{"id":"e4081bd9-3d4b-446a-a2ee-96b2f937042f","arxiv_id":"2608.00711","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SCHEMA shows that LLM science agents often reach correct answers through reasoning steps that contain factual errors, and that these errors cluster around a few highly connected biology concepts.","lead":"This paper introduces SCHEMA, a benchmark and scoring system that checks whether AI science agents reason honestly, not just whether their final answers are correct. It builds a knowledge graph of biology concepts and uses it to measure where hallucinations happen and how much damage they can do.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Decoupling claim is not measured: Appendix E.3 restricts the severity pipeline to incorrect trajectories, so HSw cannot show correct answers reached via flawed logic.","rationale":"The reader's weakest assumption concerns LLM-judge calibration; the concern raised here is more direct because it targets the quantity being measured rather than the accuracy of the labels. Even a perfectly calibrated judge cannot support the decoupling claim if the metric is computed only on incorrect trajectories. The reader's rationale does list 'hallucination severity is measured only on incorrect trajectories' as a weakening design choice, so there is partial overlap, but the weakest_assumption field focuses on judge validation instead. I retain the reader's CONDITIONAL verdict, with the condition reframed: the authors must report hallucination metrics on the correct-answer subset (or demonstrate that the reported HSw is not error-targeted) before the accuracy-honesty decoupling can be considered established.","tokens_in":28060,"tokens_out":8110,"duration_ms":74158,"concrete_test":"Inspect the released trajectory logs and pipeline configuration to determine whether Table 3's HSw was computed on all trajectories or only incorrect ones. Then re-run the hallucination pipeline (same extractor/judge as E.3) on the subset of trajectories whose terminal answer matches gold, and report HR, HS, and HSw for that correct-answer subset separately. If correct-subset hallucination severity is near zero, the decoupling/shortcut-reasoning claim is unsupported; if substantially above zero, the claim is supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline claim that 'final-answer accuracy decouples from trajectory honesty—models often reach correct conclusions through structurally flawed reasoning' (abstract; Section 4.2) requires measuring hallucinations on trajectories with correct terminal answers. The evaluation pipeline, however, is error-targeted: Appendix E.3 states it 'defaults to an error-targeted mode (is_correct=false), computing severity exclusively on trajectories that yield incorrect terminal answers.' The HS/HSw values in Table 3 and Figure 5 are therefore conditional on terminal failure. The paired example in Section 4.2—DeepSeek-V4-Flash and GPT-5.4-mini both at 0.59 accuracy with HSw 0.362 vs 0.268—shows only that failed trajectories differ in severity across models with equal accuracy. It does not establish that either model's correct trajectories contain refuted or unverifiable claims. If the default was used, the central empirical result is not measured; if it was overridden, the paper does not report the correct-trajectory breakdown needed to support the shortcut-reasoning claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SCHEMA, an automated pipeline that builds a biomedical concept graph from benchmark seeds and literature evidence, synthesizes five types of evidence-grounded questions (claim choice, boolean support, two-hop tail, essay, and experiment code), and evaluates LLM agents using terminal accuracy plus two trajectory-level diagnostics: a topology-weighted hallucination severity metric (HS/HSw) and a multi-agent counterfactual attribution module. Experiments on Protein Domain and PathVQA-Enhanced across 10 LLMs are used to claim that hallucinations concentrate at a small set of knowledge hubs and that final-answer accuracy decouples from trajectory honesty. The manuscript is unusually detailed in its appendices and plans to release code and trajectory logs.","tokens_in":28251,"tokens_out":6131,"duration_ms":56688,"significance":"If the empirical claims survive scrutiny, SCHEMA would be a valuable contribution: it makes the notion of topology-weighted hallucination severity concrete, contributes a reproducible biomedical agent benchmark, and articulates a falsifiable claim about the relationship between terminal accuracy and intermediate-trajectory quality. The pipeline is described in enough detail to be reproduced, and the planned release of per-model trajectories is a genuine strength. However, the headline empirical findings are not currently supported as stated: the hallucination metrics are computed only on incorrect trajectories, and the underlying per-claim labels come from an unvalidated LLM judge. These issues affect the central claims and require substantive revision rather than copyediting.","major_comments":[{"comment":"The hallucination pipeline 'defaults to an error-targeted mode (is_correct=false), computing severity exclusively on trajectories that yield incorrect terminal answers.' The abstract and Section 4.2 claim that 'final-answer accuracy decouples from trajectory honesty' and that models 'frequently reach correct terminal answers via flawed intermediate logic.' These claims require measuring hallucination severity on correct-answer trajectories. As reported, the HSw values in Table 3 and Figure 5 are conditional on terminal failure, so the paired DeepSeek-V4-Flash / GPT-5.4-mini example shows only that failed trajectories differ in severity; it does not show that either model's correct trajectories contain refuted or unverifiable claims. Please recompute HS/HSw on all trajectories, or at least report the correct-trajectory breakdown separately, and revise the decoupling claim accordingly.","section":"Appendix E.3; Section 4.2; Table 3; Figure 5"},{"comment":"The entire severity computation rests on the LLM judge's supported/refuted/unverifiable labels, and Appendix B acknowledges that the judge 'has not been validated against expert annotations.' This is load-bearing because every HS/HSw value in Table 3 and Figure 6 inherits the judge's verdicts. Without a calibration study on a representative subset (e.g., per-category agreement rates against domain experts), the reported differences, such as HSw 0.362 versus 0.268, cannot be separated from judge bias. Please either provide such a validation or qualify all headline comparisons as contingent on unvalidated automated verdicts.","section":"Section 3.3; Section 3.4; Appendix B"},{"comment":"The subgraph sampling procedure in Appendix E.1 is not degree-stratified, so high-degree concepts appear in more candidate triples and therefore in more questions. The finding that hallucinations 'concentrate at a small set of highly connected knowledge hubs' may consequently reflect the question distribution rather than model behavior. Table 2 reports counts by Core/Peripheral/Off-graph tiers, but no per-exposure error rates are given. Please report error rates per concept conditional on the number of questions involving that concept, or subsample subgraphs in a degree-stratified manner, before drawing the hub-concentration conclusion.","section":"Section 3.3; Table 2; Section 4.2"},{"comment":"The text states that 'ENZYMATIC/CATALYSIS alone accounts for nearly half of all hallucinations,' but Figure 2 shows n=43 out of 248 hallucinated concepts, which is about 17%. Unless a different subset is intended, this statement is inconsistent with the figure and should be corrected or clarified.","section":"Section 2.3; Figure 2"}],"minor_comments":[{"comment":"The counterfactual attribution findings in Section 2.3 are based on a single agent model (Intern-S1-Pro) and 141 failure trajectories from ProteinLMBench; this should be explicitly labeled as a case study rather than a general empirical finding.","section":"Section 2.2"},{"comment":"The HSw values along the verifiability axis range only from 0.381 to 0.436; the claim that severity 'rises monotonically' should be accompanied by confidence intervals or a statistical test, especially given the absence of error bars acknowledged in the checklist.","section":"Figure 6"},{"comment":"There are several presentation issues: Reference [24] contains a duplicated author name 'XingyuHan XingyuHan', Figure 5(c) contains the typo 'Biomaker', and the naming of 'ProteinLMBench' versus 'Protein Domain' should be made consistent throughout.","section":"References; Figure 5"},{"comment":"The fact that HS/HSw are computed only on incorrect trajectories should be stated prominently in the main text near Table 3, not only in an appendix, because it materially changes how the numbers should be interpreted.","section":"Main text; Appendix E.3"},{"comment":"The phrase 'the first evidence-grounded, topology-aware evaluation framework' is an overclaim; nearby works such as KGHaluBench also use knowledge-graph grounding and topology. Please soften the novelty claim or add a more precise comparison.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The framework and benchmark contributions are solid and within the journal's scope. The main risk is that the abstract's central decoupling claim is directly contradicted by the error-targeted mode described in Appendix E.3, and the judge-validation gap is acknowledged but not quantified. I recommend major revision rather than rejection because both issues are addressable within the manuscript's scope. Please ask the authors to clarify or re-run the severity pipeline on all trajectories."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: SCHEMA is a carefully constructed evaluation system and deserves a serious referee, but the main empirical claim is not supported by the reported measurement. Appendix E.3 says the hallucination pipeline defaults to an error-targeted mode and computes severity only on trajectories with incorrect terminal answers. That means Table 3 and Figure 5 show how bad failed trajectories are, not whether models reach correct answers through flawed reasoning. The paired DeepSeek-V4-Flash / GPT-5.4-mini example just says their failures differ in severity at equal accuracy. The abstract and Section 4.2 overstate the decoupling claim. This is a load-bearing flaw in the empirical section, and I agree with the stress-test note.\n\nWhat is genuinely new: the integration. The paper combines a graph-grounded benchmark, five task types, trajectory-level claim auditing, a topology-weighted severity score, and counterfactual attribution in one framework. The HSw metric is simple and transparent, degree-weighted with an off-graph baseline, and the authors report unweighted HS alongside it. The documentation is unusually thorough: prompts, thresholds, verdict mappings, judge-allocation rules, and budget details are all there. The limitations section is honest about the unvalidated LLM judge and the absence of expert calibration. The citation pattern looks normal; the authors engage the relevant AgentHallu, MIRAGE-Bench, and TRACE lines.\n\nSoft spots, in proportion: the decoupling claim is the big one. After that, the hub-concentration result lacks a concept-exposure baseline; if hub concepts appear more often in questions and in agent reasoning, concentration is partly an artifact of exposure. Also, the same auto-built graph seeds the benchmark questions and supplies the HSw degree weights, so the instrument is not fully independent of what it measures. That is not fatal, but it should be discussed. The missing error bars are acknowledged and are a minor issue for this kind of benchmarking.\n\nWho this is for: people building agent-evaluation benchmarks and anyone measuring hallucination in scientific agents. The framework and artifacts are reusable even if the headline findings need re-derivation. I would send it to peer review, but a referee should require correct-trajectory severity numbers, an exposure baseline, and some expert-calibrated judge evaluation before the decoupling claim is accepted. As it stands, the framework is the contribution; the strong empirical conclusions are not yet earned.","headline":"Well-built, unusually transparent evaluation framework; but the headline claim that accuracy decouples from trajectory honesty is not actually measured by the error-targeted pipeline.","tokens_in":733,"tokens_out":987,"would_cite":true,"duration_ms":38225,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Scientific agents often reach the correct final answer while reasoning through claims that are false or unverifiable, and these errors concentrate at a small set of highly connected concept hubs.","keywords":["SCHEMA","hallucination evaluation","scientific agents","knowledge graph","topology-weighted severity","counterfactual attribution","LLM-as-judge","biomedical benchmark"],"falsifier":"Sample roughly two hundred judged claims across several models and verdict categories, and have domain experts label each claim against the same evidence the judge saw, without knowing the judge's verdict. If expert labels disagree with the judge on a large share of claims, or if the disagreements concentrate on the hub concepts that dominate $HS_w$, then the hub-concentration and decoupling results do not survive human-level grounding. A secondary check is to rebuild the concept graph with a different extraction model and re-run the metric: stable hub sets would confirm the topology, while drifting sets would trace the finding to the pipeline.","tokens_in":27856,"feed_emoji":"🧬","tokens_out":12291,"duration_ms":98076,"temperature":0.7,"pith_summary":"This paper argues that judging a scientific AI agent by whether its final answer is correct leaves the main danger undetected: agents frequently produce right answers by reasoning through intermediate claims that are unsupported or false. To expose this, the authors build SCHEMA, an evaluation framework that constructs a biomedical knowledge graph from benchmark seeds and literature, generates graph-grounded tasks of five types, and audits agents' full reasoning trajectories with a topology-weighted severity metric alongside a counterfactual attribution module. The paper's empirical findings are that hallucinations are not random slips but concentrate at a small set of highly connected concept hubs, and that final-answer accuracy correlates only weakly with trajectory honesty across ten models. If these findings hold, reliability evaluation for scientific agents needs to become structural, weighing where an error sits in the knowledge topology and how far it propagates, rather than collapsing performance into a single accuracy number.","feed_headline":"AI scientists can nail the answer and still hallucinate","feed_subtitle":"Correct final answers often rest on false intermediate claims; a topology-aware metric catches what accuracy misses.","key_machinery":"The load-bearing object is the topology-weighted hallucination severity metric $HS_w$, computed over the scientific concept graph $G$. Each extracted claim $c$ is mapped to a node $\\eta(c)$ in the graph by lexical matching with an embedding-based fallback, and receives weight $w_g(c) = 1 + \\log(1 + \\deg_K(\\eta(c)))$, a log-scaled bonus that grows with the node's undirected degree; claims that map to no node fall back to $w_g = 1$. Judge verdicts (supported, refuted, unverifiable) map deterministically to severities $0, 0.5, 1$, and $HS_w$ is the weighted average over claims, so an error at a highly connected concept is penalized more than the same error at a peripheral one, reflecting its downstream blast radius. The framework pairs this scalable pipeline with a multi-agent counterfactual attribution module that follows abduction, intervention, and prediction: hallucinated claims are verified against literature, a minimally edited trajectory replaces them with corrections, and, with prior context frozen by the trajectory debugger, a candidate subset is certified as the root cause only if the rerun strictly flips the outcome.","core_discovery":"The paper's central claim is that hallucination in scientific agents is a structural phenomenon with a topological signature. On the Protein Domain subset, a majority of reasoning failures trace to flawed domain-specific priors rather than workflow problems, and hallucinated claims concentrate around a small set of recurring concept hubs, with one family, enzymatic catalysis, accounting for a large share of the hallucinated concepts, while long-tail peripheral concepts show near-zero repairability. The second headline claim is that terminal accuracy decouples from trajectory honesty: models often reach correct final answers through flawed intermediate logic, a shortcut-reasoning pattern that persists across two nearly disjoint concept spaces, so that two models can tie on accuracy yet differ sharply in topology-weighted hallucination severity. Taken together, the paper argues that terminal accuracy alone is an insufficient signal of agent reliability in high-stakes science, and that evaluation should weight claims by their structural role and trace failures back to their causal mechanism.","pith_inferences":["The hub-concentration finding suggests a repair strategy the paper does not develop: if a few high-traffic concepts carry most hallucinations, then injecting verified grounding for just those hubs could yield disproportionate reliability gains without broad retraining.","The reported partial-anchoring valley, where weak graph context hurts accuracy more than no context for most models, points to a general and testable principle: agents over-trust fragmentary retrieved structure, which would predict that how context is presented matters as much as how much of it is given.","A direct robustness test is to rebuild the concept graph with a different extraction model and judge family; if the same concepts still dominate the hallucination distribution and the accuracy–honesty decoupling survives, the findings are properties of the agents rather than of the measurement chain."],"forward_implications":["Rankings of scientific agents will change when trajectory honesty replaces terminal accuracy as the criterion; two models that tie on accuracy can differ sharply in topology-weighted severity, so deployment decisions for high-stakes science should report both families of metrics.","Hallucination benchmarks that treat all facts as interchangeable will mislead: a faithful evaluation must weight concepts by their structural role, since errors at hubs propagate through every downstream reasoning chain that passes through them.","Failure repair becomes predictable by error type: mapping errors dominate frequency but are comparatively recoverable, while fact errors have the lowest repairability because later reasoning steps inherit and amplify a false premise.","A single aggregate accuracy figure masks format-specific weaknesses: models can score near chance on boolean support judgments while excelling at experimental code generation, so results need to be reported by question type and evidence tier."],"supporting_citations":[{"why":"ProteinLMBench supplies the human-verified benchmark seeds for the Protein Domain graph and the 346 questions whose 141 failure trajectories are counterfactually analyzed.","marker":"[19]"},{"why":"PathVQA provides the second, more interconnected biomedical subdomain, grown into PathVQA-Enhanced to show the pipeline transfers across concept spaces.","marker":"[20]"},{"why":"The trajectory debugger supplies per-step tracking and state restoration, which makes the frozen-context counterfactual reruns possible.","marker":"[16]"},{"why":"Supplies the abduction–intervention–prediction procedure that the multi-agent counterfactual attribution module operationalizes.","marker":"[15]"},{"why":"Defines the ReAct-style reasoning-and-acting trajectory format that both the hallucination pipeline and the attribution module audit.","marker":"[14]"},{"why":"The prior agent hallucination attribution benchmark whose costly per-trajectory expert inspection SCHEMA's counterfactual module is designed to automate.","marker":"[10]"},{"why":"Prior work locating the exact reasoning steps where agent hallucinations arise; the trajectory-auditing baseline that SCHEMA extends toward scientific domains.","marker":"[8]"}],"fun_headline_variants":["Hallucination hubs: AI scientists' false claims cluster","Right answers, wrong reasoning: AI scientists' hidden flaw","Topology-aware score catches AI's hidden reasoning lies","AI hallucinations aren't random—they hit knowledge hubs","Accuracy can't hide AI's flawed reasoning: new metric"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the automatic fact-checking chain being reliable: if the LLM judge frequently mislabels an agent's claims as false or unsupported, and the paper itself notes the judge has not been validated against expert annotations, then the hub concentration and the decoupling of accuracy from trajectory honesty could be artifacts of the measurement rather than real properties of the agents.","fun_headline_variants_meta":{"raw":{"variants":["Hallucination hubs: AI scientists' false claims cluster","Right answers, wrong reasoning: AI scientists' hidden flaw","Topology-aware score catches AI's hidden reasoning lies","AI hallucinations aren't random—they hit knowledge hubs","Accuracy can't hide AI's flawed reasoning: new metric"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001197,"raw_usage":{"total_tokens":4951,"prompt_tokens":976,"completion_tokens":3975,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":3896}},"tokens_in":592,"tokens_out":3975,"duration_ms":25814,"temperature":1.0,"reasoning_tokens":3896,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:17:23.041622+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sample roughly two hundred judged claims across several models and verdict categories, and have domain experts label each claim against the same evidence the judge saw, without knowing the judge's verdict. If expert labels disagree with the judge on a large share of claims, or if the disagreements concentrate on the hub concepts that dominate $HS_w$, then the hub-concentration and decoupling results do not survive human-level grounding. A secondary check is to rebuild the concept graph with a different extraction model and re-run the metric: stable hub sets would confirm the topology, while drifting sets would trace the finding to the pipeline.","supporting_citations":[{"cited_title":"A fine-tuning dataset and benchmark for large language models for protein understanding","cited_arxiv_id":null,"evidence_quote":"ProteinLMBench supplies the human-verified benchmark seeds for the Protein Domain graph and the 346 questions whose 141 failure trajectories are counterfactually analyzed."},{"cited_title":"Interactive debugging and steering of multi-agent ai systems","cited_arxiv_id":null,"evidence_quote":"The trajectory debugger supplies per-step tracking and state restoration, which makes the frozen-context counterfactual reruns possible."}],"review_version":2}