{"id":"07afbabe-166f-44e7-b753-8619614468e3","arxiv_id":"2607.24236","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Explicit cognitive attribution graphs before generation contract claim–document assignment space and yield SOTA faithful inline citations on long-form QA benchmarks.","lead":"CAGE inserts an explicit answer–document support graph before long-form QA generation so citations are fixed by the graph rather than guessed during decoding. That design cuts unsupported claims and lifts TRUST scores across open and proprietary LLMs on ASQA, ELI5, and ExpertQA.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The headline EBO reduction (Table 3) is confounded by differential abstention and answer style: CAGE refuses more and writes short single-claim sentences, so its valid-response pool is easier and its atomic-fact decomposition cleaner, mechanically lowering unsupported-fact rates independent of truee","rationale":"The reader identified teacher-fidelity of GPT-5.5-synthesized maps as the weakest assumption. That is a real concern, but it is partially mitigated by the w/o-graph ablation (Figure 4), which shows the structure itself — not just teacher-flavored documents — drives TRUST gains, and by the cross-backbone consistency in Tables 5–8. The sharper, more load-bearing soft spot sits downstream in the evaluation design: the EBO numbers in Table 3, which the strongest claim explicitly cites (\"substantially reducing evidence-boundary overrun versus strong end-to-end baselines including GPT-5.5 and Claude-4.6\"), are computed over self-selected valid-response pools with refusal rates that differ systematically across compared systems, plus a decomposition-granularity asymmetry induced by CAGE's own one-sentence-per-subgraph output format. This is an internal-consistency issue in the experimental design, not a consensus disagreement, and it is cheaply checkable with existing outputs (no new training needed). It does not overturn the paper: the TRUST gains, the ablations, and the plug-and-play results stand independently, and even a halved EBO gap would still favor CAGE. Hence the verdict stays CONDITIONAL, with the condition sharpened: artifact release plus a common-support EBO re-analysis should accompany the requested human validation of synthetic maps.","tokens_in":26692,"tokens_out":1999,"duration_ms":68739,"concrete_test":"Recompute Table 3 on the intersection of ASQA questions where all five systems produced non-refusal answers (common-support restriction), and report each system's refusal rate alongside. Additionally stratify Atomic-EBO by atomic-facts-per-answer (or cap atoms per answer) to control decomposition granularity. If CAGE's A-EBO advantage over GPT-5.5/Claude-4.6 shrinks by more than half under the common-support restriction, the EBO headline is driven mainly by abstention selection and answer style rather than attribution contraction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim rests partly on Table 3: CMI+SCR cuts Atomic-EBO from 38.00%/40.85% (GPT-5.5/Claude-4.6) to 5.77–7.21%. But the EBO protocol (Appendix, Eqs. 28–30) excludes refusal and empty responses from N. CAGE has an explicit, trained abstention channel (empty-map supervision, Eq. 11) and its F1_RG gains in Tables 1–2 show it refuses more often than end-to-end baselines. Refusals are concentrated on questions where retrieved evidence is weak — precisely the questions where unsupported atomic facts are most likely. End-to-end GPT-5.5/Claude answer nearly everything, so their EBO pool includes the hard tail that CAGE opts out of. The comparison therefore mixes a treatment effect with a selection effect: CAGE's EBO is computed over a subpopulation enriched for well-supported questions. Second, the SCR training 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, whereas fluent end-to-end paragraphs decompose into more numerous, discourse-laden atoms with higher false-unsupported rates. Both confounds point the same direction, so the ~30-point EBO gap likely overstates the genuine attribution improvement. This does not touch the TRUST/F1_GC gains directly, but \"substantially reducing evidence-boundary overrun versus GPT-5.5/Claude-4.6\" is part of the strongest claim and is the least secure element of it.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","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.","tokens_in":27180,"tokens_out":3763,"duration_ms":120881,"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":[{"comment":"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).","section":"Table 3; Appendix Eqs. (28)-(30)"},{"comment":"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.","section":"Table 3; SCR training prompt (Appendix)"},{"comment":"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.","section":"VCAS, Steps 1-3 (Eqs. 10-18); Limitations"}],"minor_comments":[{"comment":"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.","section":"Preliminaries, Eqs. (7)-(8)"},{"comment":"Captions read 'ASQA, QAMPARI, and ELI5' but the columns report ExpertQA; QAMPARI appears nowhere. Likely copy-paste from another paper's table template.","section":"Tables 6-7 captions"},{"comment":"'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.","section":"Table 2 analysis paragraph"},{"comment":"'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.","section":"Table 6 analysis text"},{"comment":"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.","section":"Experimental settings"},{"comment":"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.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core TRUST results look solid and the ablations are better-controlled than typical for this subfield. My hesitation is concentrated in the EBO headline: as constructed, the metric comparison is almost certainly inflated by differential abstention and answer style, and the authors should be given the chance to re-run it on matched answered subsets before the 'substantially reduces evidence-boundary overrun' claim stands. The full reliance on GPT-5.5 as teacher, validator, and (partially) evaluator is a moderate circularity risk the authors should address with a human audit."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: CAGE is a competent two-stage citation system that externalizes claim–document support as answer-centered subgraphs before decoding. The useful idea is treating attribution as a trained intermediate object (CMI) that contracts assignment choices and plugs into heterogeneous generators, including frozen proprietary APIs.\n\nWhat is actually new is not “plan then generate” in the abstract—they cite Attribute First—but the full package: validated map synthesis (VCAS), empty-map abstention, map-aligned SCR supervision, and broad measurement (TRUST across Qwen/LLaMA scales, structure vs filter ablations, cross-backbone plug-in, and their EBO suite). The ablations matter: dropping graph structure while keeping the same documents hurts a lot more than dropping descriptions, which supports the claim that structure does real work beyond document pruning. Scale sweeps also land cleanly—map quality moves TRUST more than SCR size. Honest limitation on ~15% map error propagation is a plus.\n\nSoft spots, in proportion. Synthetic labels from GPT-5.5 generator/validator are the load-bearing supervision story; gains partly track teacher style, and artifacts are not shipped. The combinatorial “contraction to |Ω_G|=1” is definitional once a map exists—framing, not a law—but they mostly use it as motivation, not proof. The stress-test on Table 3 is fair: EBO excludes refusals, CAGE is trained to abstain on weak evidence, and SCR is prompted into short single-claim sentences, so the big Atomic-EBO drop vs end-to-end GPT-5.5/Claude mixes selection and style with true attribution improvement. I would not hang the paper on “substantially lower EBO than frontier end-to-end.” TRUST / F1_GC and the w/o-graph ablation are the sturdier evidence.\n\nWho it is for: people building cited long-form RAG and evaluation (TRUST, refusal, claim-level grounding). Worth a serious referee. I would engage, cite the pipeline and plug-and-play results if I work on attribution, and discount the headline EBO comparison until they report EBO conditional on answering the same items with matched verbosity.","headline":"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.","tokens_in":28098,"tokens_out":558,"would_cite":true,"duration_ms":15639,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Building an explicit answer–document map before writing cuts citation ambiguity and keeps long-form answers inside their evidence.","keywords":["inline citation generation","long-form question answering","attribution ambiguity","cognitive attribution map","evidence-boundary overrun","retrieval-augmented generation","claim-document alignment","TRUST evaluation"],"falsifier":"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.","tokens_in":27846,"feed_emoji":"🗺️","tokens_out":936,"duration_ms":19994,"temperature":0.7,"pith_summary":"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.","feed_headline":"Map claims to evidence first, then write the answer","feed_subtitle":"Explicit support graphs shrink citation guesswork and cut claims that outrun their sources","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["Build the attribution map before writing the answer","Contract citation space with explicit support graphs first","Map each answer unit to evidence, then generate claims","Explicit cognitive maps cut evidence-boundary overrun","Induce support subgraphs, then cite from the map"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Build the attribution map before writing the answer","Contract citation space with explicit support graphs first","Map each answer unit to evidence, then generate claims","Explicit cognitive maps cut evidence-boundary overrun","Induce support subgraphs, then cite from the map"]},"model":"grok-4.5","effort":"low","cost_usd":0.002286,"raw_usage":{"total_tokens":930,"prompt_tokens":769,"num_sources_used":0,"completion_tokens":53,"cost_in_usd_ticks":22864000,"prompt_tokens_details":{"text_tokens":769,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":108,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":769,"tokens_out":53,"duration_ms":4069,"temperature":1.0,"reasoning_tokens":108,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T20:17:50.538627+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"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.","supporting_citations":[],"review_version":1}