acceptodds
Under review as a conference paper at ICLR 2027

Tail-MDP: Multi-Dimensional Long-Tail Learning with Prompt-Guided Decoding for Trajectory Prediction

Abstract

Trajectory prediction datasets typically exhibit pronounced long-tail distributions, causing trained models to underperform in rare yet safety-critical scenarios. Existing long-tail learning methods mainly characterize long-tailedness from a single dimension, making them insufficient for driving scenarios with multi-dimensional tail properties. Moreover, they primarily focus on representation refinement while overlooking the decoding stage, leaving encoded tail cues insufficiently exploited for trajectory generation. To this end, we propose Tail-MDP, a multi-dimensional long-tail learning framework with tail-aware decoding for trajectory prediction. We first reformulate tail sample identification as an anomaly detection problem with respect to head-dominant patterns, and use autoencoder reconstruction errors as a unified criterion to quantify long-tailedness across three key dimensions affecting future motion—motion dynamics, social interactions, and topological constraints. Here, the autoencoders are refined through training sample selection and sparse reconstruction to improve estimation reliability. Based on the resulting multi-dimensional tailness estimates, we introduce dimension-wise contrastive learning to learn representations that capture diverse tail characteristics. Furthermore, we develop a prompt-guided decoding mechanism that generates prompts according to the sample’s tailness degree and injects them into the decoder via prefix tuning, thereby guiding the model to focus on tail cues during trajectory generation. Experiments on Argoverse2 and nuScenes show that Tail-MDP consistently improves performance on diverse tail scenarios, outperforms SOTA long-tail methods, and generalizes across competitive baselines.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving, Trajectory Prediction, Long-tail Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_yZw8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Tail-MDP, a multi-dimensional long-tail learning framework with tail-aware decoding for trajectory prediction. The authors reformulate tail sample identification as an anomaly detection problem relative to head-dominant patterns, and use autoencoder reconstruction errors as a unified criterion to quantify tailness across three key dimensions affecting future motion. In addition, a lightweight prompt-guided decoding strategy is proposed to direct model attention toward tail cues during generation. Experiments are conducted on Argoverse2 and nuScenes.

Critical strengths

1.Long-tail identification is formulated as a unified multi-dimensional anomaly detection problem, with SAE head prototypes used to amplify reconstruction errors for tail samples.

2.Multi-dimensional tailness is explicitly incorporated into the decoding stage through lightweight prompt-guided tail-aware generation.

Critical weaknesses

1.Dataset coverage is somewhat limited. The empirical study only uses Argoverse2 and nuScenes, without including other widely used benchmarks such as Waymo Open Motion or INTERACTION.

  1. The novelty is mainly compositional, and the distinction from existing multi-dimensional long-tail work is not sharp enough. The framework combines AE/SAE-based anomaly detection, contrastive learning, prompt pools, and Prefix-Tuning, all of which have precedents. Multi-dimensional long-tail modeling has also been explored in recent work such as CDKFormer, AMD, and SAML. The paper would be stronger if it more explicitly clarified its unique contribution beyond these methods, especially the SAE head-prototype mechanism and the coupling between multi-dimensional tailness and prompt-guided decoding.
Questions for authors

1.Dataset generalization and evaluation fairness: Could the authors clarify why only Argoverse2 and nuScenes are used, and whether at least one additional widely used benchmark such as Waymo Open Motion could be added?

2.Novelty positioning relative to prior multi-dimensional long-tail work: How does Tail-MDP fundamentally differ from recent multi-dimensional long-tail methods such as CDKFormer, AMD, SAML, and SAIL? Is the main contribution the SAE head-prototype mechanism, the prompt-guided decoding, or their coupling, rather than multi-dimensional long-tail modeling itself?

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

The paper studies long-tail trajectory prediction. It argues that rare driving scenes may involve unusual motion, interaction or road topology, whereas many existing approaches use a single rarity dimension and concentrate on encoder learning. Tail-MDP builds on Forecast-MAE and trains three autoencoders on samples that a consensus of pretrained predictors considers easy. Reconstruction error provides a rarity score for each dimension. The scores guide dimension-wise contrastive losses and assign head, middle and tail tiers for decoder prompts. Training uses the reconstruction-based tier, while inference uses a classifier's predicted tier. The authors evaluate baseline-error, Kalman-filter-error and state-based tail subsets on Argoverse 2 and nuScenes. They report gains under all three criteria, improvements over long-tail methods ported to the same backbone, and further results on EMP and SHARP.

Critical strengths

A sensible decomposition, evaluated broadly. Motion, interaction and topology capture different reasons why a driving scene may be difficult. The paper evaluates three tail criteria on both AV2 and nuScenes. Improvements also appear under KF-error and state-based criteria, which are independent of the evaluated predictor. The nuScenes KF Top-1% minFDE improvement from 7.67 to 5.61 is particularly meaningful.

The method transfers across three forecasting backbones. Table 3 shows tail gains on Forecast-MAE, EMP and SHARP, rather than only the backbone used to develop the method. Forecast-MAE Top-1% minFDE falls from 14.80 to 9.18; SHARP improves from 11.78 to 9.40. The smaller relative gain on SHARP still supports applicability to a stronger predictor.

Critical weaknesses

The benefit of the proposed rarity scores and routed prompts is not established. Multi-factor rarity and adaptive decoding have precedents in SAML (AAAI 2026), AMD (ICCV 2025), FEND (CVPR 2023) and PerReg+ (CVPR 2025). Tail-MDP's specific choices are reconstruction-based scores and tier-routed prompts. Table 6 shows that contrastive learning helps and that prompt decoding improves Top-1% minFDE from 9.86 to 9.18. However, the reconstruction score is never replaced by KF error or a simple multi-factor score under the same losses. The routed prompts are also not compared with a shared prompt of the same length. Tables 4–5 test diagnostic autoencoders trained on KF/state-selected head sets, so they only indirectly validate the deployed consensus-selected scores. Score substitution and matched prompt/routing controls would establish what these choices add.

Questions for authors

How were the ported long-tail methods obtained and tuned? What exactly do +Resample and +Reweight do, and which dataset, thresholds and group sizes are used in Tables 1 and 3?

With the contrastive losses fixed, what happens if reconstruction scores are replaced by KF error or a simple multi-factor score? How do shared prompts and random/oracle tier routing compare at matched prompt length and parameter count?

Reviewer_ijJNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies long-tailed distributions in trajectory prediction. First, the paper proposes a multi-dimensional learning framework to quantify long-tailedness across dimensions. Then, the paper introduces dimension-wise contrastive learning to capture diverse tail characteristics. Lastly, the paper develops a prompt-guided decoding mechanism to improve trajectory generation in long-tail scenarios. Experimental results demonstrate that the proposed method achieves state-of-the-art performance in long-tail trajectory prediction.

Critical strengths

The prompt-guided decoding mechanism for improving long-tail trajectory prediction appears to be a novel and promising contribution.

Critical weaknesses

In my view, the paper introduces too many components, including multi-dimensional long-tail identification, dimension-wise contrastive learning, and prompt-guided decoding, which looks like a combination of three different works. Although each component has its own motivation, integrating them into a single framework makes the overall method overly complex and potentially difficult to reproduce and implement. I recommend that the authors release the source code to improve reproducibility and facilitate further research.

Questions for authors
  1. Could you please provide more details on how the head prototypes are generated?
  2. How long does it take to complete the entire pipeline of the proposed method?
3 comments

Thank you for addressing my questions. Your responses resolved my concerns, and I appreciate your efforts to release the code. I will raise my score to 3.

Dear Reviewer ijJN,

Thank you very much for your feedback. We sincerely appreciate your time and constructive comments, which helped us improve our work greatly.

We wish you all the best.

Dear Reviewer ijJN,

Thank you again for your feedback. May we kindly ask if you have any remaining concerns? If not, we really appreciate it if you could update your score before the deadline.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.