acceptodds
Withdrawn from ICLR 2027

iCluedo: Clue-Driven Agentic Investigation for Evidence-Grounded Traffic-Violation Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Traffic-violation analysis from dashcam video is a hypothesis-driven, sparse evidence reasoning problem. The decisive observations may be brief, temporally dispersed, and uncertain, yet a reliable judgment must establish who acted, where and when the behavior occurred, what happened, and which visual evidence supports the conclusion. We first introduce CVCV, a benchmark of 4.65K real-world dashcam videos covering 20 fine-grained behaviors across six California Vehicle Code (CVC) violation families. Each video is represented as a case file containing its violation labels, responsible road user, decisive timestamp, accident context, human-reviewed evidence, and grounding to the corresponding CVC section. We also introduce iCluedo, a clue driven agentic framework that addresses this problem through a progression from structured case memory, to competing hypotheses, to selective evidence acquisition, and finally map the violation to the law-grounded decision. iCluedo first organize perception tools output into a persistent, ego-centric spatio-temporal case memory that represents road users, traffic controls, events, and their evolving relationships. Segment-level hypothesis agent use this memory to generate competing violation hypotheses, which a violation investigation agent verifies through a selective READ–ASK–INSPECT–DECIDE loop. It retrieves relevant structured evidence, queries specialized scouts to resolve uncertainty, revisits the original video only when a decision-critical clue is missing, contradictory, or unreliable, and selects the interpretation best supported by the accumulated evidence. A visual witness provides observable facts from targeted video intervals, while explicit uncertainty determines whether additional evidence must be acquired. The resulting hypothesis is checked for consistency across the actor, location, time, behavior, and supporting evidence, and is then mapped deterministically to a traffic-law violation. iCluedo achieves 50.6% accuracy, outperforming four of five direct VLM baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Traffic Violation Reasoning, Scene Graphs, Vision-Language Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Zn8ZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a benchmark of approximately 4,650 real-world dashcam videos covering 20 fine-grained behaviors grouped into six California Vehicle Code violation families. Its case-file annotations include behavior labels, responsible actors, decisive timestamps, statute references and human-reviewed rationales/evidence.

It also proposes iCluedo, a training-free pipeline combining frozen perception tools, persistent spatio-temporal memory, competing hypotheses and a READ–ASK–INSPECT–DECIDE investigation loop. Additional evidence is acquired to resolve missing or unreliable clues, followed by consistency verification and deterministic mapping to a statute. On the reported evaluation subset, classification accuracy improves from 30.6% to 50.6%, while macro F1 improves from 0.382 to 0.423. Further experiments characterize actor/time grounding and top-two hypothesis recovery.

Critical strengths
  1. the formulation sounds useful, instead of just a label-only traffic-video analysis. The case-file representation not only asks what happened, but also who acted, where the decisive event occurred and what evidence supports the conclusion.
  2. the paper proposes a task-aligned investigation architecture with promising initial results.
Critical weaknesses
  1. evaluation coverage and alignment with the case-level claims are not very clear. Section 3 describes six violation families plus no-violation as closed-set categories, but Table 2 reports only five families, with supports totaling 3,264 clips. The manuscript does not clearly reconcile these supports with the approximately 4,650-video dataset or account for Special Stops, no-violation, Unknown and invalid predictions. Section 5.1 evaluates only clips for which every method returns a valid prediction. This provides matched coverage within the selected subset, but exclusion counts, class composition and per-model validity rates are necessary to interpret the headline score. This is not evidence of deliberate selection bias; it is an unresolved protocol question. These gaps affect what the benchmark and reported gains actually establish.
  2. The claimed benefit of clue-directed investigation is not isolated, and some of the reproduction details are missing. The method combines structured perception, tracking, scene memory, multiple model calls and adaptive video inspection, while the main baselines use fixed-context raw-video prompting. This can demonstrate a system-level improvement, but it does not identify the contribution of clue-conditioned investigation rather than better inputs, different models, additional observations or greater inference expenditure.
Questions for authors
  1. how does iCluedo compare with a similar computation budget, similar architecture system with non-clue-directed inspection policy? It would help isolate the proposed investigation mechanisum
  2. Could you clarify the temporal-evaluation population and provide corresponding baseline actor/time results? What systematic evaluation supports the evidence-grounding claim beyond the qualitative examples?
3 comments
AuthorsOpenReview ↗

W1: Evaluation coverage, exclusions, and interpretation of the reported results

Reviewer’s concern: The five-family evaluation does not clearly account for the full dataset, omitted categories, or invalid predictions.

Response:

We would like to clarify that the CVCV dataset contains 4,636 videos, while the original evaluation reported results on 3,264 clips. The original table presents the overall performance and the results for each evaluated violation family. Videos labeled as unknown didn't use in the evaluation. It will be used to support future research on ambiguous and long-tail driving scenarios, which we also plan to investigate in future work.

We understand the reviewer's concern regarding evaluation coverage. To provide a broader evaluation, we conducted additional experiments covering the violation categories beyond the original family labels with no violation. Specifically, we evaluated the methods on a 421-video subset of CVCV, reporting MRR, Top-1 and Top-3 accuracy, precision, recall, and F1-score. The iFinder and iCluedo results in Table 1 correspond to their GLM-based configurations.

Table 1. Violation-family recognition performance (%). Best results in each column are highlighted in bold.

ModelMRR↑Top-1 Acc.P@1R@1F1@1Top-3 Acc.P@3R@3F1@3
Cosmos-Reason2-8B20.89.510.915.74.030.770.837.034.3
Qwen3-VL-8B37.57.829.316.37.563.185.964.965.5
iFinder [1]46.821.438.123.116.869.681.467.469.5
iCluedo (Ours)60.140.835.941.535.274.964.868.564.9

Compared with iFinder, iCluedo improves MRR by 13.3 percentage points, Top-1 accuracy by 19.4 points, and Top-3 accuracy by 5.3 points. F1@1 also improves from 16.8% to 35.2%, although F1@3 is lower (64.9% versus 69.5%). These results support the benefit of the complete iCluedo framework over a structured-perception baseline.

We additionally report the per-category performance across all seven CVCV categories, including No Violation, to provide a more detailed comparison between iCluedo, iFinder–GLM, and direct VLM baselines.

Table 2. Chapter-level F1@1 (%) comparison across seven CVCV categories. Best results in each row are highlighted in bold.

Violation CategoryCosmos-2BQwen3-VL-8BiFinderiCluedo
Traffic Signal/Sign39.138.96.540.0
Driving/Overtaking16.344.15.351.2
Right-of-Way11.830.013.344.4
Pedestrian0.024.018.237.8
Turning/Stopping0.00.022.225.8
Special Stops0.053.316.738.9
No Violation10.38.320.88.0
Macro-F1@111.128.414.735.2

iCluedo achieves the highest macro-F1 of 35.2%, outperforming iFinder–GLM by 20.5 percentage points, Qwen3-VL-8B by 6.8 points, and Cosmos-2B by 24.1 points. It achieves the best performance in five of the seven categories, with particularly notable improvements in Driving/Overtaking, Right-of-Way, and Pedestrian violations. However, Qwen3-VL-8B performs better on Special Stops, while iFinder–GLM achieves the highest F1@1 for No Violation. These results further support the benefit of iCluedo over the structured-perception baseline while highlighting the remaining challenges in certain violation categories. Reference

[1] Yao et al. iFinder: Structured Zero-Shot Vision-Based LLM Grounding for Dash-Cam Video Reasoning.

AuthorsOpenReview ↗

W2/Q1: Isolating clue-directed investigation and clarifying implementation details

Reviewer’s question: How does iCluedo compare with a similarly structured system using non-clue-directed inspection under a comparable computation budget?

Response:

We agree that comparisons with fixed-context, raw-video VLMs alone cannot isolate the contribution of clue-directed investigation. To strengthen the comparison, we added iFinder [1], a training-free structured-perception pipeline that does not employ the scene graph and clue-driven investigation used by iCluedo. The iFinder use perception toolchain. The perception models are pretrained and frozen, with no fine-tuning on CVCV.

We have already presented the comparison across the seven CVCV violation categories in our response to the first reviewer concern. We further expanded the evaluation beyond the original family-level setting by conducting additional experiments covering fine-grained recognition across 19 primary violation labels, reporting MRR, Top-1, Top-3, and Top-10 accuracy, precision, recall, and F1-score.

Table 1. Fine-grained primary violation recognition (%). Best results in each column are highlighted in bold.

ModelMRR↑Top-1 Acc.P@1R@1F1@1Top-3 Acc.P@3R@3F1@3Top-10 Acc.P@10R@10F1@10
Cosmos-Reason2-8B23.919.125.518.615.225.739.925.022.238.272.237.138.6
Qwen3-VL-8B37.219.925.919.315.547.855.548.344.280.186.180.779.4
iFinder39.423.815.913.110.945.247.336.937.281.092.881.082.3
iCluedo (Ours)42.225.750.426.026.153.769.654.052.882.487.182.481.6

At the primary-label level, iCluedo improves over iFinder from 23.8% to 25.7% Top-1 accuracy, from 45.2% to 53.7% Top-3 accuracy, and from 39.4% to 42.2% MRR. F1@1 also improves from 10.9% to 26.1%. These results extend the comparison to fine-grained behavior recognition and support the benefit of the complete iCluedo framework over the structured-perception baseline.

Implementation details. The clue-driven investigation operates through four actions: READ, ASK, INSPECT, and DECIDE. The implementation settings are summarized below.

Table 2. iCluedo implementation settings.

ComponentSetting
Reasoning modelsGLM-5.2, Qwen3.6-35B
Inspection VLMCosmos-Reason2-8B
Inspection sampling rate16 FPS
Maximum expert actions6
Maximum pixel inspections4
Expert votes2
Scout / uncertainty token limit10,000
Expert / revision / voting token limit20,000
Primary-audit token limit12,000

The table summarizes the reasoning models, visual inspection settings, and computational limits used in iCluedo. The reported iFinder and iCluedo results correspond to their GLM configurations.

Reference

[1] Yao et al. iFinder: Structured Zero-Shot Vision-Based LLM Grounding for Dash-Cam Video Reasoning.

AuthorsOpenReview ↗

Q2 / W1: Temporal-evaluation population, actor attribution, and evidence grounding

Reviewer’s question: Could you clarify the temporal-evaluation population and provide corresponding baseline actor/time results? What systematic evaluation supports evidence grounding beyond qualitative examples?

Response: We agree that the case-level claims require evaluation beyond behavior recognition. To address this concern, we conducted additional experiments comparing iCluedo with direct VLM baselines on temporal localization, actor attribution, reasoning quality, and evidence grounding.

Temporal Localization

We clarify that this evaluation focuses on violation timestamps rather than accident timestamps. We measure temporal localization performance as the percentage of predicted violation timestamps falling within two and three seconds of the human-annotated violation time.

Table 2. Temporal localization performance (%). Higher values indicate better performance. Bold values indicate the best result in each column, including ties.

ModelWithin 3s ↑Within 2s ↑
Cosmos-Reason2-2B44.426.0
Cosmos-Reason2-8B52.739.6
Qwen3-VL-8B33.724.3
iFinder11.28.9
iCluedo (Ours)53.839.6

iCluedo achieves the highest temporal localization performance within three seconds (53.8%) and ties Cosmos-Reason2-8B within two seconds (39.6%). Compared with iFinder, iCluedo improves localization within three seconds by 42.6 percentage points and within two seconds by 30.7 percentage points. These results demonstrate improved temporal localization, although precise identification of violation timestamps remains challenging.

Actor Attribution

The task distinguishes whether the responsible road user is the ego vehicle or another road user, rather than identifying an exact tracked vehicle among multiple surrounding actors. We additionally evaluated the models on this task, as shown in Table 3.

Table 3. Actor attribution performance (%). Higher values indicate better performance.

ModelAcc. ↑Other Car ↑Ego Car ↑
Qwen3-VL-8B8.46.24.5
Cosmos-Reason2-8B5.01.44.5
iCluedo (Ours)45.344.177.3

iCluedo achieves the highest overall actor attribution accuracy (45.3%), compared with 8.4% for Qwen3-VL-8B and 5.0% for Cosmos-Reason2-8B. It also achieves the highest performance for identifying other road users (44.1%), while iCluedo–Qwen achieves the highest ego-vehicle attribution score (86.4%). These results demonstrate substantial improvements over direct VLM baselines, although identifying the responsible actor remains challenging.

Reasoning and Evidence Grounding

We additionally evaluate the quality of the generated reasoning and supporting visual evidence.

Table 4. Reasoning and evidence-grounding performance (%). Higher values indicate better performance.

ModelReason ↑Evidence ↑
Cosmos-Reason2-8B39.650.3
Qwen3-VL-8B37.649.3
iCluedo (Ours)39.152.2

iCluedo achieves the highest evidence-grounding score (52.2%), compared with 50.3% for Cosmos-Reason2-8B and 49.3% for Qwen3-VL-8B. For reasoning quality, iCluedo–GLM achieves 39.1%, close to the highest score of 39.6% obtained by Cosmos-Reason2-8B. These results show that iCluedo improves evidence grounding while maintaining reasoning performance comparable to the strongest direct VLM baseline.

Reviewer_29irOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary
  1. Introduces CVCV, a dashcam benchmark linking traffic behaviors to actors, timestamps, evidence, and legal categories.
  2. Proposes structured scene memory and hypothesis-driven requests for additional evidence.
  3. Reports 50.6% classification accuracy, while actor attribution and precise timing remain difficult.
Critical strengths
  1. Useful task formulation. Linking behavior, actor, time, and evidence supports more complete video understanding.
  2. Promising investigation design. Structured memory and competing hypotheses suit cases with sparse or ambiguous evidence.
Critical weaknesses
  1. The central mechanism is not isolated. No ablations separate clue-driven investigation from the underlying perception stack.
  2. Implementation and comparisons are incomplete. The agent model, prompts, budgets, and checkpoints are insufficiently specified.
  3. Annotation reliability is unclear. Annotator qualifications, agreement, and adjudication procedures are not adequately documented.
  4. Evaluation coverage is incomplete. Only five families are reported, exclusions remain unclear, and evidence grounding lacks quantitative evaluation.
Questions for authors
  1. What happens when inspection, hypothesis comparison, or structured memory is removed under matched budgets?
  2. Which models, prompts, checkpoints, and action budgets were used?
  3. What annotation procedures and agreement statistics support the dataset labels?
  4. Can you report full-category results, exclusion counts, and baseline comparisons for actor, timing, and evidence grounding?
4 comments
AuthorsOpenReview ↗

W1/Q1: Isolating clue-driven investigation from the perception stack

Reviewer’s question: What happens when inspection, hypothesis comparison, or structured memory is removed under matched budgets?

Response: This is a good question. We agree that comparisons with direct VLMs alone are insufficient to establish the benefit of clue-driven investigation. To strengthen the comparison, we added iFinder [1], a training-free pipeline that uses structured perception with VLM without our clue-driven investigation. The iFinder use perception toolchain, providing a comparison beyond raw-video baselines. In this experiments, We evaluated the methods on balance subset of 421-video from CVCV Dataset, reporting MRR, Top-1 and Top-3 accuracy, precision, recall, and F1-score. The evaluation include all the family labels with no vioatlion The iFinder and iCluedo results below shown in Table 1. Table 1. Chapter-level violation recognition performance (%). Best results in each column are highlighted in bold.

ModelMRR↑Top-1 Acc.P@1R@1F1@1Top-3 Acc.P@3R@3F1@3
Cosmos-Reason2-8B20.89.510.915.74.030.770.837.034.3
Qwen3-VL-8B37.57.829.316.37.563.185.964.965.5
iFinder [1]46.821.438.123.116.869.681.467.469.5
iCluedo (Ours)60.140.835.941.535.274.964.868.564.9

Compared with iFinder, iCluedo improves MRR by 13.3 percentage points, Top-1 accuracy by 19.4 points, and Top-3 accuracy by 5.3 points. F1@1 also improves from 16.8% to 35.2%, although F1@3 is lower (64.9% vs. 69.5%). These results support the benefit of the complete iCluedo framework over a structured-perception baseline.

[1] Yao et al. iFinder: Structured Zero-Shot Vision-Based LLM Grounding for Dash-Cam Video Reasoning.

AuthorsOpenReview ↗

W2/Q2: Implementation details and completeness of the comparisons

Reviewer's question: Implementation and comparisons are incomplete. The agent model, prompts, budgets, and checkpoints are insufficiently specified.

Response As described in Appendix A.1 (Table 4), our pipeline separates pretrained perception from the clue-driven investigation process. The perception stack uses OWL-V2 for object detection, ByteTrack for persistent object identities, CenterTrack for object yaw estimation, Metric3D v2 and SAM for distance estimation, OMR for lane detection, GeoCalib for camera calibration, and DROID-SLAM for ego-motion estimation. InternVL provides scene and object attributes, while VideoLLaMA2 generates video-level descriptions. All perception models are frozen and are not fine-tuned on CVCV.

The investigation operates over the resulting case memory through four actions as mentioned, including READ, ASK, INSPECT, and DECIDE. A subsequent consistency check verifies the behavior, responsible actor, location, time, and supporting evidence before mapping the behavior to the CVC taxonomy.

Implementation settings:

ComponentSetting
GLM reasoning modelglm-5.2
Qwen reasoning modelqwen3.6-35b
Additional reasoning backends supported in codegpt-4.1, gpt-4o, gpt-5.4
Inspection VLMCosmos-Reason2-8B
Inspection sampling rate16 fps
Expert action budget6
Maximum pixel inspections4
Expert votes2
Scout / uncertainty token limit10,000
Expert / revision / vote token limit20,000
Primary-audit token limit12,000

The iFinder and iCluedo results reported in our response tables correspond to their GLM-based configurations.

AuthorsOpenReview ↗

W3/Q3: Annotation reliability and quality control

Reviewer’s question: What annotation procedures and agreement statistics support the dataset labels?

Response: As mentioned in Line 245, we would like to clarify the distinction between human annotation and VLM-assisted text generation. Human annotators provide the behavior, responsible actor, accident, and timestamp labels. A VLM generates initial drafts of the rationale and evidence descriptions, which are subsequently reviewed by human annotators. Therefore, the core event labels are human-annotated rather than automatically generated by the VLM. Examples of dataset annotations are provided in Appendix A.3 (Table 8).

AuthorsOpenReview ↗

W4/Q4: Evaluation coverage, exclusions, and grounding

Reviewer’s question: Can you report full-category results, exclusion counts, and baseline comparisons for actor, timing, and evidence grounding?

Response:

We expanded the evaluation beyond the original violation-family setting by conducting additional experiments covering fine-grained recognition across all 19 primary violation labels. We compare iCluedo against direct VLM baselines and iFinder, a structured-perception pipeline without clue-driven investigation.

Table 1. Fine-grained primary violation recognition (%). Best results in each column are highlighted in bold.

ModelMRR↑Top-1 Acc.P@1R@1F1@1Top-3 Acc.P@3R@3F1@3Top-10 Acc.P@10R@10F1@10
Cosmos-Reason2-8B23.919.125.518.615.225.739.925.022.238.272.237.138.6
Qwen3-VL-8B37.219.925.919.315.547.855.548.344.280.186.180.779.4
iFinder39.423.815.913.110.945.247.336.937.281.092.881.082.3
iCluedo (Ours)42.225.750.426.026.153.769.654.052.882.487.182.481.6

At the primary-label level, iCluedo with GLM 5.3 model improves over iFinden from 23.8% to 25.7% Top-1 accuracy and from 45.2% to 53.7% Top-3 accuracy. MRR also increases from 39.4% to 42.2%. These results extend our evaluation to fine-grained behavior recognition, while showing that identifying the correct first-choice violation label remains challenging.

Temporal localization. We additionally evaluate violation-time localization by measuring the percentage of predictions falling within two and three seconds of the annotated timestamp.

Table 2. Temporal localization performance (%). Higher values indicate better performance.

ModelWithin 3s ↑Within 2s ↑
Cosmos-Reason2-2B44.426.0
Cosmos-Reason2-8B52.739.6
Qwen3-VL-8B33.724.3
iFinder11.28.9
iCluedo (Ours)53.839.6

iCluedo achieves the highest performance within three seconds (53.8%) and ties Cosmos-Reason2-8B for the highest performance within two seconds (39.6%). Compared with iFinder, iCluedo improves localization within three seconds from 11.2% to 51.5% and within two seconds from 8.9% to 37.9%.

Actor attribution. The task distinguishes whether the responsible road user is the ego vehicle or another road user, rather than identifying an exact tracked vehicle among multiple surrounding actors. We additionally evaluated the models on this task, as shown in Table 3.

Table 3. Actor attribution performance (%). Higher values indicate better performance.

ModelAcc. ↑Other Car ↑Ego Car ↑
Qwen3-VL-8B8.46.24.5
Cosmos-Reason2-8B5.01.44.5
iCluedo - Qwen36.331.786.4
iCluedo - GLM(Ours)45.344.177.3

iCluedo achieves the highest overall actor attribution accuracy (45.3%), compared with 8.4% for Qwen3-VL-8B and 5.0% for Cosmos-Reason2-8B. It also achieves the highest reported performance for identifying other road users (44.1%), while iCluedo performs best for ego-vehicle attribution (86.4%). These results demonstrate substantial improvements over direct VLM baselines, although identifying the responsible actor remains challenging.

Reasoning and evidence grounding. We additionally evaluate the quality of the generated reasoning and supporting visual evidence.

ModelReason ↑Evidence ↑
Cosmos-Reason2-8B39.650.3
Qwen3-VL-8B37.649.3
iCluedo (Ours)39.152.2

iCluedo achieves the highest evidence-grounding score (52.2%), compared with 50.3% for Cosmos-Reason2-8B and 49.3% for Qwen3-VL-8B. For reasoning quality, iCluedo achieves 39.1%, close to the highest score of 39.6% obtained by Cosmos-Reason2-8B. These results show that iCluedo improves evidence grounding while maintaining reasoning performance comparable to the strongest direct VLM baseline.

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

The paper treats traffic-violation reasoning from dashcam video as a hypothesis-driven problem over sparse evidence rather than a single-pass classification. It contributes a benchmark and a method. CVCV holds 4,650 real-world clips from Nexar, RoadSocial and CCD, labelled with 20 fine-grained behaviours that map to specific California Vehicle Code sections across six chapter-level families, plus no-violation and unknown. Each clip carries a case file rather than a label, naming the violated section, the responsible road user, the decisive time, the accident context and a human-reviewed rationale with its visual evidence.

iCluedo is the method. A frozen perception toolchain (OWL-V2, ByteTrack, Metric3D v2, OMR, DROID-SLAM, InternVL and VideoLLaMA2 among them) builds an ego-centric spatio-temporal scene graph with deterministic relation rules and per-frame evidence and uncertainty scores. Segment-level scouts propose competing hypotheses from that graph, and a violation expert resolves them through a bounded READ-ASK-INSPECT-DECIDE loop that returns to raw frames only when a clue is missing, contradictory or unreliable. The decision is checked for consistency across actor, location, time, behaviour and evidence, then mapped to a CVC section.

On the matched-coverage subset, iCluedo reports 50.6% video-weighted accuracy and 0.423 macro F1 against 30.6% and 0.382 for the strongest of four direct VLM baselines, Cosmos3-Nano. Actor attribution reaches 55.0% balanced accuracy and the predicted decisive time lands within 5 s on 63.0% of clips, within 1 s on 20.9%. Counting the second-ranked hypothesis raises accuracy to 64.6%.

Critical strengths
  • CVCV asks for something no prior benchmark in the related-work table asks for. DriveCombo runs in CARLA simulation, DecaTARA and CAViAR tie their labels to crashes, and C-TRAIL covers roughly 1K clips under Chinese law. None requires a model to name the California Vehicle Code section, attribute the violation to the right road user, localize the decisive moment and cite the evidence in one case file (Eq. 1). Framing the task as case reconstruction rather than label prediction matches how insurers and collision reports actually work, and that framing is the paper's durable contribution.
  • The reporting on the method's own failures is honest. Per-class F1 in Table 2 shows iCluedo ahead on overtaking (0.634 against 0.401) and turning (0.725 against 0.532) and behind on signal (0.236 against 0.338) and pedestrian (0.218 against 0.394), and the authors attribute those two losses to upstream perception failures on small or briefly visible entities instead of smoothing them over. The top-2 diagnostic in Section 5.2 is the same kind of reporting. Accuracy rises from 0.506 to 0.646, which locates 28% of the top-1 errors in hypothesis ranking rather than hypothesis generation, and points the next piece of work at the right stage.
Critical weaknesses
  • The 20.0-point gain cannot be attributed to the mechanism the paper claims for it. iCluedo receives a structured scene graph from seven or more frozen perception models; the four baselines receive raw video and nothing else. No variant isolates the parts, so feeding the same perception records to a single VLM call, running the loop without the clue graph, or dropping the hypothesis scouts are all untested, in Section 5 and in the appendix. The agentic video systems cited in Section 2, VideoAgent and DriveAgent-R1, are never run. On top of that, the matched-coverage subset keeps only clips on which every method returns a valid prediction, and the paper does not say how many of the 4,650 clips survive. iCluedo fails by producing no hypothesis when perception misses the decisive entity, while a VLM always emits a label, so the filter can remove exactly the clips where the method is weakest.
  • One of the three evaluation levels the paper defines is never measured. Eq. 1 and Section 1 promise recognition, event grounding and evidence grounding, and Section 5 reports the first two. No table scores whether the rationale is supported by the evidence it cites, which is the level that distinguishes case reconstruction from classification and the reason the benchmark is interesting. Actor attribution and temporal grounding have the same shape of gap: Table 3 reports them for iCluedo alone, so the 55.0% and the 63.0% within 5 s stand without a comparison.
Questions for authors

How large is the matched-coverage subset, and what does the full-dataset number look like? Section 5.1 defines the subset as the clips on which every method returns a valid prediction, and gives neither its size nor the per-class support after filtering. Scoring a missing prediction as incorrect over all 4,650 clips would answer the obvious worry, which is that the filter removes the clips where the perception pipeline failed and iCluedo would have lost.

What are the policy prompts and the budgets? The investigation loop is governed by a policy that Appendix A.1 calls prompted rather than learned, bounded by an action budget B and an INSPECT sub-budget B_pix, and neither the prompts nor the two values appear in the paper. The agentic contribution is precisely how the expert chooses to READ, ASK, INSPECT or DECIDE, so without them the method cannot be reproduced. An ablation over B would also show whether the result depends on a generous budget.

How much do the numbers move across seeds? The scouts and the expert are sampled LLM calls, the per-class supports are small (104 for pedestrian, 221 for signal), and Table 2 reports one run with no interval. Three seeds with standard deviations, on the per-class F1 in particular, would show whether the 20.0-point gap and the per-class ordering are stable.

How are severity, reliability and the down-weighting factor in Eq. 7 computed? All three are described as deterministic functions of the perception records, and no formula, threshold or worked example is given for any of them. Table 5 lists the relation rules but not how a relation becomes a number in [0,1], so the clue graph cannot be rebuilt from the paper.

Two smaller points. The abstract says iCluedo outperforms four of five direct VLM baselines while Table 2 lists four and iCluedo leads all four, so one of the two is wrong. And what does the pipeline cost? Ten frozen models, several LLM calls per step and optional re-inspection are reported with no wall-clock time, GPU-hours or per-clip figure.

2 comments
AuthorsOpenReview ↗

W1: Isolating the contribution of clue-driven investigation

Reviewer’s concern: The original comparison does not separate the investigation mechanism from structured perception, additional model calls, and adaptive inspection.

Response: We agree that the original 20.0-percentage-point improvement demonstrates a system-level gain rather than isolating the contribution of clue-driven investigation. To strengthen the comparison, we added iFinder [1], a training-free pipeline that uses structured perception without our clue-driven investigation. iFinder use perception toolchain, whose models are pretrained and frozen, with no fine-tuning on CVCV. This provides a stronger comparison beyond raw-video VLM baselines.

We conducted additional experiments across seven CVCV categories, including No Violation. The iFinder and iCluedo results correspond to their GLM configurations.

Table 1. Seven-category violation recognition performance (%). Best results in each column are highlighted in bold.

ModelMRR↑Top-1 Acc.P@1R@1F1@1Top-3 Acc.P@3R@3F1@3
Cosmos-Reason2-8B20.89.510.915.74.030.770.837.034.3
Qwen3-VL-8B37.57.829.316.37.563.185.964.965.5
iFinder [1]46.821.438.123.116.869.681.467.469.5
iCluedo (Ours)60.140.835.941.535.274.964.868.564.9

Compared with iFinder, iCluedo improves MRR by 13.3 percentage points, Top-1 accuracy by 19.4 points, and Top-3 accuracy by 5.3 points. F1@1 also improves from 16.8% to 35.2%, although F1@3 is lower (64.9% versus 69.5%). These results support the benefit of the complete iCluedo framework over a structured-perception baseline.

Reference

[1] Yao et al. iFinder: Structured Zero-Shot Vision-Based LLM Grounding for Dash-Cam Video Reasoning.

AuthorsOpenReview ↗

W2: Evidence grounding and baseline actor/time comparisons

We agree that recognition accuracy alone does not establish the quality of a reconstructed case. Our additional experiments include baseline comparisons for violation timing, actor attribution, and the reported Reason and Evidence scores.

Temporal Localization

We clarify that this evaluation focuses on violation timestamps rather than accident timestamps. We measure temporal localization performance as the percentage of predicted violation timestamps falling within two and three seconds of the human-annotated violation time.

Table 2. Temporal localization performance (%). Higher values indicate better performance. Bold values indicate the best result in each column, including ties.

ModelWithin 3s ↑Within 2s ↑
Cosmos-Reason2-2B44.426.0
Cosmos-Reason2-8B52.739.6
Qwen3-VL-8B33.724.3
iFinder11.28.9
iCluedo (Ours)53.839.6

iCluedo achieves the highest temporal localization performance within three seconds (53.8%) and ties Cosmos-Reason2-8B for the highest performance within two seconds (39.6%). Compared with iFinder, iCluedo improves localization within three seconds by 42.6 percentage points and within two seconds by 30.7 percentage points. These results demonstrate improved temporal localization, although precise identification of violation timestamps remains challenging.

Actor Attribution

The task distinguishes whether the responsible road user is the ego vehicle or another road user, rather than identifying an exact tracked vehicle among multiple surrounding actors. We additionally evaluated the models on this task, as shown in Table 3.

Table 3. Actor attribution performance (%). Higher values indicate better performance.

ModelAcc. ↑Other Car ↑Ego Car ↑
Qwen3-VL-8B8.46.24.5
Cosmos-Reason2-8B5.01.44.5
iCluedo–Qwen36.331.786.4
iCluedo–GLM (Ours)45.344.177.3

iCluedo–GLM achieves the highest overall actor attribution accuracy (45.3%), compared with 8.4% for Qwen3-VL-8B and 5.0% for Cosmos-Reason2-8B. It also achieves the highest performance for identifying other road users (44.1%), while iCluedo–Qwen achieves the highest ego-vehicle attribution score (86.4%). These results demonstrate substantial improvements over direct VLM baselines, although identifying the responsible actor remains challenging.

Reasoning and Evidence Grounding

We additionally evaluate the quality of the generated reasoning and supporting visual evidence.

Table 4. Reasoning and evidence-grounding performance (%). Higher values indicate better performance.

ModelReason ↑Evidence ↑
Cosmos-Reason2-8B39.650.3
Qwen3-VL-8B37.649.3
iCluedo (Ours)39.152.2

iCluedo achieves the highest evidence-grounding score (52.2%), compared with 50.3% for Cosmos-Reason2-8B and 49.3% for Qwen3-VL-8B. For reasoning quality, iCluedo–GLM achieves 39.1%, close to the highest score of 39.6% obtained by Cosmos-Reason2-8B. These results show that iCluedo improves evidence grounding while maintaining reasoning performance comparable to the strongest direct VLM baseline.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Q4: Computing severity, reliability, and the down-weighting factor

Reviewer’s question: How are the three quantities in Eq. 7 computed, and what formulas or thresholds make the clue graph reproducible?

Response:

We agree that describing these functions as deterministic is not sufficient for reproducibility. The three quantities in Eq. 7 serve different purposes:

QuantityMeaning
Severity, sk,ts_{k,t}Measures how strongly a clue indicates a potential traffic violation.
Reliability, ρk,t\rho_{k,t}Measures the trustworthiness of the perception observations supporting the clue.
Down-weighting factor, wk,tw_{k,t}Reduces the contribution of distant or roadway-irrelevant observations.

Severity and reliability are computed from the structured perception records before language-model reasoning. The resulting evidence and uncertainty summaries are not calibrated probabilities, but are used to guide the investigation toward potentially relevant and uncertain observations.

AuthorsOpenReview ↗

Summary of Rebuttal and Manuscript Revisions

We thank all reviewers for their careful reading and constructive feedback. We have updated the manuscript and provided point-by-point responses to the concerns raised.

The revision adds iFinder comparisons, expands recognition evaluation to seven closed-set categories and 19 primary labels, and includes additional comparisons for actor attribution, temporal localization, and reason/evidence scores.

Our responses also clarify the annotation workflow, model configurations, and investigation budgets. We distinguish the reported system-level improvements from conclusions requiring further controlled experiments and acknowledge the remaining limitations. We appreciate the reviewers’ suggestions, which helped us improve the paper’s clarity and evaluation.

Area_Chair_QJFVOpenReview ↗

Reminder: Please Review the Author Responses

Dear Reviewers,

As the discussion period is coming to an end, please take a moment to review the authors' responses and update your ratings if your assessment has changed. If you have any remaining concerns, please feel free to share them in the discussion.

Thank you for your efforts!

AC

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

Related papers

Discussion (0)

Sign in to comment.