{"id":"14d2a7e8-e621-4497-8f77-cfd04f812424","arxiv_id":"2608.00200","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Attribution-guided DAG reasoning traces distilled into a compact sensor-conditioned LM give 84.43% average accuracy and 81.24% F1 across seven HAR datasets, with SNM as an LLM-judge diagnostic for trace fidelity.","lead":"TRACE-TS trains a compact language model to output both an activity label and a structured, evidence-linked reasoning trace for wearable sensor windows, using saliency maps from an expert classifier to build the training data. It reports top average accuracy/F1 across seven HAR benchmarks and introduces an LLM-judge metric, SNM, that detects reasoning errors standard text metrics miss.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 'signal-grounded' claim rests on teacher-generated references that are both training targets and SNM references; a signal-level grounding audit is missing, so reasoning fidelity is not independently verified.","rationale":"The reader's weakest assumption is exactly the load-bearing issue. The paper's central claim is not merely high accuracy; it is that TRACE-TS produces traceable reasoning grounded in sensor evidence. That claim requires an evaluation independent of the teacher. SNM, by construction, compares against teacher traces that are also the training targets (Eq. 3), so it cannot distinguish 'grounded in signal' from 'matches teacher.' The Table 8 perturbation analysis is strong evidence that SNM measures something about evidence structure, but it does not test whether teacher/student observations correspond to actual signal content. The human evaluation is the only external check, but one window per class is insufficient to certify grounding, and no per-observation signal-level verification or annotator agreement is reported. The authors are transparent about this in Section 6, which makes this a missing verification rather than a hidden flaw. This does not overturn the classification results or the framework's utility; the accuracy gains and ablations stand. But the 'attribution-grounded and traceable reasoning' headline is conditional on an external grounding audit. Since the reader already set CONDITIONAL and identified this as the weakest assumption, no verdict change is needed. Secondary issues such as single training runs and modified benchmark protocols affect the precision of the accuracy claims but not the core reasoning-grounding claim.","tokens_in":38974,"tokens_out":4831,"duration_ms":57529,"concrete_test":"Run a signal-level grounding audit on a held-out set stratified across all seven datasets (e.g., 50 windows per dataset, not used for training). For each generated observation node, extract the cited channel and temporal interval from the raw sensor window and check mechanically, or via an independent annotator blind to model/teacher, whether the claimed pattern actually occurs: peak, oscillation, near-baseline, sustained elevation. Apply the identical audit to the teacher reference traces. Report per-dataset agreement/precision and chance baseline. If student observation grounding is near chance while SNM remains high, the self-referential evaluation is masking ungrounded reasoning and the grounding claim fails; if student and teacher grounding are both high, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"TRACE-TS's core contribution is that observations in generated DAG traces are grounded in expert-classifier attribution, not just fluent text. The only quantitative check of this is SNM, but SNM scores a predicted trace against the same teacher-generated reference traces used as supervision targets in Eq. 3. A student can therefore achieve high SNM by imitating the teacher's vocabulary and structure, and SNM cannot detect cases where both teacher and student fabricate an observation because the reference contains the same fabrication. The perturbation study in Table 8 shows SNM is sensitive to rewired channels and based_on edges, but that establishes metric sensitivity, not that student traces are signal-faithful. The teacher prompt itself provides the predicted activity label (Figure 8), so teacher traces may rationalize a label rather than derive from the signal. Human evaluation (Section 4.3, Table 4) samples only one window per activity class per dataset and reports no inter-annotator agreement; it is too small and selective to certify grounding across seven benchmarks. Section 6 openly states that trace quality is bounded by teacher and attribution pipeline, and SNM is not absolute verification. Thus, the key claim that reasoning is attribution-grounded is currently underdetermined by the evaluation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TRACE-TS, a two-stage framework for joint activity classification and structured DAG reasoning over wearable sensor windows. In Stage 1, an expert HAR classifier's IG+SHAP attributions are serialized and used to prompt a teacher LLM to produce DAG traces (observation/inference/synthesis/activity nodes with based_on provenance edges). In Stage 2, a frozen student LM is adapted via a projector and gated cross-attention to generate these traces from raw sensor inputs; at inference only the student runs, without attribution or teacher guidance. The paper also introduces SNM, an LLM-as-judge metric comparing predicted and reference traces at observation, inference, and synthesis levels. Experiments on seven HAR benchmarks report the best average accuracy and macro-F1 among evaluated methods (84.43%/81.24%), with extensive ablations over attribution source, reasoning format, teacher quality, sensor encoder, token count, adapter rank, and SNM perturbation behavior.","tokens_in":39251,"tokens_out":5271,"duration_ms":64256,"significance":"If the grounding claim were verified, this would be a substantial contribution: it addresses a real limitation of existing LLM-based explanations for time-series data, provides traceable structured reasoning with explicit provenance, and SNM is a sensible diagnostic idea. The paper's strengths include a freely released implementation, coverage of seven benchmarks, controlled ablations, a four-judge robustness analysis for SNM, and a perturbation study showing that SNM is sensitive to certain kinds of misgrounding. However, the central claim that the generated reasoning is \"attribution-grounded\" is currently underdetermined: the supervision and the evaluation share the same teacher-generated references, the teacher prompt supplies the predicted activity label, the human grounding audit is very small, and classification results are single runs. These issues are partly acknowledged in Section 6 and Appendix D.1, but they need to be addressed experimentally before the paper's main claim can be accepted.","major_comments":[{"comment":"The supervision pipeline and the SNM evaluation use the same teacher references. Figure 8 shows that the teacher prompt includes 'Predicted Activity: <ACTIVITY_LABEL>' and asks the model to generate a trace 'for <ACTIVITY_LABEL>'; Eq. (3) trains the student to imitate these label-conditioned traces; and SNM (Sec. 4.1) scores predicted traces against the same teacher-generated references. A student can therefore achieve high SNM by imitating teacher vocabulary and structure, and SNM cannot detect cases where both teacher and student fabricate an observation. Table 8's 'Misgrounded' perturbation demonstrates SNM sensitivity at the inference and synthesis levels (SNM-IF1 −24.35, SNM-SA −72.01), but SNM-OF1 actually increases (+1.18), so observation-level grounding is not independently verified. The paper acknowledges this in Section 6 ('not an absolute verification of reasoning faithfulness","section":"§3.1, Fig. 8, Eq. (3), §4.1, §6"},{"comment":"All classification results are from a single training run per backbone×dataset combination (Appendix D.1). The central recognition claim rests on average accuracy/F1, but several per-dataset gaps are small (e.g., UCI-HAR 96.67 vs. 97.50 for Attend & Discriminate; USC-HAD 72.35 vs. 73.09 for MantisV2; MHEALTH 98.17 vs. 98.10 for SensorLLM), and OPPORTUNITY is substantially below the specialist best (77.13 vs. 84.33). Without multiple seeds, error bars, or paired significance tests, the claim of 'best average' performance may be within run-to-run variation. Please report at least three seeds with mean±std, or paired tests on the comparisons that matter for the headline claim. The current SNM std values in Table 3 are across judges, not training runs, and therefore do not address this.","section":"Appendix D.1, Table 1"},{"comment":"The human grounding evaluation samples one window per activity class per dataset, with five annotators and no inter-annotator agreement reported. It rates only TRACE-TS traces and does not include baseline traces for comparison. This is too small and too selective to certify grounding across seven benchmarks. The blind pairwise comparison in Table 19 (attribution-grounded vs. no-attribution, 250 judgments) is valuable, but it measures preference between supervision conditions, not whether the cited observations correspond to actual signal events. Please expand the human audit (or a signal-statistics audit) and report IAA; otherwise, claims about 'grounded reasoning' should be softened to expert preference for attribution-conditioned traces.","section":"§4.3, Table 4"},{"comment":"SNM is a new metric and the paper's main diagnostic instrument, but its validity is established only indirectly: four judges give stable rankings and the perturbation study shows sensitivity to selected rewiring. No correlation with human judgments of trace quality is reported, and the perturbation study does not cover systematic teacher hallucinations (e.g., fluent observations that cite the correct channel but describe a pattern absent from the signal). Given that SNM is used both as a headline contribution and as evidence for the attribution-grounded claim, a direct validation against human annotations or against an independent signal-level audit is needed.","section":"§4.1, Appendix H"}],"minor_comments":[{"comment":"The reproduced prompt has a numbering error: the rule list skips index '11'. The caption notes this, but the authors should either renumber or fix the source prompt in the released code.","section":"Figure 8"},{"comment":"The caption 'Early (ℓ ≤ 1)' is confusing; it presumably means early layers rather than only layer 1. Please clarify the layer grouping.","section":"Figure 5(a)"},{"comment":"The text says that among faithful attribution sources, SNM differences fall within run-to-run variation, but the paper's reporting convention is single-run. Either add multi-run evidence or rephrase this as exploratory observation.","section":"§4.4, Table 5"},{"comment":"The threshold and serialization choices (e.g., 'global p90', 'top-k regions', temporal vocabulary) are described as hyperparameters but their sensitivity is not systematically studied. At minimum, state whether the reported results are robust to reasonable changes in these choices.","section":"Appendix D.2 / Figure 8"},{"comment":"The human evaluation aggregates ratings over a very small number of samples; the per-dataset standard deviations are large. It would help to state the total number of rated traces explicitly in the main text and to report the sample size per dataset.","section":"§4.3 / Table 4"}],"recommendation":"major_revision","confidential_remarks":"The circularity concern is not a minor caveat: the paper's core claim is that reasoning is grounded in signal-level attribution, yet the only quantitative grounding measure compares student output to the same teacher-generated references used for training, and the teacher prompt contains the predicted label. I would make the requested independent grounding audit and the multi-seed recognition results conditions for acceptance. The paper is otherwise well executed, with strong ablations and a useful diagnostic metric, but the evidence base for the headline claim needs to be broadened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about TRACE-TS. First, it is a real contribution to sensor-language reasoning: attribution from an expert classifier is used to supervise structured DAG traces with explicit provenance, a compact student LM learns to generate these traces from raw sensor input via gated cross-attention, and the SNM metric actually localizes reasoning errors better than BERTScore or METEOR. Second, the central \"signal-grounded\" claim is not independently verified. SNM compares student traces to the same teacher-generated references used as training targets, so high SNM can partly reflect imitation fidelity. The paper is honest about this in Section 6, but the limitation is real and not fully addressed by the perturbation study, which shows metric sensitivity, not external grounding.\n\nWhat is genuinely good: the framework is modular, the ablations are thorough (attribution source, reasoning format, encoder, teacher, adapter size), the blind pairwise human evaluation between attribution-grounded and no-attribution traces is a nice piece of evidence (82.4% raw preference, 91.2% decisive), and the code is promised. The SNM diagnostic clearly separates correct from incorrect traces at inference and synthesis levels, which is more than standard NLG metrics do. The qualitative error analysis is also unusually careful.\n\nThe soft spots, in proportion: (1) the self-referential evaluation is the big one, as above; (2) the teacher prompt includes the predicted activity label, so teacher traces may rationalize a label rather than derive from signal; (3) all classification results are single runs, with no error bars or significance tests; (4) the protocol is modified (null classes dropped, CAPTURE-24 subset), disclosed but making across-paper comparisons risky; (5) the human grounding evaluation samples only one window per activity class per dataset, no inter-annotator agreement reported. None of these are fatal, and the authors flag most of them, but together they mean the headline claims should be treated as conditional.\n\nWho this is for: anyone working on LLMs for wearable time series, explainable HAR, or structured reasoning evaluation. It deserves a serious referee. I would send it out rather than desk-reject, with reviewers asked to push for multi-seed results and an independent grounding audit (e.g., human verification on a larger sample, or SNM tested against traces generated from mislabeled teacher evidence).","headline":"A serious, well-ablated sensor-language reasoning framework whose one load-bearing weakness is that reasoning fidelity is measured against the very teacher traces used for supervision.","tokens_in":715,"tokens_out":799,"would_cite":true,"duration_ms":25248,"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":"By grounding language-model reasoning in the sensor signal, TRACE-TS jointly predicts activity labels and explains them, setting the best average accuracy/F1 across seven wearable benchmarks.","keywords":["sensor-language reasoning","human activity recognition","wearable time series","attribution grounding","DAG reasoning traces","cross-attention adapters","LLM-as-judge","explainable AI"],"falsifier":"Provide a teacher with deliberately misleading attribution maps (e.g., top regions from a different activity class) and show that the student still produces traces humans rate as well grounded; or show that SNM gives high scores to traces whose cited channels are silent while the text is plausible. The paper's own perturbation study suggests SNM would catch the latter, so the stronger test is whether human raters can distinguish attribution-grounded from misgrounded traces at scale, beyond the single-window sample.","tokens_in":38859,"feed_emoji":"📊","tokens_out":3581,"duration_ms":40433,"temperature":0.7,"pith_summary":"TRACE-TS tries to establish that wearable-activity reasoning can be both accurate and verifiable: it uses attribution maps from an expert classifier to pin reasoning traces to actual sensor regions, distills those structured DAG traces into a compact language model, and shows this grounded supervision improves both classification and trace quality. The framework trains a frozen LLM with gated cross-attention over sensor memory tokens, so at inference a single pass yields the activity label and a provenance-linked explanation without any teacher model or attribution computation. It also introduces Semantic Node Match (SNM), an LLM-judge metric that separates sound reasoning from hallucinated observations better than standard NLG metrics. The key claim is that grounding and structure are not costs but levers: they raise accuracy and macro-F1 above all evaluated baselines.","feed_headline":"Grounded reasoning lifts wearable activity accuracy to 84.4%","feed_subtitle":"A compact model predicts the activity and explains it with DAG traces; 17.96% F1 ahead of the best LLM-based baseline.","key_machinery":"The load-bearing mechanism is the attribution-grounded DAG trace. Integrated Gradients and SHAP are fused by geometric mean into a per-timestep, per-channel salience map; the top-k spatio-temporal regions are serialized into a closed evidence vocabulary; the teacher LLM converts these into a four-level DAG (Observation, Inference, Synthesis, Activity) with based_on provenance edges. This trace serves both as supervision for a compact student LM via gated cross-attention adapters over 8 sensor memory tokens, and as the reference for the proposed SNM metric, which matches nodes by Hungarian assignment and LLM-judged semantic equivalence.","core_discovery":"The paper's central claim is that reasoning traces for wearable time series can be made trustworthy by construction: attribute first, reason over the attributed regions, and distill the result into a small model. TRACE-TS shows that a teacher LLM, prompted only with serialized IG+SHAP salient regions, produces DAG reasoning traces whose observation nodes cite specific channels and temporal intervals, and that a frozen student LLM with gated cross-attention adapters can learn to reproduce these traces from raw sensor windows alone. The DAG structure—Observation → Inference → Synthesis → Activity with based_on provenance edges—turns out to be supervisory sugar: removing it and using free-form","pith_inferences":["The same attribution→trace→distill pipeline could transfer to other multivariate time-series domains (clinical monitoring, industrial sensors), where the closed-evidence format would need dataset-specific channel vocabularies.","SNM could become a general diagnostic for any structured reasoning format, not just sensor traces, if node types and provenance edges are defined.","The ablation pattern suggests that reasoning structure acts as a regularizer: by forcing the model to ground each claim, it prevents the LLM from falling back on language priors—a mechanism that could be tested by probing hidden states during reasoning generation.","A testable extension: verify whether the student's attention to sensor tokens aligns with the expert classifier's attribution maps on unseen windows—the paper shows alignment qualitatively, but a quantitative correlation would strengthen the grounding claim."],"forward_implications":["Wearable models can output both a label and a provenance-linked explanation in a single inference pass, with no attribution or teacher compute.","Grounded reasoning supervision is not a drag on classification: removing the DAG structure drops average accuracy by over 20 points.","Attribution quality directly controls reasoning quality: random or missing attribution collapses observation grounding (SNM-OF1 from 21.49 to 0).","Structure-aware metrics like SNM can localize hallucinations to observation, inference, or synthesis level, something surface metrics cannot.","Stronger teachers give better traces and better student accuracy, so trace quality scales with teacher reasoning ability."],"fun_headline_variants":["Wearable AI: traceable reasoning hits 84.4% accuracy","84.4% accuracy, verifiable explanations for wearable data","Grounded DAG traces beat LLM baselines by 17.96% F1 on wearables","Attribution-grounded reasoning makes sensor explanations verifiable"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reference traces and the SNM judge both come from the same teacher LLM, so if the teacher produces fluent but unsupported reasoning (or the attribution map misses the truly relevant channels), the student inherits that error and SNM cannot catch it—the human grounding check covers only one window per activity class (Section 6).","fun_headline_variants_meta":{"raw":{"variants":["Wearable AI: traceable reasoning hits 84.4% accuracy","84.4% accuracy, verifiable explanations for wearable data","Grounded DAG traces beat LLM baselines by 17.96% F1 on wearables","Attribution-grounded reasoning makes sensor explanations verifiable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001803,"raw_usage":{"total_tokens":6958,"prompt_tokens":788,"completion_tokens":6170,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":6089}},"tokens_in":532,"tokens_out":6170,"duration_ms":47215,"temperature":1.0,"reasoning_tokens":6089,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T01:00:31.792219+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Provide a teacher with deliberately misleading attribution maps (e.g., top regions from a different activity class) and show that the student still produces traces humans rate as well grounded; or show that SNM gives high scores to traces whose cited channels are silent while the text is plausible. The paper's own perturbation study suggests SNM would catch the latter, so the stronger test is whether human raters can distinguish attribution-grounded from misgrounded traces at scale, beyond the single-window sample.","supporting_citations":[],"review_version":1}