acceptodds
Under review as a conference paper at ICLR 2027

WaveAnchor-Agent: From Waveform Detection to Interactive EEG Analysis

Abstract

Fine-grained electroencephalography (EEG) interpretation requires localizing waveform events and explaining their spatial and temporal relationships. We introduce WaveAnchor-Agent (WA-Agent), an evidence-grounded workflow for fine-grained, generative EEG interpretation, together with a rubric-based benchmark for evaluating generated descriptions and dialogue. We build a physician-annotated dataset covering nine waveform classes and train a multi-scale temporal detector to localize events by channel and time. A planner-responder architecture connects waveform detection, deterministic measurement, and visualization tools to natural-language responses, supporting interactive analysis grounded in explicit signal-level evidence. Our benchmark covers waveform detection, selection, description, multi-turn conversation, and visualization. Its description and conversation tasks use 1,550 manually reviewed atomic rubric criteria to assess factual content beyond lexical similarity. On our benchmark, the detector achieves 45.22% test mAP at temporal IoU 0.5. WA-Agent achieves 93.71% selection accuracy and LLM-judged partial-credit rubric scores of 68.73% and 55.52% for description and conversation, respectively. Together, the workflow and benchmark provide a foundation for generating and evaluating fine-grained EEG explanations grounded in localized waveform evidence.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, Waveform detection, Large Language Models, LLM agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ciugOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper builds a system that answers questions about 10-s clinical EEG segments. WA-Detector (5.14M parameters, self-supervised pretraining on about 1,170 h) finds nine waveform classes per channel and time and gets 45.22% mAP at tIoU 0.5 on 175 test windows from 26 patient-disjoint recordings (Table 1). In WA-Agent, a planner calls typed tools (metadata, background spectra, detection, sharp-transient measurement, plotting) and a responder answers from their structured outputs. WA-Bench tests selection, description and multi-turn conversation on the same 175 windows, using 1,550 AI-drafted, human-reviewed atomic rubric criteria that an LLM judge scores, plus a 30-case visualization task. WA-Agent reaches 93.71% on selection and 68.73% / 55.52% partial-credit rubric scores, well above image-only LLM baselines (Table 4).

Critical strengths
  1. The benchmark keeps three kinds of correctness apart. Detection mAP, segment-level rubric coverage and tool routing get separate scores (Appendix A.8), and the rubric criteria are atomic, with strict and partial credit. The Oracle rows of Table 4 (detections replaced by annotations) are a good check: partial-credit description goes from 68.73% to 94.84%. So the detector is the bottleneck, not the language model.

  2. The detector is studied closely and reported honestly. Tables 15-16 ablate pretraining stages, cross-channel modules, pyramid levels and augmentations. Figure 9 and Section 6.1.1 say plainly that every class is recall-limited, with about 60% of abnormal slow and fast events missed. If released, the annotated nine-class, channel-level dataset (28,634 training and 7,344 test events) would help other groups.

Critical weaknesses
  1. Evaluation favours the agent and uses its own model as judge. The language tasks favour the agent by design. References, selection options and rubric criteria were drafted by an LLM from the structured annotation evidence (Appendix A.6, C.1) and checked by one medical trainee (L1009). WA-Agent's responder gets the same kind of structured class/channel/interval table from its tools. The image baselines have to read channel names and intervals off a plot (L407-409). The judge is DeepSeek-V4-Flash (L371-372), the same model the agent uses for both roles (L364). The sensitivity check swaps in DeepSeek-V4-Pro (Table 18), from the same family. Two results suggest the format matters. On selection, image-only GPT-5.5 already matches WA-Agent at 93.71% (Table 4). On CerebraGloss-Bench, which the authors did not build, WA-Agent is behind CerebraGloss on selection (70.00% vs 80.00%) and summary ROUGE-1. It is ahead only on the conversation score given by the same DeepSeek judge (4.27 vs 4.10, Table 34). The limitations paragraph admits that rubric scoring may miss unsupported statements (L480-481), and no expert rated the generated answers. I'd want a judge from another model family, an expert rating of a random sample of answers with judge-expert agreement, and a count of unsupported or hallucinated statements per answer. If possible, a second annotator should also write rubrics from the raw EEG, not from the structured evidence.

  2. Test-set early stopping, single run. Appendix E.1.2 says the ablations "use the same training and test splits and test [email protected] for early stopping" (L1239-1240), and the main text says the variants share the full model's early-stopping rule (L356). Sadly, I could not find a validation split anywhere. The configuration lists one random seed (Table 14). The 45.22% headline, the 10.66-point gain over the adapted CerebraGloss-YOLO (Table 2) and ablation differences of 0.9-10.6 points (Tables 15-16) all come from test-selected checkpoints on 175 windows from 26 recordings of one hospital, and none has an interval. The agent results use this same checkpoint. The CerebraGloss-Bench detector comparison puts a new checkpoint against a published number under that paper's protocol (Appendix I.1). I'd ask for a patient-level validation split for early stopping and re-reported test numbers, with several seeds and bootstrap intervals over recordings.

Smaller points that don't decide my recommendation: the annotations come from one EEG specialist, checked by a medical trainee (L128-129), with no inter-rater agreement. This affects how to read both mAP and the rubrics. Selection uses "synthetic silver" answer keys (L370). The language results are from one run per configuration, with no repeated sampling.

Questions for authors
  1. How do the description and conversation scores, and the ranking against the image baselines, change with a judge from a different model family? How well does the judge agree with an EEG expert on a sample of answers?
  2. Was any validation split used to pick the WA-Detector checkpoint? If not, what is test mAP when early stopping uses a held-out set of training patients?
  3. What are the seed-to-seed spread and recording-level bootstrap intervals for the mAP values in Tables 1-3?
  4. Did the hospital data have ethics-board approval or a consent waiver? Does it cover the planned public release? And if the baselines were queried through third-party model APIs, does it cover sending EEG plots to them?
1 comment
AuthorsOpenReview ↗

Response to Reviewer ciug

We thank the reviewer for the detailed feedback. We added cross-family rejudging, unsupported-statement counts, a text-only selection analysis, and validation-split, multi-seed, and confidence-interval analyses.

Response to W1 and Q1

Cross-family judge and expert check

GPT-5.6 Sol re-scored the frozen answers with the same rubrics. Partial-credit scores, DeepSeek/GPT (%):

SystemDescriptionConversation
WA-Agent, structured only60.14 / 66.3052.42 / 57.68
WA-Agent, summary priority68.73 / 72.3555.52 / 61.74
Gemini 3.8 Flash36.23 / 44.4630.40 / 38.75
GPT-5.531.59 / 35.2033.46 / 36.63
Claude Sonnet 533.83 / 41.5631.04 / 35.35
DeepSeek-V4.1-Flash28.13 / 32.1720.58 / 25.43
CerebraGloss25.84 / 30.4517.36 / 23.96

Although GPT scores are higher, rankings of the seven systems agree closely (Spearman: description 1.000, conversation 0.893), with summary-priority and structured-only WA-Agent first and second under both judges. Under GPT, summary-priority WA-Agent leads the strongest image baseline, Gemini, by 27.88/22.98 description/conversation points (95% recording CIs [21.22, 31.90] and [12.74, 27.83]). Item-level agreement is substantial (78.11% exact agreement; linear-weighted kappa 0.7236), and an EEG expert reviewing a random sample of the judge's scores found no concerns.

Unsupported statements per answer

DeepSeek Flash extracted 74,400 claims from 6,160 answers and checked them against reference annotations and metadata. Per-answer counts and the partial-credit score after weighted penalties for contradicted and unsupported event claims, description/conversation:

SystemSupported eventsContradictedPenalized score (%)
WA-Agent5.69 / 5.420.04 / 0.3441.09 / 35.81
GPT-5.50.75 / 1.560.73 / 0.189.05 / 23.61
Gemini 3.8 Flash0.65 / 1.870.54 / 0.4012.32 / 19.46

WA-Agent reports far more supported events per answer, the fewest description contradictions, and the highest penalized scores on both tasks.

Selection performance of GPT-5.5

One reason GPT-5.5 approaches WA-Agent on selection is that the items are fairly easy: questions are direct and options rarely contain traps. We asked GPT-6.1 Sol to answer from the stem and options only, without EEG, and classified its rationales by cue sufficiency:

Text-only outcomeCountRate (%)
Correct overall147/17584.00
Correct, cues judged sufficient28/17516.00
Correct, cues judged insufficient119/17568.00
Incorrect28/17516.00

Candidate options thus let partial cues rule out distractors, explaining the high selection scores of most image baselines in Table 4. Open-ended answers require reporting waveform type and location directly, which better separates systems; this is where WA-Agent opens a clear gap.

Response to W2, Q2, and Q3

The submitted checkpoint was selected by test mAP without a validation split. We repeated the original protocol over four seeds and reran selection on four random patient-held-out splits:

Protocol[email protected], mean +/- sample SD (%)
Original protocol, varying training seed43.01 +/- 1.65
Patient-held-out validation, varying split seed40.31 +/- 0.64

Each split holds out six of 29 report-linked groups from the 34 training recordings; checkpoints are selected on validation and scored on the unchanged test windows. Even so, the mean mAP (40.31%) exceeds the test-selected CerebraGloss-YOLO (34.56%) and LaBraM (36.15%), a comparison favoring the baselines.

We report recording-level 95% CIs from 10,000 paired bootstrap draws over 26 test recordings; point estimates follow Tables 1-3:

ModelmAP 95% CI (%)WA-Detector gain, 95% CI (points)
WA-Detector[31.12, 49.20]
CerebraGloss-YOLO[22.25, 36.51][3.13, 16.72]
w/o pretraining[23.58, 38.56][4.91, 13.98]
w/o additional TU training data[27.63, 43.55][-5.41, 13.88]
w/o channel self-attention[23.65, 39.06][3.41, 12.61]
w/o homologous-pair interaction[30.14, 44.19][-2.75, 8.21]
w/o top-down pyramid fusion[29.45, 45.19][-2.01, 7.47]
w/o local temporal attention[30.21, 45.05][-3.06, 7.20]

Every ablation lowers mAP; the gain over CerebraGloss-YOLO and the contributions of pretraining and channel self-attention exclude zero, while the remaining ablations have wider intervals given the small test set. The CerebraGloss-Bench comparison uses a published score, so no interval is available.

The specialist's waveform annotations were checked by a medical trainee, who reported no major discrepancies.

Response to Q4 and Ethics Concern

Ethics review is pending. Only image-based evaluations sent EEG plots, stripped of all identifying information; other API calls used structured evidence and text.

Reviewer_SL3pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

WaveAnchor-Agent: From Waveform Detection to Interactive EEG Analysis proposes an integrated agentic framework for EEG analysis, using both off-the-shelf components (DeepSeek LLMs) and new models (WA-Detector). However, the most interesting contribution is WA-Bench, though the current evaluation is potentially misaligned.

Critical strengths
  1. The paper is well-written and the research question is relevant, with notable activity in the field at the moment
  2. The effort in compiling a rubric using real-world experts is commendable, and the result is interesting
  3. The amount of ablations and results is impressive, especially in the appendices, and help assessing the robustness of the contributions (though some could be better targeted, as below)
  4. The rubric is sure to become a very useful dataset once open-sourced (hopefully?)
Critical weaknesses
  1. The novelty of agentic EEG frameworks with tool use is limited; the paper correctly points out previous literature (e.g., EEGAgent), but does not sufficiently expand on the differences with the current work
  2. Since WA-Bench is designed in tandem with WA-Agent, its tools are specifically meant to answer the rubric questions, and vice-versa; this potentially inflates the difference in the rubric scores (indeed, the selection scores are much closer with and without detector and background tool), and while the chosen criteria are certainly relevant, they are neither a necessary nor sufficient list in real-world clinical usage
  3. More comparisons are needed for WA-Detector, as its architecture is again not particularly novel, and it is compared only with one other model
Questions for authors
  1. Table 4 and 5 identify the detection interface as both a source of significant gain and a bottleneck, so evaluating larger models (Brant, LaBraM, MVPFormer, etc…), would partially clarify this aspect
  2. A second related ablation needs to compare whether the agentic framework or the detector are the source of the good performance, e.g., by giving the structured detector outputs to the LLM directly as a prompt without the agentic interface
2 comments
AuthorsOpenReview ↗

Response to Reviewer SL3p

We thank the reviewer for the helpful suggestions. We clarified the EEGAgent comparison, expanded detector comparisons, corrected the detector-removal ablation, and added direct prompting.

Response to W1

Related Work ("EEG with large language models and agents") now clarifies how our waveform detection and annotation-grounded QA evaluation differ from EEGAgent.

Response to W2

WA-Agent and WA-Bench share an EEG analysis framework (waveform class, location, timing, background, and artifacts), but the rubric derives from reviewed reference answers generated from physicians' annotations (Appendix C), not from WA-Agent's responses. The two therefore share a vocabulary but differ in coverage: the rubric assesses vertex waves separately, whereas the detector groups them under sharp waves, and summary-priority WA-Agent strictly passes only 1 of 68 vertex-wave criteria.

The ablations affect the tasks differently because their topic mixes differ: rubric criteria spread across non-epileptiform activity, sharp transients, sleep, artifacts, and recording setup, whereas 112 of the 175 selection items concern sharp transients or epileptiform discharges. Credited criteria (pass or partial) and correct selections by topic:

Rubric topicCriteriaWithout detection interfaceBackground removed
Sharp-transient candidates3930304
Non-epileptiform activity3260199
Background rhythms135294
Sleep waveforms2980187
Artifacts2890219
Recording setup1098864
Selection topicItemsWithout detection interfaceBackground removed
Sharp transients / epileptiform discharges11255107
Non-epileptiform activity / background341828
Artifacts / signal quality231123
Sleep waveforms626

Without the detection interface, all criteria except background rhythms and recording setup receive no credit, yet 55 of the 112 sharp-transient/discharge selections remain correct, mostly from stem cues or option elimination (36) or indirect inference from background measurements (15). Candidate options thus allow answering from partial evidence, whereas open-ended questions require reporting specific waveform findings. Removing background analysis mainly affects background-rhythm criteria (4 of 135 credited) and leaves selection at 164/175.

WA-Bench targets the waveform review clinicians perform when preparing an EEG report. We agree that clinical interpretation also draws on the full recording, patient history, event semiology, and other tests, which lie beyond this assessment.

Response to W3 and Q1

Beyond CerebraGloss-YOLO (Gu et al., 2026), we adapted LaBraM (Jiang et al., 2024), BRANT (Zhang et al., 2023), and MVPFormer-S (Carzaniga et al., 2026) to nine-class channel/time detection with a high-resolution stem, temporal pyramid, and decomposed-eventness head. These comparisons are now in the main-text detector table:

Detector[email protected] (%)
WA-Detector45.22
CerebraGloss-YOLO34.56
LaBraM-base (adapted)36.15
BRANT-large (adapted)32.16
MVPFormer-S (adapted)31.82

These adaptations do not improve mAP over WA-Detector. Component ablations further support the EEG-specific design: removing pretraining or channel self-attention lowers mAP by 9.46 and 9.24 points, with recording-level 95% CIs excluding zero ([4.91, 13.98] and [3.41, 12.61]). With structured-only evidence, Oracle annotations raise description/conversation scores from 60.14%/52.42% to 79.75%/83.10% (Table 4), leaving room for better detection.

We corrected the original ablation, which mistakenly retained detector-dependent measurements. The revised "without detection interface" condition in Table 5 scores 9.64%/0.15% on description/conversation (partial credit) versus the structured-only baseline above, supporting the importance of detector-derived evidence.

Response to Q2

We ran the suggested baseline, giving all precomputed structured outputs to the same LLM in one call without a planner, tool access, or postprocessing, and compared it with a contemporaneous WA-Agent rerun sharing the detector, LLM, prompts, and generation settings.

RunConfigurationSelectionDCTool calls
OriginalFixed routing93.1459.3854.483.000/request
OriginalWA-Agent93.7160.1452.421.637/request
NewDirect prompt94.2964.7858.864/segment, offline
NewWA-Agent93.1464.5757.651.657/request

D/C: description/conversation partial credit (%); compare only within run pairs.

The two new arms show no significant difference (paired 95% recording-level CIs include zero). The agentic framework instead retrieves evidence on demand: relative to fixed routing, adaptive routing reduces tool calls by 45.43% with small changes in QA scores.

The ablation performed in Q2 by the Authors confirms that the agent framework does not improve performance over the detector + LLM combination, significantly weakening the contribution. Moreover, while adaptive routing can reduce tool calls, they are usually much less expensive than LLM calls. To properly evaluate whether the agent makes the analysis computationally cheaper, a token count or latency measurement would be more appropriate. Also in Q2, it is interesting to see that D and C increase notably from the original to the new result, and reinforces Reviewer ciug's reservation about single runs.

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

The authors propose a system to detect events in EEG signals and identify when and where they occur, along with an agentic LLM framework to follow up on those findings and answer questions, and a dataset and benchmark to evaluate models on this task.

Critical strengths
  1. This tackles a useful problem. Scrubbing through EEG recordings to find significant events and especially to weed out artifacts is laborious. A framework that allows models to perform better on this task is a welcome contribution.
  2. Separating the analysis into an explicit detection step and question-answering step provides a layer of interpretability which is useful for clinical scenarios.
Critical weaknesses
  1. My main concern is the number of held-out recordings is quite limited, with 26 held-out recordings and 175 potentially correlated 10 second chunks. Given this, could the authors provide confidence intervals or statistical significance tests for the main detector and agent improvements using patient-level (or recording-level) resampling?
  2. How well would this system generalize to long, continuous EEG recordings containing much more background activity and fewer significant events? The evaluation uses waveform-enriched 10 second chunks so it is unclear whether the reported detection and downstream QA performance would stay the same under the distribution shift of more realistic recording conditions.
Questions for authors
  1. The authors acknowledge in the conclusion that the LLM rubric may not capture unsupported claims beyond the specified criteria. Why not then modify the rubric to penalize unsupported claims given that the authors have the labels for waveform classes, channels, and intervals?
  2. The planner-responder architecture reduces tool calls according to table 5, does it also improve performance over a single LLM receiving the same detector outputs and measurements, not using a separate planner? This information would help disentangle the contributions of structured evidence, tool selection, and the agent architecture.
2 comments
AuthorsOpenReview ↗

Response to Reviewer sazz

We thank the reviewer for the constructive feedback. We report recording-level confidence intervals, background examples, and an unsupported-claim audit, and clarify the planner's role in tool-use efficiency.

Response to W1

Paired 95% recording-bootstrap CIs use 10,000 draws, retaining all chunks per sampled recording. Gains are WA-Detector/Agent minus the comparator, in percentage points:

MetricComparatorGain95% CI
[email protected]CerebraGloss-YOLO+10.66[3.13, 16.72]
DescriptionGemini 3.8 Flash+32.50[25.83, 35.84]
DescriptionGPT-5.5+37.14[32.30, 40.70]
ConversationGemini 3.8 Flash+25.11[12.74, 30.76]
ConversationGPT-5.5+22.06[11.30, 27.97]

Agent scores use partial-credit rubrics and detected evidence with concise summary priority. Undefined class APs are excluded in detector draws with missing classes. All unadjusted paired CIs exclude zero, supporting the statistical robustness of the observed detector and QA gains under recording-level resampling that accounts for within-recording dependence.

Response to W2

We agree that deliberately enriching the 10-s evaluation segments for target waveforms changes their balance with background activity. Figure 11 in Appendix E.5 of the revised manuscript shows WA-Detector predictions on four TU EEG segments dominated by background activity. Manual review found no epileptiform discharges. At the original confidence threshold of 0.202, the four segments contained nine channel-level false-positive boxes, comprising five sharp-wave boxes, three abnormal slow-wave boxes, and one sleep-spindle box. The figure marks where these errors occur and illustrates what false positives look like on background activity. Assessing whether detection and downstream question answering maintain their performance on routine continuous EEG requires annotated continuous recordings, which are currently unavailable to us.

Response to Q1

DeepSeek Flash extracted 74,400 factual claims from 6,160 saved responses across nine conditions. Each event's class, location, and time form one claim. We penalize both contradicted event claims and event claims without matching annotations, excluding unmeasured background and clinical properties. Critical/minor claims incur penalties of 0.10/0.05, capped at 0.30 per response, with zero clipping and rubric-item-weighted aggregation within cases.

The table compares rubric scores before and after these penalties; each entry shows original / penalized partial-credit scores (%). WA-Agent uses detected evidence with concise summary priority.

ModelDescription (%)Conversation (%)
WA-Agent68.73 / 41.0955.52 / 35.81
GPT-5.531.59 / 9.0533.46 / 23.61
Gemini 3.8 Flash36.23 / 12.3230.40 / 19.46

Penalization lowers scores for all three systems, but WA-Agent remains highest on both tasks. Its paired penalty-adjusted gains over GPT-5.5 are 32.05 description points (95% recording CI [27.06, 34.68]) and 12.19 conversation points ([2.68, 17.46]), supporting its advantage under this penalty-based evaluation.

Response to Q2

The planner targets tool-use efficiency, not higher QA scores. Table 5's fixed-routing control invokes all analysis tools and passes their structured outputs to one responder without a planner, using the same detector, tools, responder backend, and response-processing rules as adaptive routing. Structured-only partial-credit description/conversation scores are 59.38%/54.48% with fixed routing versus 60.14%/52.42% with adaptive routing. Tool calls fall 45.43%, from 3.000 to 1.637 per request, with a modest conversation-score decrease. Fixed routing supplies all outputs while adaptive routing supplies selected outputs, so evidence is not identical. Thus, the planner's demonstrated benefit is fewer tool calls, with task-dependent changes in QA scores.

Thank you for the detailed response and the additional experiments.

The recording-level bootstrap CIs and the fixed routing comparison address some of my points.

However I still have some concerns:

For the rubric, WA-Agent loses 27.6 description points out of a maximum 30-point penalty, so most responses appear to hit the cap. The cap and zero-clipping limit how much any system can lose, so the comparison doesn’t fully reflect differences in unsupported claims.

Nine false-positive boxes in four background segments is roughly two per 10s window, which suggests a high false-positive rate on routine EEG. A false-positive rate per hour on held-out normal recordings would not require new annotations.

As such, I will be maintaining my current score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.