Reason or Recite: Complementary Specialists with an LLM Judge for Traffic Anomaly Reasoning

• Closed-form questions reward evidence-grounded reasoning, while seven open-ended tasks use BERTScore against one reference, making annotation-register matching important.A model that reasons better than its supervision can score worse when its answers diverge from that reference style.Source: §1 · pp. 1–2
• RoR separates the objectives with a structured-trace reasoning specialist, an answer-only specialist, and a frozen LLM judge that chooses per question.The reasoning model is a LoRA-tuned 9B VLM; the answer model is a QLoRA-fine-tuned 27B VLM.Source: §1 · p. 2
• TAR-RE rewrites official reasoning traces into an explicit, temporally grounded schema, substantially improving closed-form QA after fine-tuning.Source: §1 · p. 2
• The answer-only QLoRA specialist outperforms reasoning-tuned models on BERTScore-scored open-ended tasks, while unconstrained judging scores below the stronger specialist and conservative arbitration gains over both.Source: §1 · p. 2
• 0.6788 mean score earned RoR first place on the in-domain TAR leaderboard, including perfect binary event verification.Source: §1 · p. 2

2.1 Vision-Language Models for Video Understanding

• Rather than encoding time in video architecture or retrieval, RoR makes timestamped evidence explicit in structured supervision.The surrounding work includes timestamp-bound features, event boundaries, causal traces, and temporally indexed evidence banks.Source: §2.1 · pp. 2–3

2.2 From Anomaly Detection to Reasoning Supervision

• RoR restructures weak auto-labeled reasoning traces into an evidence-grounded schema, a cheaper alternative to RL rollouts that remains compatible with lightweight LoRA fine-tuning.Prior work spans chain-of-thought prompting, trace structure, rationale distillation, trace rewriting, and reinforcement learning.Source: §2.2 · p. 3

2.3 Output Selection and LLM-as-Judge

• Unlike pre-inference model routers, RoR’s training-free judge arbitrates after generation between two heterogeneous fine-tuned specialists for each question.Related selection approaches include self-consistency, output ranking and fusion, automated judging, and cost-quality routing.Source: §2.3 · p. 3

2.4 VLM Solutions in the AI City Challenge

• Unlike prior AI City systems using static task routing or caption decomposition, RoR judges candidates per question and restructures reasoning traces.Closest cited systems route by task type or question type using precomputed routing choices.Source: §2.4 · p. 3

3.1 The TAR Benchmark

• The in-domain test set has 960 human-verified annotations from 80 clips, and the unweighted leaderboard mean covers nine scored task types.The benchmark has ten task types; temporal localization is excluded from organizer scoring.Source: §3.1 · p. 4

3.2 How the Annotations Were Made, and Why It Matters

• The training set contains 44,040 annotations over 3,670 CCTV videos, and its Q&A reasoning traces are machine-generated by a three-stage VLM pipeline.For 910 videos, preexisting NVIDIA human annotations supplied supplementary context; test references are human-verified.Source: §3.2 · p. 4
• Some traces justify answers using captions or annotator notes unavailable to the model; conservative phrase matching flags 10.1% of scene-description reasoning and about 1% overall.The authors treat these rates as a lower bound because only explicit wording is detectable.Source: §3.2 · p. 4
• Timestamps occur in 66.1% of annotations but use four inconsistent conventions, including 911 timestamps with colon decimal separators.113 gold answers mix two timestamp conventions within one JSON object.Source: §3.2 · p. 4
• Manual inspection found annotations with incorrect premises or timing, including a collision question family about an event that never occurs; one upstream hallucination can propagate across tasks for a video.The authors do not relabel these cases because gold outcomes anchor reasoning regeneration, so this noise bounds what can be extracted from the training distribution.Source: §3.2 · p. 5
• Excluding scene description, traces anchor a median of two observations to timestamps, and 33.1% cite no moment; 46.1% of “No” verification traces cite none.The paper notes that establishing a negative binary answer requires scanning the whole video.Source: §3.2 · p. 5
• Traces average 109–244 words across task types and often restate their answers, with reasoning-answer ROUGE-L reaching 0.33.The paper identifies this shallowness alongside the four defect classes as motivation for regenerating reasoning supervision.Source: §3.2 · p. 5

4.1 Overview

• For MCQ-OE, RoR reuses the matched option text from the corresponding MCQ prediction rather than generating an answer, exploiting the benchmark’s paired items.Under single-reference BERTScore, exact wording is the strongest submission when the underlying MCQ prediction is correct.Source: §4.1 · p. 6

4.2 Structured Reasoning Enhancement (TAR-RE)

• TAR-RE canonicalizes timestamps, then regenerates traces with scene, query, timestamped evidence, reason, and conclusion while preserving the gold outcome.Validation enforces structural tags and depth floors and rejects traces containing the gold answer verbatim; the schema separates visible evidence from inference.Source: §4.2 · pp. 6–7, Figure 2 · p. 8
• ROUGE-L overlap roughly halved on every open-ended task; over 99% of traces use unified timestamps, and traces without a cited moment fell from 33.1% to 0.2%.The median number of time-anchored observations rose from two to 16–30; these statistics measure structure, not factual grounding.Source: §4.2 · p. 7

4.3 Reasoning Specialist: 9B LoRA on TAR-RE

• The reasoning specialist is a Qwen 3.5 9B VLM trained to produce the full TAR-RE trace followed by the answer across all ten task types.Question and video tokens are masked from the loss; greedy decoding splits at </think> to extract the answer.Source: §4.3 · p. 8

4.4 Answer Specialist: 27B QLoRA, Answer-Only SFT

• The answer specialist uses QLoRA to fine-tune a Qwen 3.6 27B model on answers alone, targeting the annotation distribution for the seven open-ended tasks.The model is weaker on closed-form questions, where it may repeat a salient prior instead of verifying the video; BERTScore barely separates the false and true verdicts (0.57 vs. 0.61).Source: §4.4 · pp. 9–10

4.5 Judge-Guided Selection

• A frozen, training-free Qwen 3.6 27B judge compares both candidates per question and keeps the task-type specialist with the higher validation score unless video evidence contradicts it.A deterministic pre-filter settles uncontested items, while the judge handles disagreements.Source: §4.1 · p. 6, §4.5 · p. 10

5.1 Implementation Details

• The training pipeline used chunked cross-entropy to avoid materializing the full vocabulary logits tensor and kept prompt construction byte-identical between training and inference.The authors implemented the pipeline because widely used fine-tuning frameworks did not then offer stable video-input support for these Qwen models.Source: §5.1 · p. 11
• Training used rank-64 LoRA adapters, 32 video frames, and a 5120-token maximum sequence length; the 27B base was NF4-quantized while its vision tower remained in bf16.The 9B and 27B models trained for four and two epochs, respectively, with greedy decoding at inference.Source: §5.1 · p. 11
• Both specialists trained on the video-level 80% split, while validation guided development and the internal test split was reserved for ablations without development-time use.Splitting was stratified by source dataset, with all annotations from each video assigned to one split.Source: §5.1 · p. 11

5.2 Main Results

• RoR remained within 0.025 of the per-column best on every other task, with its per-column pattern matching the specialists’ division of labor on validation.The leaderboard scores the combined submission.Source: §5.2 · p. 12

5.3 Ablation Study

• On the 9B backbone, TAR-RE trace supervision reached 100% multiple-choice accuracy and improved binary verification, while answer-only training led BCQ-OE by 5.1 BERTScore points.Open QA and scene description were within two points; trace supervision favored exact-scored tasks but cost similarity on reference-scored tasks.Source: §5.3 · p. 12
• Judge-guided selection scored 0.5043 versus 0.4958 for the 27B-only ceiling and 0.4887 for a neutral judge rubric on the internal test split.An oracle reached 0.5347, and the constrained judge recovered roughly a fifth of that gap.Source: §5.3 · pp. 12–13
• Per-question flips offered oracle gains across five free-text task types, ranging from +0.060 per flipped BCQ-OE item to +0.008 for scene description; video summarization had no flips.The recoverable score concentrated in reasoning-heavy tasks that static task-to-model routing would discard.Source: §5.3 · p. 13

5.4 What Did Not Work

• Rewriting, cleanup, fusion, and truncation regressed against a single reference, while adding verifiable location overlays or event-window anchors improved leaderboard scores by +0.004 and +0.008, respectively.For scene description, the remaining constraint was precision of what the model asserted it saw: deployable recall gains brought larger precision losses, and tested capacity levers tied or regressed.Source: §5.4 · p. 14

6 Conclusion

• RoR ranked first in the 2026 AI City Challenge Traffic Anomaly Reasoning track with a mean score of 0.6788.The entry combined an annotation audit, structured timestamp-grounded reasoning supervision, two complementary specialists, and per-question LLM judging.Source: §6 · p. 14
• Unconstrained multimodal judging was unreliable, while position-bias control made arbitration profitable.Source: §6 · p. 14
• With auto-labeled supervision and single-reference similarity scoring, reasoning quality and reference matching are separable objectives.The paper concludes that conflating them in one model leaves accuracy on the table.Source: §6 · p. 14