judgments, all of which can be mapped to TIR's evaluation format. We evaluated the same LongCite-trained checkpoint without any additional training:

    BenchmarkBaseTIRTIR vs BaseMixed vs Base
    Attributed-QA: 23 real RAG systems, 9,483 answers with human-verified evidence77.8979.61+1.72+0.73
    QASPER: 300 questions, human evidence paragraphs76.7277.03+0.31+1.52
    GaRAGe: 1,939 questions, human grounding annotations86.9087.13+0.22+0.26

    We further evaluated a separately trained TIR model using a LongCite + HotpotQA training mixture. Compared with the untrained baseline (The Mixed vs Base in table): 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 +0.73pp, QASPER +1.52pp, GaRAGe +0.26pp).

We will add the matched loss comparison, the head sweep with retrieval-head background, and the three benchmarks, and fix the transposed InfoNCE columns in Table 3.

We would be grateful if the reviewer would consider raising the score.

Reviewer_3cTYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TIR is a fine-grained attribution method built on the principle that an LLM's attention from an answer span back to the source document can be trained into an evidence ranker. It argues that naive training fails for two reasons: the useful signal sits in a few hard negatives that InfoNCE handles poorly, and a score that multiplies attention by the value-vector norm lets the model lower its loss by inflating value norms. TIR scores each document unit by summing attention times value norm over preselected heads. It fine-tunes every layer up to the deepest selected head on a 10K-token version of LongCite-45K, using a hinge ranking loss weighted toward hard negatives and a stop-gradient on the value norm. At inference, it runs one forward pass over the document, question, and answer, scores each unit, and returns the top-ranked units as evidence. On LongCite, HotpotQA, and VERI-GRAN with Llama-3.1-8B and Qwen3-4B, it reports LongCite Recall@GoldK gains of about 6 to 7 points over untrained internal signals such as AT2.

Critical strengths
  1. Training the attention readout itself, instead of only selecting heads from a frozen model, is a well-motivated idea. It's interesting to see that the underlying attention can be trained to infer attribution with competitive performance against the baselines.
  2. The training recipe is well matched to the task. Removing the stop-gradient and applying the hard-negative weighting are reasonable techniques suited for this task.
Critical weaknesses
  1. The evaluation does not show that training the internal representations is what produces the gain, and the closest prior work is not discussed. Specifically, every baseline in Table 1 is either untrained or trained for a different target (AT2 predicting ablation effects). No baseline uses the same LongCite labels without the attention readout. In addition, BlockRank (https://arxiv.org/abs/2510.05396) already fine-tunes an LLM with a contrastive loss on mid-layer query-to-document attention so that attention ranks documents. This contradicts the claim in the abstract and Section 1 that no prior work trains internal representations for this purpose.
  2. The main paper's own results do not support the claim of generalization for TIR. It looks like TIR mostly work in-domain from Table 2, because training on HotpotQA degrades the performance on LongCite. In fact, every single-dataset trained model ends up below the untrained model on the other datasets. As a result, this strategy looks more like domain adaptation than generalization.
Questions for authors
  1. The "w/o hard-negative weighting" (75.34) scores below the best InfoNCE run (77.35, Table 3). What if we use InfoNCE with hard-negative weighting?
  2. Can we report confidence interval for table 2?
  3. Which parameters does TIR train, and how many?
3 comments
AuthorsOpenReview ↗

About Weekness

We thank the reviewer for the constructive feedback. We provide additional results and clarify our work below:

  • W1 —Whether training the internal representations produces the gain. We trained a linear interaction scorer with the same labels, loss, update budget as TIR: The token feature h\mathbf h is the hidden state at the output of the layer-16 input_layernorm; uu is the mean of h\mathbf h over all tokens of the candidate sentence, and tt is the mean of h\mathbf h over all tokens of the target answer span. Linear(16384→1) maps [u,t,u⊙t,∣u−t∣[u,t,u\odot t,|u-t| to the score of that sentence. LongCite Recall@GoldK:

    ReadoutBackbone frozenBackbone trained
    Attention73.5180.31(TIR)
    Linear readout70.4980.39

    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\mathbf 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.

  • W1 — Regarding BlockRank. We acknowledge that BlockRank is a relevant antecedent that uses contrastive supervision to train query-to-document attention for document-level in-context ranking. It combines this objective with structured sparse attention to improve both ranking efficiency and attention-based relevance estimation. TIR addresses a different problem: answer-span-conditioned fine-grained evidence attribution, where individual document units must be ranked according to their support for a specific answer span. Beyond this difference in task formulation and supervision granularity, TIR introduces a value-norm stop-gradient and hard-negative-weighted margin objective tailored to this setting. We will revise the related work and narrow our novelty claim to clarify this distinction.

  • W2 — generalization. In response to multiple reviewers, we added evaluations on Attributed-QA, QASPER and GaRAGe (Attributed-QA is human-rated attribution of real RAG systems (23 systems, 9,483 samples)). Training jointly on LongCite and HotpotQA (the Mixed model) is at or above the untrained baseline on every one of them:

    DatasetBaseTIR (LongCite only)Mixed
    LongCite73.5280.3177.51
    HotpotQA76.6272.3687.86
    VERI-GRAN (320)78.7575.3179.38
    Attributed-QA (9,483)77.8979.6178.61
    QASPER (300)76.7277.0378.23
    GaRAGe (1,939)86.9087.1387.16

    Table 2 shows that LongCite and HotpotQA exhibit distinct evidence-position distributions, which may explain the cross-domain degradation observed with single-dataset training. Joint training on both datasets substantially mitigates this degradation, improving performance on HotpotQA while retaining gains over the untrained baseline on LongCite. Moreover, the Mixed model maintains or slightly improves performance across all three additional benchmarks (Attributed-QA, QASPER, and GaRAGe), as well as VERI-GRAN. These results demonstrate that combining complementary training distributions using TIR is an effective way to improve cross-dataset robustness.

AuthorsOpenReview ↗

About Questions.

  • Q1 — InfoNCE with hard-negative weighting. The weighting uses the same detached softmax over negatives as the margin loss, with τ=0.0\tau=0.0:

    ObjectiveR@GoldK
    Margin, uniform negatives75.34
    InfoNCE(best config)77.35
    InfoNCE + hard-negative weighting79.46
    Margin + hard-negative weighting (TIR)80.31

    Hard-negative weighting yields clear improvements for both losses, with the margin loss showing a small advantage of 0.85 percentage points over InfoNCE. The ordering also explains the Table 3 result the reviewer noticed: plain InfoNCE already beats the unweighted margin because the softmax in its denominator is an implicit hard-negative weighting.

  • Q2. Yes. We report the requested confidence intervals in the table below. All intervals are 95% paired cluster bootstrap over 10,000 draws, and they quantify test-sampling uncertainty for these fixed checkpoints. Clusters are the source document for LongCite (8,374 clusters), the question for Attributed-QA (884), the paper for QASPER (189), the source context for VERI-GRAN (132), and the sample itself for HotpotQA and GaRAGe. Recall@GoldK:

    DatasetBaseTIR (LongCite only)Mixed (LongCite + HotpotQA)
    LongCite (51,077)73.5280.31, +6.80 [+6.53, +7.07]77.51, +4.00 [+3.76, +4.22]
    HotpotQA (1,000)76.6272.36, −4.26 [−5.73, −3.06]87.86, +11.23 [+9.69, +12.76]
    VERI-GRAN (320)78.7575.31, −3.44 [−7.47, 0.00]79.38, +0.63 [−3.63, +4.93]
    Attributed-QA (9,483)77.8979.61, +1.72 [+0.35, +3.09]78.61, +0.73 [−0.76, +2.19]
    QASPER (300)76.7277.03, +0.31 [−3.06, +3.81]78.23, +1.52 [−2.17, +5.24]
    GaRAGe (1,939)86.9087.13, +0.22 [−0.29, +0.71]87.16, +0.26 [−0.20, +0.71]
  • Q3. Full fine-tuning updates the token embedding and every transformer layer up to and including the deepest readout layer. For Llama-3.1-8B, 4.036B parameters receive gradient — the layer-16 value projection is excluded because the value norm is detached, which is the point of the stop-gradient. The interaction control trains 4.015B plus its 16,385 new parameters.

  • Clarification on Strength 2. We would like to clarify that TIR applies a stop-gradient to the value-norm term, blocking gradients through the value norm during optimization.

We hope this addresses the review and would be grateful if the reviewer would consider raising the score.

Thank you very much for your valuable feedback, which has been highly useful in refining our manuscript. It would be great if you could review our replies and give us additional comments on whether further refinement could be made in our next revision. We would be very grateful for your time and effort.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.