Pith. sign in

REVIEW 3 major objections 6 minor 57 references

Building an explicit answer–document map before writing cuts citation ambiguity and keeps long-form answers inside their evidence.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Explicit cognitive attribution graphs before generation contract claim–document assignment space and yield SOTA faithful inline citations on long-form QA benchmarks.

T0 review reviewed 2026-07-31 challenge →

load-bearing objection Solid systems paper: explicit answer–document support graphs before generation help citation faithfulness; TRUST gains look real, EBO vs GPT/Claude is partly confounded by abstention and style. the 3 major comments →

arxiv 2607.24236 v2 pith:4V5FRW6V submitted 2026-07-27 cs.CL

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

Zhichao Yan , Shizhao Li , Jiapu Wang , Haoran Luo , Qingang Zhang , Jiaoyan Chen , Ru Li , Jeff Z. Pan This is my paper
classification cs.CL
keywords inline citation generationlong-form question answeringattribution ambiguitycognitive attribution mapevidence-boundary overrunretrieval-augmented generationclaim-document alignmentTRUST evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Long-form answers with inline citations often look grounded while still overrunning what the cited documents can strictly support. The paper argues this is not only a capacity problem: when a model must invent claims and pick supporting documents in one pass, it faces a huge combinatorial assignment space and can treat topical relevance as if it were sufficient evidence. CAGE inserts a middle step that first builds a cognitive attribution map—answer-centered support subgraphs that say which documents back which semantic answer units, or that the evidence is empty and the system should refuse. A second model then turns those units into sentences whose citations match the map. On ASQA, ELI5, and ExpertQA the approach leads reported TRUST scores, improves both open and proprietary generators without retraining them, and sharply lowers evidence-boundary overrun relative to strong end-to-end systems.

Core claim

The authors claim that attribution ambiguity is structural in end-to-end citation generation, and that contracting the citation-assignment space with an explicit cognitive attribution map before decoding produces more faithful inline citations. Conditioned on the induced map, each answer unit has one graph-specified citation target instead of up to (2^n−1)^m unconstrained assignments, which they show yields state-of-the-art TRUST and large drops in evidence-boundary overrun across ASQA, ELI5, and ExpertQA.

What carries the argument

Cognitive attribution map: a collection of answer-centered support subgraphs linking the query, supporting documents, and a semantic answer unit via explicit support edges (plus empty maps for abstention). A plug-and-play Map Induction Model builds the map; a Structured Citation Reasoning Model realizes units as cited sentences aligned to the map.

Load-bearing premise

The synthetic maps and map-aligned answers used as training targets are faithful enough proxies for real evidential support that teaching models to follow them truly contracts attribution rather than copying a teacher’s style—even though map induction still drops supporting evidence on a non-trivial share of cases.

What would settle it

Hold the retrieved top-k fixed and compare map-guided generation to strong end-to-end baselines on atomic-fact support against those documents: if Atomic-EBO, EBO-Severity, and citation TRUST stop improving—or if ablating the graph structure while keeping the same documents erases the gains—the attribution-space contraction claim fails.

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

If this is right

  • Citation-grounded long-form QA can separate “what can be supported” from “how to word the answer,” with the first stage reusable across generators.
  • Proprietary or frozen answer models can gain citation faithfulness by consuming an external attribution map without generator-specific retraining.
  • Empty maps give a structured refusal path when retrieved evidence cannot support any answer unit.
  • Gains should track map quality more than answer-model scale once a high-quality map is available.
  • Evidence-boundary overrun becomes measurable and reducible by constraining claims to graph-specified support sets.

Where Pith is reading between the lines

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

  • Any pipeline that still jointly decodes claims and citations without an intermediate support structure is likely to keep paying a combinatorial ambiguity tax even as base models get stronger.
  • Improving map induction (connectivity, sufficiency checks, multi-hop unit split) may matter more for deployment than further fine-tuning of the final writer.
  • The same answer-centered support graph idea could transfer to other grounded generation settings where topical retrieval is routinely mistaken for warrant.
  • Error propagation from bad maps suggests future systems may need joint repair or verification loops between map and answer stages.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes CAGE, a two-stage framework for inline citation generation in long-form QA. A Cognitive Map Induction Model (CMI) first constructs answer-centered support subgraphs linking semantic answer units to supporting documents (including an empty-map abstention channel); a Structured Citation Reasoning Model (SCR) then realizes each unit as a sentence-level claim with map-fixed citations. Supervision for both stages is synthesized by a GPT-5.5 generator/validator pipeline (VCAS) that checks subgraph connectivity and evidential sufficiency. Experiments on ASQA, ELI5, and ExpertQA report state-of-the-art TRUST scores across Qwen and LLaMA SCR backbones at multiple scales, plug-and-play gains when CMI is prepended to GPT-5.5 and Claude-4.6, ablations isolating graph structure from document filtering, and large reductions in a new evidence-boundary overrun (EBO) metric versus end-to-end proprietary baselines.

Significance. If the results hold, the work offers a practically useful decomposition: a trained, model-agnostic attribution module that improves heterogeneous generators, including proprietary APIs, without generator-specific retraining. Specific strengths worth crediting: the w/o-graph ablation (Fig. 4) retains the same documents while removing structure, cleanly separating attribution structure from mere document filtering; the scale analysis (Fig. 3) identifies map induction, not generator capacity, as the bottleneck; results are consistent across three datasets, two model families, and five scales; and the authors disclose a 15.42% map-induction failure rate rather than hiding error propagation. The EBO metrics are a reasonable attempt to operationalize "claims exceeding cited support." Main limitations to significance: all supervision and part of the evaluation depend on a single proprietary teacher (GPT-5.5), no code/data release is indicated, and the headline EBO comparison is subject to a selection confound detailed below.

major comments (3)
  1. [Table 3; Appendix Eqs. (28)-(30)] The EBO protocol excludes refusal and empty responses from N, but CAGE has a trained abstention channel (empty-map supervision, Eq. 11) and its F1_RG gains (Tables 1-2) show it refuses more than end-to-end GPT-5.5/Claude-4.6, which answer nearly everything. Refusals concentrate on weak-evidence questions — exactly where unsupported atomic facts would cluster. The 38.00%/40.85% → 5.77-7.21% Atomic-EBO gap therefore mixes a treatment effect with a selection effect over an easier answered subpopulation. Required: per-method refusal rates on the 948 questions, and EBO recomputed on the intersection of questions all methods answer (and/or answerable-only EBO with unjustified refusals penalized).
  2. [Table 3; SCR training prompt (Appendix)] A second, independent confound in the EBO comparison: the SCR answer prompt forces 'exactly one independent answer sentence' per subgraph, 'short and focused on one factual claim,' with citations fixed by the map. GPT-5.5-based FActScore decomposition of such templated sentences yields fewer, cleaner atoms that TRUE can verify, while fluent end-to-end paragraphs decompose into more discourse-laden atoms with higher false-unsupported rates. Both confounds inflate the ~30-point EBO gap in the same direction. A style-matched control (e.g., prompting GPT-5.5/Claude to emit one claim per sentence) or human verification of a sample of 'unsupported' atoms per method would bound this effect.
  3. [VCAS, Steps 1-3 (Eqs. 10-18); Limitations] All CMI/SCR supervision derives from GPT-5.5 as both generator and validator, with predicates (Connected, D_r |=_Q a_r, permutation alignment) also LLM-judged; GPT-5.5 additionally serves as the EBO decomposer, so teacher and evaluator errors are correlated. The acknowledged 15.42% map-induction failure rate is not analyzed by direction (over-accepting weak support vs. dropping real support), which determines whether TRUST/EBO gains partly reflect teacher style. A human audit of a sample of validated maps and map-aligned answers (acceptance precision, sufficiency agreement) is needed to anchor the central supervision-quality assumption.
minor comments (6)
  1. [Preliminaries, Eqs. (7)-(8)] The contraction |Ω_unstr|=(2^n−1)^m → |Ω_G|=1 is a definitional identity once a map is fixed, not an empirical or theoretical result; the text hedges this in one place but elsewhere claims results 'support our theoretical analysis.' Please present it uniformly as motivation.
  2. [Tables 6-7 captions] Captions read 'ASQA, QAMPARI, and ELI5' but the columns report ExpertQA; QAMPARI appears nowhere. Likely copy-paste from another paper's table template.
  3. [Table 2 analysis paragraph] 'CAGE improves TRUST by 8.47 and 11.10 points on ASQA' — the Claude-4.6 ASQA gain is 7.21 (11.10 is the ExpertQA figure). Also worth discussing: EM_F1_AC drops when CMI is prepended (66.18→61.41; 70.46→65.29), so the TRUST gain trades answer correctness against refusal/grounding.
  4. [Table 6 analysis text] 'CAGE_DPO further improves TRUST to 62.19%' for the 1.5B setting is lower than CAGE's 65.05 on ASQA; 'further improves' is inaccurate there.
  5. [Experimental settings] Exclusion of START, LongCite, and SelfCite is justified only as 'differences in reproducibility, task formulation, and evaluation protocol' — please specify concretely; LongCite in particular is a close fine-grained-citation baseline. Also report run variance: a single seed (42) is used, and some emphasized deltas (e.g., 0.71% in Fig. 3 discussion) are within typical fine-tuning noise.
  6. [General] Notation is inconsistent: F1CG vs F1GC, TRUST vs Trust, EM_F1_AC vs F1AC across tables; Table 1's final row has garbled formatting; EBO is evaluated only on ASQA — one sentence on dataset choice would help. No code/data release statement is given; releasing the VCAS corpus would substantially aid reproducibility.

Circularity Check

1 steps flagged

No load-bearing circular derivation: attribution-space contraction is a definitional framing of the method, and efficacy is tested on external TRUST/EBO benchmarks.

specific steps
  1. self definitional [Preliminaries, Attribution-space contraction (Eqs. 6–8); Abstract closing claim]
    "Without an attribution map, each of the m claims may select any non-empty subset of the n retained documents, yielding up to |Ω_unstr|=(2^n −1)^m. Once induced, the attribution map G fixes one citation set for each answer node: Ω_G={...}, |Ω_G|=1. Thus, conditioned on the induced map, CAGE reduces the downstream citation-assignment space from up to (2^n −1)^m unconstrained choices to one graph-specified target"

    Mild framing only: |Ω_G|=1 is true by definition of fixing α_G(a_r) per answer node, not an independent derivation. The paper pairs this identity with empirical TRUST/EBO gains and does not treat the identity alone as proof of better answers; score kept minimal because the load-bearing claim is experimental, not this tautology.

full rationale

CAGE’s central chain is methodological, not a first-principles prediction that collapses into its inputs. The identity |Ω_unstr|=(2^n−1)^m → |Ω_G|=1 (Preliminaries, Eqs. 7–8) holds by construction once a map fixes one citation set per answer node; the paper uses it as design rationale for contracting claim–document assignments, not as an independently discovered empirical law that is then “confirmed.” Efficacy is argued from held-out ASQA/ELI5/ExpertQA TRUST scores and EBO rates versus ICL, FRONT, T-ALIGN, Ground-GRPO, GPT-5.5, and Claude-4.6—external benchmarks and baselines, not refits of the same fitted quantity. VCAS supervision (GPT-5.5 generator/validator) and GPT-5.5 atomic decomposition for EBO are synthetic-label and evaluator choices that can bias style or selection (e.g., abstention), but they are not self-definitional reductions or uniqueness theorems imported from overlapping authors to forbid alternatives. SCR’s training objective aligns citations to α_G(a_r) by design of the two-stage pipeline; that is intentional supervision, then checked with TRUE citation metrics, not a circular proof that map induction is faithful. Author self-citations (e.g., Yan et al. 2025 on atomic facts) are background, not load-bearing uniqueness. Residual map error (15.42%) is openly reported. Overall: framing identity plus standard distill-and-evaluate; no step where a claimed prediction equals the fitted or defined input by construction in a way that carries the main result.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 3 invented entities

The load-bearing content is an engineering method plus LLM-as-judge synthetic supervision, not a parameter-free theory. Claims rest on standard RAG task assumptions, the TRUST/TRUE evaluation stack from prior work, teacher-LLM validity for maps/answers, and several design knobs (k=5, graph textual encoding, one-claim-per-subgraph realization).

free parameters (3)
  • top-k retrieved documents (k=5) = 5
    Fixed to 5 for VCAS and evaluation 'consistent with prior study'; changes evidence available to maps and EBO denominators.
  • CMI/SCR SFT hyperparameters = lr=1e-5, 3 epochs, seed=42
    Learning rate, epochs, batch sizes, warmup, AdamW settings chosen for training; affect whether maps/answers match VCAS targets.
  • VCAS corpus sizes after filtering = 7.8K maps / 2.3K answers
    Accepted map/answer counts (7.8K SCMI, 2.3K SSCR) depend on generator/validator thresholds and discard rules; supervision distribution is a free construction choice.
axioms (5)
  • domain assumption Attribution ambiguity—combinatorial claim–document assignments in end-to-end decoding—is a structural cause of citation error and evidence-boundary overrun.
    Stated in Introduction as the problem diagnosis; motivates map induction but is not independently proven beyond motivating examples and later EBO correlations.
  • ad hoc to paper Validator predicates Connected(G_r) and D_r |=_Q a_r (plus answer–node permutation alignment) are adequate operationalizations of sufficient evidential support.
    VCAS Steps 1–3; correctness of all supervision hinges on these LLM-checked predicates.
  • domain assumption TRUST = (1/3)(EM F1_AC + F1_RG + F1_CG) with TRUE-based citation NLI is an appropriate overall measure of citation-grounded trustworthiness.
    Adopted from Song et al. 2025; SOTA claims are relative to this composite.
  • domain assumption Cognitive-map theories justify externalizing latent answer–evidence relations as graphs before decoding.
    Cited (Mark et al., Whittington et al., Tan et al.) as inspiration in Introduction; not used as a testable neuroscientific claim.
  • standard math Autoregressive MLE on textual graph encodings and claim–citation sequences learns the intended structured attribution behavior.
    Standard seq2seq training assumption for CMI/SCR losses (Eqs. 22, 26).
invented entities (3)
  • Cognitive attribution map G = {G_r} of answer-centered support subgraphs no independent evidence
    purpose: Intermediate structure that fixes α_G(a_r) citation targets and contracts assignment space before answer realization.
    Core postulated representation (Preliminaries); schema and empty-map abstention are paper-specific.
  • Evidence-boundary overrun (A-EBO / S-EBO / R-EBO) independent evidence
    purpose: Quantify claims/atomic facts exceeding support of cited or retrieved documents.
    Named failure mode and metric suite in analysis/appendix; useful but defined for this evaluation setup.
  • VCAS / CMI / SCR two-stage pipeline no independent evidence
    purpose: Synthesize supervision and factor map induction from map-conditioned citation generation.
    Methodological decomposition introduced by the paper; plug-and-play CMI is the deployable artifact.

reviewed 2026-07-31 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CAGE: Cognitive Attribution Graphs for Faithful Inline Citation Generation in Long-Form Question Answering." pith.science (2026). https://pith.science/paper/4V5FRW6V

@misc{pith2026260724236,
  author       = {Pith},
  title        = {Pith review of: CAGE: Cognitive Attribution Graphs for Faithful Inline Citation Generation in Long-Form Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4V5FRW6V}},
  note         = {Machine review of arXiv:2607.24236}
}
Share X Bluesky LinkedIn Reddit HN
read the original 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.

Figures

Figures reproduced from arXiv: 2607.24236 by Haoran Luo, Jeff Z. Pan, Jiaoyan Chen, Jiapu Wang, Qingang Zhang, Ru Li, Shizhao Li, Zhichao Yan.

Figure 1
Figure 1. Figure 1: Motivation of CAGE. The human brain abstracts [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of CAGE. Stage 1 Validated Cognitive Attribution Synthesis constructs validated cognitive attribution map and map-aligned answer supervision, which are used to train the Cognitive Attribution Map Induction Model and the Structured Citation Reasoning Model, respectively, in Stage 2. Attribution-space contraction. The graph induces the re￾tained document set D+ = Sm r=1 Dr, with n = |D+|, and specif… view at source ↗
Figure 3
Figure 3. Figure 3: Effect of model scale on TRUST performance. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study. We fix the structured citation rea [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: A case study illustrating graph-based reasoning for multi-source information integration. Subgraph 0 identifies Anatoli [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

57 extracted references · 6 linked inside Pith

  1. [1]

    Unifying Large Language Models and Knowledge Graphs: A Roadmap , year=

    Pan, Shirui and Luo, Linhao and Wang, Yufei and Chen, Chen and Wang, Jiapu and Wu, Xindong , journal=. Unifying Large Language Models and Knowledge Graphs: A Roadmap , year=

  2. [2]

    Transactions on Graph Data and Knowledge , year=

    Large Language Models and Knowledge Graphs: Opportunities and Challenges , author=. Transactions on Graph Data and Knowledge , year=

  3. [3]

    Nature , volume=

    Detecting hallucinations in large language models using semantic entropy , author=. Nature , volume=. 2024 , publisher=

    2024

  4. [4]

    IEEE Transactions on Knowledge and Data Engineering , year=

    Atomic fact decomposition helps attributed question answering , author=. IEEE Transactions on Knowledge and Data Engineering , year=

  5. [5]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , year=

    Complex Claim Verification with Evidence Retrieved in the Wild , author=. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , year=

    2024

  6. [6]

    arXiv preprint arXiv:2212.08037 , year=

    Attributed question answering: Evaluation and modeling for attributed large language models , author=. arXiv preprint arXiv:2212.08037 , year=

    Pith/arXiv arXiv

  7. [7]

    Findings of the Association for Computational Linguistics ACL 2024 , pages=

    Learning Fine-Grained Grounded Citations for Attributed Large Language Models , author=. Findings of the Association for Computational Linguistics ACL 2024 , pages=

    2024

  8. [8]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Longcite: Enabling llms to generate fine-grained citations in long-context qa , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  9. [9]

    Enabling Large Language Models to Generate Text with Citations

    Gao, Tianyu and Yen, Howard and Yu, Jiatong and Chen, Danqi. Enabling Large Language Models to Generate Text with Citations. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023

    2023

  10. [10]

    ArXiv , year=

    RARR: Researching and Revising What Language Models Say, Using Language Models , author=. ArXiv , year=

  11. [11]

    Measuring and Enhancing Trustworthiness of

    Maojia Song and Shang Hong Sim and Rishabh Bhardwaj and Hai Leong Chieu and Navonil Majumder and Soujanya Poria , booktitle=. Measuring and Enhancing Trustworthiness of. 2025 , url=

    2025

  12. [12]

    Nature Machine Intelligence , pages=

    Towards unveiling sensitive and decisive patterns in explainable AI with a case study in geometric deep learning , author=. Nature Machine Intelligence , pages=. 2025 , publisher=

    2025

  13. [13]

    2024 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML) , pages=

    Unifying corroborative and contributive attributions in large language models , author=. 2024 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML) , pages=. 2024 , organization=

    2024

  14. [14]

    Advances in Neural Information Processing Systems , volume=

    Contextcite: Attributing model generation to context , author=. Advances in Neural Information Processing Systems , volume=

  15. [15]

    2003 , publisher=

    The description logic handbook: Theory, implementation and applications , author=. 2003 , publisher=

    2003

  16. [16]

    2024 , eprint=

    Improving Attributed Text Generation of Large Language Models via Preference Learning , author=. 2024 , eprint=

    2024

  17. [17]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    Effective large language model adaptation for improved grounding and citation generation , author=. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2024

  18. [18]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Advancing large language model attribution through self-improving , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  19. [19]

    International Conference on Machine Learning , pages=

    SelfCite: Self-Supervised Alignment for Context Attribution in Large Language Models , author=. International Conference on Machine Learning , pages=. 2025 , organization=

    2025

  20. [20]

    arXiv preprint arXiv:2601.06021 , year=

    Chaining the Evidence: Robust Reinforcement Learning for Deep Search Agents with Citation-Aware Rubric Rewards , author=. arXiv preprint arXiv:2601.06021 , year=

    arXiv

  21. [21]

    First Conference on Language Modeling , year=

    Source-Aware Training Enables Knowledge Attribution in Language Models , author=. First Conference on Language Modeling , year=

  22. [22]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Transparentize the internal and external knowledge utilization in LLMs with trustworthy citation , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  23. [23]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    Chain-of-Thought Improves Text Generation with Citations in Large Language Models , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2024 , pages=

    2024

  24. [24]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Think&Cite: Improving attributed text generation with self-guided tree search and progress reward modeling , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  25. [25]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Attribute first, then generate: Locally-attributable grounded text generation , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  26. [26]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Laquer: Localized attribution queries in content-grounded generation , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  27. [27]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Rarr: Researching and revising what language models say, using language models , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  28. [28]

    arXiv preprint arXiv:2503.21460 , year=

    Large language model agent: A survey on methodology, applications and challenges , author=. arXiv preprint arXiv:2503.21460 , year=

    Pith/arXiv arXiv

  29. [29]

    Advances in neural information processing systems , volume=

    Direct preference optimization: Your language model is secretly a reward model , author=. Advances in neural information processing systems , volume=

  30. [30]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Large language models are neurosymbolic reasoners , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  31. [31]

    Findings of the association for computational linguistics: NAACL 2024 , pages=

    Language models can be deductive solvers , author=. Findings of the association for computational linguistics: NAACL 2024 , pages=

    2024

  32. [32]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Faithful logical reasoning via symbolic chain-of-thought , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  33. [33]

    Nature , volume=

    Solving olympiad geometry without human demonstrations , author=. Nature , volume=. 2024 , publisher=

    2024

  34. [34]

    International Conference on Learning Representations , volume=

    Large language models meet symbolic provers for logical reasoning evaluation , author=. International Conference on Learning Representations , volume=

  35. [35]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Interactive evolution: A neural-symbolic self-training framework for large language models , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  36. [36]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Aristotle: Mastering logical reasoning with a logic-complete decompose-search-resolve framework , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  37. [37]

    arXiv preprint arXiv:2512.18196 , year=

    LogicReward: Incentivizing LLM Reasoning via Step-Wise Logical Supervision , author=. arXiv preprint arXiv:2512.18196 , year=

    arXiv

  38. [38]

    ASQA : Factoid Questions Meet Long-Form Answers

    Stelmakh, Ivan and Luan, Yi and Dhingra, Bhuwan and Chang, Ming-Wei. ASQA : Factoid Questions Meet Long-Form Answers. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022

    2022

  39. [39]

    ELI 5: Long Form Question Answering

    Fan, Angela and Jernite, Yacine and Perez, Ethan and Grangier, David and Weston, Jason and Auli, Michael. ELI 5: Long Form Question Answering. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019

    2019

  40. [40]

    E xpert QA : Expert-Curated Questions and Attributed Answers

    Malaviya, Chaitanya and Lee, Subin and Chen, Sihao and Sieber, Elizabeth and Yatskar, Mark and Roth, Dan. E xpert QA : Expert-Curated Questions and Attributed Answers. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024

    2024

  41. [41]

    2025 , eprint=

    Lessons from Training Grounded LLMs with Verifiable Rewards , author=. 2025 , eprint=

    2025

  42. [42]

    arXiv preprint arXiv:2508.08942 , year=

    Jointly Generating and Attributing Answers using Logits of Document-Identifier Tokens , author=. arXiv preprint arXiv:2508.08942 , year=

    Pith/arXiv arXiv

  43. [43]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Deciphering the interplay of parametric and non-parametric memory in retrieval-augmented language models , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  44. [44]

    International Conference on Learning Representations , volume=

    Redeep: Detecting hallucination in retrieval-augmented generation via mechanistic interpretability , author=. International Conference on Learning Representations , volume=

  45. [45]

    Communications Biology , volume=

    The medial and lateral orbitofrontal cortex jointly represent the cognitive map of task space , author=. Communications Biology , volume=. 2025 , publisher=

    2025

  46. [46]

    Nature communications , volume=

    Transferring structural knowledge across cognitive maps in humans and models , author=. Nature communications , volume=. 2020 , publisher=

    2020

  47. [47]

    Nature neuroscience , volume=

    How to build a cognitive map , author=. Nature neuroscience , volume=. 2022 , publisher=

    2022

  48. [48]

    arXiv preprint arXiv:2605.28044 , year=

    Relevant is not warranted: Evidence-force calibration for cited RAG , author=. arXiv preprint arXiv:2605.28044 , year=

    Pith/arXiv arXiv

  49. [49]

    FA ct S core: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation

    Min, Sewon and Krishna, Kalpesh and Lyu, Xinxi and Lewis, Mike and Yih, Wen-tau and Koh, Pang and Iyyer, Mohit and Zettlemoyer, Luke and Hajishirzi, Hannaneh. FA ct S core: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023

    2023

  50. [50]

    Proceedings of the 38th International Conference on Neural Information Processing Systems , pages=

    Large language models-guided dynamic adaptation for temporal knowledge graph reasoning , author=. Proceedings of the 38th International Conference on Neural Information Processing Systems , pages=

  51. [51]

    arXiv preprint arXiv:2308.02457 , year=

    A survey on temporal knowledge graph completion: Taxonomy, progress, and prospects , author=. arXiv preprint arXiv:2308.02457 , year=

    Pith/arXiv arXiv

  52. [52]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    GraphRAG-induced dual knowledge structure graphs for personalized learning path recommendation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  53. [53]

    Proceedings of the ACM on Web Conference 2025 , pages=

    Effective instruction parsing plugin for complex logical query answering on knowledge graphs , author=. Proceedings of the ACM on Web Conference 2025 , pages=

    2025

  54. [54]

    arXiv preprint arXiv:2508.01290 , year=

    Prompting large language models with partial knowledge for answering questions with unseen entities , author=. arXiv preprint arXiv:2508.01290 , year=

    Pith/arXiv arXiv

  55. [55]

    arXiv preprint arXiv:2509.00765 , year=

    Decomposing and Revising What Language Models Generate , author=. arXiv preprint arXiv:2509.00765 , year=

    arXiv

  56. [56]

    EFSP-TE: End-to-End Frame-Semantic Parsing with Table Encoder , year=

    Su, Xuefeng and Li, Ru and Li, Xiaoli and Yan, Zhichao , journal=. EFSP-TE: End-to-End Frame-Semantic Parsing with Table Encoder , year=

  57. [57]

    Proceedings of the 31st International Conference on Computational Linguistics , pages=

    LOG: A Local-to-Global Optimization Approach for Retrieval-based Explainable Multi-Hop Question Answering , author=. Proceedings of the 31st International Conference on Computational Linguistics , pages=

This paper was first reviewed by grok-4.5 on July 31, 2026.