acceptodds
Under review as a conference paper at ICLR 2027

ALERT: A Self-evolving Agent for Latent Risk Reasoning in Autonomous Driving

Abstract

Safety-critical events in autonomous driving often develop progressively before explicit conflicts emerge. However, existing approaches mainly focus on detecting manifested hazards or predicting short-term outcomes, leaving the early-stage risk evolution insufficiently explored. In this work, we introduce ALERT, a self-evolving agent framework for latent risk reasoning that aims to identify how observable cues may evolve into future driving risks. ALERT employs a mechanism-oriented risk routing taxonomy to organize diverse risk formation processes, uses tool-augmented reasoning to acquire driving-specific evidence, and continuously improves its Skill Pool through controlled evolution. To enable systematic evaluation, we construct a latent risk reasoning benchmark from real-world takeover data covering risk stages, risk types, risk subjects and causal reasoning. ALERT outperforms the strongest VLM-only baseline by 11.7% in overall score and further surpasses knowledge-access baselines, particularly in risk-subject identification and reasoning. Controlled Skill evolution yields an additional 4.9% improvement over the warm-up stage, demonstrating the value of accumulating reusable reasoning experience for latent risk understanding.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving, Latent Risk Reasoning, Self-evolving Agent

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_YMqnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ALERT, an LLM agent that reasons about latent risks in autonomous driving, which are risks that develop before an explicit conflict emerges. The key idea is to route risk hypotheses through a 10-category taxonomy organized by risk formation mechanisms, gather evidence with driving-specific tools to verify each hypothesis, and accumulate reasoning experience as reusable Skill documents. Each Skill edit is merged only after it passes a validation check. For evaluation, the authors construct a benchmark by sampling scenes backward in time from real-world takeover logs and annotating them with GPT-5.5. On this benchmark, the paper reports higher overall scores than VLM-only and knowledge-retrieval baselines, particularly in risk-subject identification and reasoning.

Critical strengths
  1. The accumulated knowledge is human-readable and editable. Without retraining the backbone LLM, the method stores risk-specific knowledge as text Skill documents, so one can directly inspect which rules were added or revised.
Critical weaknesses
  1. Annotation and evaluation both rely on the same LLM (GPT-5.5). The ground truth is written by GPT-5.5 with access to future frames, and the risk-subject and reasoning scores, which make up 55%55\% of the overall score, are graded by GPT-5.5 using rubrics derived from that ground truth. ALERT's Skill evolution is also designed to maximize this score against training-set annotations produced in the same way (Eq. 5). Indeed, relative to the HyDE-style baseline, the gains are concentrated in the LLM-graded subject and reasoning dimensions, while the risk-type score is essentially unchanged. It is therefore unclear whether the improvement reflects better latent-risk reasoning or a closer fit to the annotator's style.

  2. The effect of controlled Skill evolution, the core contribution, is not isolated. Removing the tools lowers the overall score by about 99 points, yet there is no baseline that uses the tools without any accumulated knowledge (neither Skills nor retrieved experience). The gap to the HyDE-style baseline, which uses the same Skill Pool, is only 1.751.75 points, with no variance reported. The evidence for the validation gate (accepted edits +9.97+9.97, rejected edits −3.65-3.65) is a selection effect produced by the acceptance rule itself, which accepts an edit only if its resampled score on the same scene improves. Moreover, each edit is derived from the ground truth of a scene and validated on that same scene, so this does not show that the edits generalize to other scenes.

Questions for authors
  1. Some takeovers happen for reasons other than safety risk, such as ride comfort or system limitations. How were risk-relevant takeovers selected, and how is risk severity handled? What is the inter-annotator agreement on the stage and type labels?

  2. What does the BEV image contain (e.g., perception outputs, map, trajectories)? This is needed to judge how much information the VLM-only baselines receive.

  3. Could the authors report a baseline that uses the driving tools without any accumulated knowledge (neither Skills nor retrieved experience)? It would correspond to tool-augmented driving agents such as AgentThink or Agent-Driver, and is needed to separate the contribution of the Skills from that of the tools.

  4. Do the authors plan to release the benchmark (at least the test split and its annotations) and the code?

Reviewer_Lu7uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ALERT, a self-evolving agent framework for latent risk reasoning in autonomous driving. ALERT combines risk mechanism routing, tool-augmented reasoning, and a hierarchical Skill Pool. It continuously updates external reasoning knowledge through multi-dimensional evaluation, counterfactual validation, and conflict-aware merging, while keeping the backbone frozen. The authors also construct a latent risk reasoning benchmark based on real-world driving takeover events, covering 10 risk mechanisms. ALERT achieves an overall score of 76.75, outperforming the HyDE-style baseline (75.00) and the strongest VLM-only baseline (68.73).

Critical strengths
  1. The problem is practically relevant, and the method is systematically designed. The paper focuses on identifying latent risks before explicit conflicts occur, inferring how risks may develop from currently observable cues rather than simply classifying static scenes. ALERT integrates risk routing, tool-augmented reasoning, and continuous evolution of external Skills into a unified framework. It accumulates reusable reasoning knowledge while keeping the backbone frozen, which offers practical value.

  2. The paper provides interpretable reasoning examples and failure analysis. Through concrete driving scenarios, the paper shows how ALERT uses risk subtype knowledge, vehicle motion prediction, and driving corridor checks to identify latent risks. Comparisons before and after evolution illustrate how Skill updates affect risk routing and evidence acquisition. The paper also presents a failure case where the correct risk mechanism is identified but a secondary risk subject is missed, making the agent's behavior and limitations easier to understand.

Critical weaknesses
  1. The core Skill Evolution mechanism lacks direct validation, and the relationship between the evolution objective and the final evaluation target remains unclear. Controlled Skill Evolution is a core contribution, but the existing ablations mainly compare different evolution stages and the removal of subtype Skills, tools, or BEV. The paper does not directly compare a static Skill Pool with the full evolution method, or evolution with and without the counterfactual validation gate under the same training conditions. Thus, it is difficult to determine whether the gains come from continuous Skill accumulation, candidate edit validation, or other components. Meanwhile, counterfactual validation mainly resamples trajectories on the scene that triggered an edit to assess local improvements. This is insufficient to establish that edits generalize to unseen scenes or prevent long-term knowledge degradation. In addition, the final benchmark score consists of risk stage, risk type, risk subjects, and reasoning quality, while the evolution-stage Evaluation Agent uses a different scoring scheme that additionally evaluates evidence quality and tool-call effectiveness, with different weights across dimensions. The acceptance or rejection of candidate Skill edits directly depends on score changes under this evolution-stage evaluator. Although process-level scores may help develop more reliable reasoning strategies, the paper does not sufficiently verify whether these additional signals align with improvements in the final benchmark score or analyze their individual contributions. Therefore, there is insufficient experimental evidence that the signals driving Skill Evolution effectively optimize the final objective defined in the paper.

  2. The generalizability, statistical robustness, and external validity of the benchmark are insufficiently validated. All ALERT experiments use a frozen Qwen3.7-Plus backbone. Although the paper compares other models, including Qwen3.7-Flash, GPT-5.6-Sol, and DeepSeek-V4.1-Flash, they are only used as VLM-only baselines without Skills or tools; The full ALERT framework is not evaluated on them. Therefore, the current results do not establish whether ALERT's gains transfer to different model families or depend mainly on its adaptation to a specific backbone. Regarding experimental robustness, the test set contains 458 takeover clips and 1,825 scenes, but scenes from the same clip are correlated and cannot simply be treated as fully independent samples. The paper reports no clip-level clustered confidence intervals, paired significance tests, or results from multiple independent Skill Evolution runs. In particular, ALERT's overall advantage over HyDE-style is only 1.75 points, and the current results do not establish whether this advantage is statistically stable. Moreover, although the paper covers 10 risk mechanisms, it does not sufficiently report per-category performance, confusion patterns, or results for long-tail risk types. It is therefore difficult to determine whether the overall gains extend consistently across different risk mechanisms. The benchmark design also limits the external validity of the conclusions. The dataset uses real-world driving takeover events as temporal anchors and samples different risk stages backward in time. No-risk samples also come from clips that eventually involve a takeover, rather than being independently sampled from ordinary driving scenes without a takeover. This design helps analyze how risks evolve from early cues to explicit conflicts, but does not fully represent general road driving. Thus, the current results mainly support ALERT's latent risk reasoning ability in takeover-related scenarios, but do not establish that it can reliably distinguish no risk, latent risk, and explicit risk in general driving scenarios, particularly in terms of false alarm control. Although the authors instruct annotators to judge based on currently visible evidence and manually review the test set, the potential effects of using future context during annotation have not been sufficiently analyzed.

Questions for authors
  1. Regarding the independent contribution of Skill Evolution and its optimization objective: How do the results differ between a static Skill Pool, Skill Evolution without the counterfactual validation gate, and full ALERT? In addition, the multi-dimensional scores used during evolution do not fully match the final benchmark metrics. Have the authors analyzed the correlation between the two and verified the actual contributions of the additional evidence quality and tool-call effectiveness scores to final performance?

  2. Regarding generalizability and experimental robustness: Can the full ALERT framework transfer to backbones from different model families? Furthermore, can statistical analyses that account for correlations among scenes within the same takeover clip, together with multiple independent Skill Evolution runs, establish whether its performance advantage over HyDE-style is stable? How does ALERT perform across different risk mechanisms, especially long-tail categories?

  3. Regarding the external validity of the benchmark: Since no-risk samples also come from events that eventually involve a takeover, how well does the current benchmark reflect the risk distribution in general road driving? Can ALERT maintain reliable risk identification and reasonable false alarm rates on independent ordinary driving clips without a takeover?

Reviewer_inPbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ALERT anticipates latent driving risks through a taxonomy of ten risk mechanisms, a benchmark built from real human takeover events, and skill updates checked through counterfactual validation. It reports overall score improvements of 8.02 over VLM baselines and 1.75 over knowledge-access baselines.

Critical strengths
  1. Comparing original and updated skills on the same scene helps filter out harmful changes and makes experience accumulation more reliable.
  2. Using subsequent takeover events to annotate early risks, while restricting models to current observations, is a sensible and practically relevant design.
Critical weaknesses
  1. Restricted data access makes it difficult for external researchers to reproduce the results or independently assess annotation quality.
  2. Without cross-dataset experiments, it remains unclear whether the skills overfit to the source distribution and its annotation conventions.
Questions for authors

The cited entries in Table 7 give equal scores to errors in opposite directions. Given the different safety costs of missed risks and false alarms, what motivates this choice? Would the method rankings and main conclusions hold if missed risks were penalized more heavily?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.