acceptodds
Withdrawn from ICLR 2027

CAGE: Cognitive Attribution Graphs for Faithful Inline Citation Generation in Long-Form Question Answering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-form question answering increasingly relies on retrieved evidence to make LLM outputs verifiable, with inline citations tracing claims to source documents. However, existing systems often attach citations that are topically related but insufficient to support their claims. We identify attribution ambiguity as a structural challenge: end-to-end generation must implicitly resolve combinatorial claim–document assignments, obscuring evidential boundaries and increasing the risk of evidence-boundary overrun, where claims exceed cited support. To address this challenge, we propose CAGE (Cognitive Attribution Graphs for Citation Generation), a two-stage framework that introduces an explicit cognitive attribution map before answer generation. CAGE first trains a plug-and-play Cognitive Map Induction Model to construct answer-centered support subgraphs, aligning each semantic answer unit with supporting documents through explicit relations. A Structured Citation Reasoning Model then realizes these units as sentence-level claims with map-aligned citations. Experiments on ASQA, ELI5, and ExpertQA show that CAGE achieves state-of-the-art performance, demonstrating the effectiveness of attribution-space contraction and map-guided citation generation.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Citation Generation, Question Answering, Large Language Models

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_Dc1MOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CAGE, a two-stage framework for generating inline citations in long-form question answering. A Cognitive Attribution Map Induction Model (CMI) first constructs support subgraphs linking answer units to retrieved documents. A Structured Citation Reasoning Model (SCR) then generates an answer with citations guided by these assignments. The authors use LLM-generated and validated supervision to train the two components. Experiments on ASQA, ELI5, and ExpertQA report improvements in aggregate TRUST scores, particularly citation groundedness. Model-scale analyses, structural ablations, and experiments with proprietary generators provide further evidence of the framework's effectiveness and applicability across model backbones.

Critical strengths

S1. Strong empirical performance. The framework achieves competitive aggregate results across several datasets and model scales, with substantial improvements in citation groundedness. The structural ablations also provide evidence that explicit support assignments contribute beyond document selection and textual descriptions. These findings give the proposed implementation practical value.

S2. Useful modular design. The attribution module can guide different downstream generators, including proprietary models, without fine-tuning those generators. This separation makes the method applicable when the downstream model cannot be retrained and shows that the benefit is not restricted to one model family..

Critical weaknesses

W1. Limited substantive novelty. Using a graph to organize answer units and their supporting evidence is a fairly common modeling idea. Here, the graph primarily makes answer–document assignments explicit within a two-stage generation pipeline. I acknowledge that the baseline comparisons and structural ablations demonstrate the effectiveness of this design. However, they do not clearly establish what substantive modeling advance the graph provides beyond making these relationships explicit. In particular, removing support edges changes the information available to the generator, so the ablation does not isolate the graph representation from the support assignments it encodes. The same-backbone experiments are useful, but the main results also involve a 9B attribution model and supervision synthesized by GPT-5.5. These factors may contribute to the performance gains, particularly for smaller downstream generators, and the current analysis does not sufficiently separate their effects from the contribution of the graph representation itself.

W2. Insufficient analysis of answer quality and the scope of EBO. Citation groundedness improves, but Table 2 also shows lower answer-calibration scores (EMACF1) for some generators. TRUST includes correctness and refusal behavior, yet the paper does not sufficiently explain whether these declines reflect changes in answering decisions or reduced coverage of relevant information. This leaves the claim of improved grounding without compromising response quality insufficiently supported. Figure 5 also illustrates a possible coverage limitation: D5 explicitly mentions both Rob Hall and Scott Fischer, but the induced answer units and final response omit Rob Hall.In addition, Appendix A.3 checks atomic facts against all top-5 retrieved documents jointly, whereas EBO is introduced as a claim exceeding what its cited documents justify. This evaluation measures support within the retrieved evidence pool and does not directly test support within each claim's actual citations. Although the main citation-groundedness metric provides evidence on citation quality, the additional EBO analysis does not fully substantiate the more specific interpretation. Excluding refusals and empty responses also leaves possible effects of response selection and factual coverage unresolved.

Questions for authors

Q1. Clarification of the technical contribution. Beyond making answer–evidence relationships explicit, what substantive modeling capability does the graph introduce? Which existing comparison most directly distinguishes the value of the graph representation from the support assignments it encodes? How can its contribution to the observed gains be distinguished from the capacity of the CMI model and the quality of the GPT-5.5-generated supervision? Q2. Clarification of the answer-quality trade-off. How do the authors explain the lower EMACF1 scores in Table 2 despite improved citation groundedness? Could the existing evaluation be broken down by answering behavior and coverage of relevant information to clarify whether either contributes to the decline?

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

This paper proposes CAGE, which organises the answer points and their supporting documents into an attribution graph, and then produces an answer with citations on the basis of that graph, for long-form question answering. Supervision is built by a generate-and-verify pipeline, and two models are trained separately: an attribution model and an answer generator. The attribution model is also plugged into closed-source generators that receive no extra training. Experiments cover ASQA, ExpertQA and ELI5, and report answer correctness together with answering behaviour, abstention and citation quality, along with comparisons across model configurations and ablations.

Critical strengths

The same induction model plugs into Qwen 0.5B–7B, the LLaMA family, and GPT-5.5 / Claude-4.6 without retraining the generator, and it improves over the strongest baseline at every scale.

Critical weaknesses
  1. The paper says the map fixes the answer units, so the model knows what to say before writing. But the prompt in the appendix only gives the generator the citation numbers for each sentence, and it says the answer content must come from the documents, not from the map. So the map decides which documents to cite, not what to say. This is much weaker than the claim in Section 3.
  2. The prompt forces one short single-claim sentence per subgraph, forbids merging, and requires a citation on every sentence. F1_CG and EBO are computed per sentence and per atomic fact, so this format alone inflates them while baselines write natural long-form answers. No ablation isolates format from graph.
  3. Table 4 excludes refusals from EBO. CAGE abstains on the hard cases, so they leave its denominator, while GPT-5.5 keeps them. The 38.00 vs 5.77 comparison is not on the same questions, and refusal rates are not reported.
Questions for authors

No efficiency numbers are reported. CAGE runs two models and generates the map as text before answering, so can you report latency and token usage per question against the single-pass baselines?

Reviewer_yreYOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

CAGE separates citation generation into two stages: constructing a graph of answer units and supporting documents, then generating an answer with graph-aligned citations. Both stages are trained using GPT-5.5-generated and automatically validated supervision. Evaluation on ASQA, ELI5, and ExpertQA reports improvements in TRUST and citation groundedness across several model families, alongside an analysis of unsupported atomic facts.

Critical strengths

The empirical improvements in citation groundedness are substantial, particularly on ExpertQA. The ability to use the same attribution model with different downstream generators, including frozen proprietary models, is also practically useful. These results provide reasonable evidence that explicitly planning evidence attribution can improve citation generation.

Critical weaknesses
  1. The contribution beyond existing attribution planning is not sufficiently established. The graph has a fixed query–document–answer structure, and its essential information could also be represented as answer units paired with supporting document IDs. The “w/o graph” ablation removes the explicit attribution links; it does not compare against a simpler representation preserving the same information. Attribute First, then Generate is discussed but not evaluated. Likewise, reducing the assignment space to one target follows from fixing an attribution plan, while the difficult attribution decision remains in the first stage. I am not convinced that the experiments isolate the benefit of the proposed graph formulation.
  2. The faithfulness gains come with an unresolved loss in answer quality and coverage. In Table 2, adding CAGE reduces ExpertQA answer calibration from 46.74 to 32.88 for GPT and from 47.66 to 36.70 for Claude. The EBO analysis does not resolve this concern: it excludes refusals and checks support against all five retrieved documents rather than the documents actually cited. Figure 5 is also troubling as a positive example. For a question about deaths in Into Thin Air, the answer includes Boukreev’s later death on Annapurna, while omitting Rob Hall despite explicit evidence in D5. This illustrates why supported statements alone do not establish a correct, sufficiently complete answer.
Questions for authors
  1. Controlled comparison: How does CAGE compare with a plain answer-unit/citation plan that preserves the same attribution links, using the same planner, supervision, and inference budget? This would directly clarify the contribution of the graph formulation.
  2. Evaluation clarification: How much of the answer-calibration decline comes from false refusals versus missing correct answer content? Do the EBO improvements persist on questions answered by both systems when support is checked against the actual cited documents?
  3. Implementation clarification: Equation 24 requires fixed answer semantics and exact citation sets, but the appendix prompt lets the generator determine factual content from documents and omit unsupported listed citations. Which constraints govern the reported experiments, and how often do generated answers depart from the induced map?

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

Related papers

Discussion (0)

Sign in to comment.