pith. sign in

arxiv: 2509.20102 · v2 · submitted 2025-09-24 · 💻 cs.AI · cs.RO

Steerable Adversarial Scenario Generation through Test-Time Preference Alignment

Tong Nie , Yuewen Mei , Yihong Tang , Junlin He , Jie Sun , Haotian Shi , Wei Ma , Jian Sun This is my paper

Pith reviewed 2026-05-18 14:08 UTC · model grok-4.3

classification 💻 cs.AI cs.RO
keywords adversarial scenario generationpreference alignmentautonomous drivingtest-time controlpolicy interpolationlinear mode connectivitymulti-objective optimization
0
0 comments X p. Extension

The pith

SAGE steers adversarial driving scenarios at test time by linearly interpolating weights between two expert policies trained on opposing preferences.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper reframes adversarial scenario generation for autonomous driving as a multi-objective preference alignment problem. Instead of producing one fixed model, SAGE fine-tunes two separate expert policies, one favoring adversariality and one favoring realism. At inference time it creates any desired balance by taking a weighted average of their parameters. This removes the need to retrain when requirements change and supports more effective closed-loop policy training. A sympathetic reader would see value in the resulting flexibility for safety assessment of self-driving systems.

Core claim

SAGE fine-tunes two expert policies on opposing preferences using hierarchical group-based preference optimization, then constructs a continuous spectrum of policies at inference time by linearly interpolating their weights, with justification from linear mode connectivity, thereby enabling fine-grained control over the adversariality-realism trade-off without retraining.

What carries the argument

Test-time linear interpolation of weights between two expert policies, each fine-tuned on a single preference pole via hierarchical group-based preference optimization.

If this is right

  • Scenarios can be generated with any chosen balance of adversariality and realism on demand at inference time.
  • Closed-loop training of driving policies improves because scenarios can be tailored to the current training stage without retraining the generator.
  • A single pair of expert models replaces the need for multiple separately trained generators for different trade-offs.
  • The approach decouples hard feasibility constraints from soft preferences during offline alignment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same interpolation construction could be tested in other multi-objective generation settings such as safety-constrained content or robotics planning.
  • Checking whether the interpolated policies maintain stability when deployed in full closed-loop simulators would be a direct next experiment.
  • Extending the method to more than two experts might produce richer non-linear trade-off surfaces.

Load-bearing premise

Linear interpolation of the weights of two separately fine-tuned expert policies produces a continuous and meaningful spectrum that correctly trades off adversariality against realism.

What would settle it

Measuring adversariality and realism metrics while sweeping the interpolation coefficient and finding no smooth monotonic trade-off, or finding that intermediate points are dominated by one of the endpoint experts.

Figures

Figures reproduced from arXiv: 2509.20102 by Haotian Shi, Jian Sun, Jie Sun, Junlin He, Tong Nie, Wei Ma, Yihong Tang, Yuewen Mei.

Figure 1
Figure 1. Figure 1: Limitation of existing adversarial generation methods, our solution, and its application. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Behavioral realism comparison. Qualitative results in [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Pareto front and continuous performance transition at test time. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: More aggressive behaviors are generated by increasing [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Empirical evidence for LMC and preference vector extrapolation in the weight space. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation studies. We conduct ablation studies to validate key de￾sign choices [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Training performances of the agent under different scenario generation methods. [PITH_FULL_IMAGE:figures/full_fig_p033_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of vector-based extrapolation and direct weight extrapolation. [PITH_FULL_IMAGE:figures/full_fig_p034_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Examples of generated scenarios by different methods. [PITH_FULL_IMAGE:figures/full_fig_p035_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Examples of generated scenarios by different methods. [PITH_FULL_IMAGE:figures/full_fig_p036_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Examples of generated scenarios (SAGE) under different adversarial weights. [PITH_FULL_IMAGE:figures/full_fig_p036_11.png] view at source ↗
read the original abstract

Adversarial scenario generation is a cost-effective approach for safety assessment of autonomous driving systems. However, existing methods are often constrained to a single, fixed trade-off between competing objectives such as adversariality and realism. This yields behavior-specific models that cannot be steered at inference time, lacking the efficiency and flexibility to generate tailored scenarios for diverse training and testing requirements. In view of this, we reframe the task of adversarial scenario generation as a multi-objective preference alignment problem and introduce a new framework named \textbf{S}teerable \textbf{A}dversarial scenario \textbf{GE}nerator (SAGE). SAGE enables fine-grained test-time control over the trade-off between adversariality and realism without any retraining. We first propose hierarchical group-based preference optimization, a data-efficient offline alignment method that learns to balance competing objectives by decoupling hard feasibility constraints from soft preferences. Instead of training a fixed model, SAGE fine-tunes two experts on opposing preferences and constructs a continuous spectrum of policies at inference time by linearly interpolating their weights. We provide theoretical justification for this framework through the lens of linear mode connectivity. Extensive experiments demonstrate that SAGE not only generates scenarios with a superior balance of adversariality and realism but also enables more effective closed-loop training of driving policies. Project page: https://tongnie.github.io/SAGE/.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces SAGE, a framework that reframes adversarial scenario generation for autonomous driving as a multi-objective preference alignment task. It proposes hierarchical group-based preference optimization as a data-efficient offline method to fine-tune two expert policies on opposing preferences (high adversariality vs. high realism), then constructs a continuous spectrum of policies at inference time via linear interpolation of their weights, justified by linear mode connectivity. The central claims are that this enables fine-grained test-time steering of the adversariality-realism trade-off without retraining and yields scenarios with superior balance that improve closed-loop driving policy training.

Significance. If the linear interpolation reliably produces controllable intermediate policies and the experimental gains hold, SAGE would provide an efficient, flexible alternative to retraining separate models for different scenario requirements in AV safety testing and training. The offline preference optimization approach could also generalize to other multi-objective generative tasks in simulation.

major comments (2)
  1. [theoretical justification section] Theoretical justification section: the appeal to linear mode connectivity to justify that weight interpolation between opposing-preference experts yields a continuous, meaningful trade-off spectrum is load-bearing for the steerability claim, yet the manuscript does not appear to provide a direct empirical check (e.g., plots of adversariality and realism metrics vs. interpolation coefficient) or a proof that the loss surface remains low along the line when the objectives are antagonistic; without this, the construction risks producing incoherent policies.
  2. [Experiments section] Experiments section (and abstract): the claims of 'superior balance' and 'more effective closed-loop training' rest on unshown quantitative results; the manuscript should include explicit tables with baselines, effect sizes, standard errors, and ablation on the interpolation step to allow verification that the spectrum is monotonic and useful.
minor comments (1)
  1. [Abstract] Abstract: the description of how 'hierarchical group-based preference optimization' decouples hard feasibility constraints from soft preferences could be clarified with a brief algorithmic outline or pseudocode reference.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and indicate the changes planned for the revised manuscript.

read point-by-point responses

  1. Referee: [theoretical justification section] Theoretical justification section: the appeal to linear mode connectivity to justify that weight interpolation between opposing-preference experts yields a continuous, meaningful trade-off spectrum is load-bearing for the steerability claim, yet the manuscript does not appear to provide a direct empirical check (e.g., plots of adversariality and realism metrics vs. interpolation coefficient) or a proof that the loss surface remains low along the line when the objectives are antagonistic; without this, the construction risks producing incoherent policies.

    Authors: We thank the referee for highlighting this key aspect of our steerability claim. The manuscript justifies weight interpolation via linear mode connectivity, an empirically observed property of neural network loss landscapes that has been documented across many architectures and tasks. To directly address the request for empirical verification, the revised manuscript will add plots of adversariality and realism metrics versus the interpolation coefficient. These will show that the trade-off varies continuously and produces coherent intermediate policies. A formal proof that the loss surface remains low for antagonistic objectives is not provided, as establishing such a guarantee for arbitrary opposing preferences remains an open theoretical question; we will expand the discussion section to explicitly note this limitation while emphasizing the supporting experimental evidence. revision: partial

  2. Referee: [Experiments section] Experiments section (and abstract): the claims of 'superior balance' and 'more effective closed-loop training' rest on unshown quantitative results; the manuscript should include explicit tables with baselines, effect sizes, standard errors, and ablation on the interpolation step to allow verification that the spectrum is monotonic and useful.

    Authors: We agree that explicit quantitative reporting strengthens verifiability. Although the manuscript already presents experimental outcomes demonstrating superior balance and improved closed-loop training, we will revise the Experiments section and update the abstract to include detailed tables. These tables will report comparisons against baselines, effect sizes, standard errors computed over multiple random seeds, and an ablation isolating the interpolation step. The added results will confirm monotonicity of the spectrum and its practical utility. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on independent linear mode connectivity literature and empirical validation

full rationale

The paper's core construction—fine-tuning two opposing-preference expert policies then forming a continuous spectrum via linear weight interpolation at inference time—is presented as a practical method whose justification invokes linear mode connectivity as an external lens rather than a result derived or fitted inside the paper. No equations reduce the interpolation outcome to a tautological fit of the same data, no self-citation chain bears the central claim, and the method is not defined in terms of its own predictions. The framework remains self-contained against external benchmarks and falsifiable via the reported experiments on scenario quality and downstream policy training.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review performed on abstract only; no explicit free parameters, axioms, or invented entities are stated in the provided text. The linear mode connectivity justification is treated as an external theoretical assumption rather than a paper-specific invention.

pith-pipeline@v0.9.0 · 5785 in / 1260 out tokens · 31868 ms · 2026-05-18T14:08:37.986991+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages · 4 internal anchors

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

    work page

  2. [2]

    K., Hayase, J., and Srinivasa, S

    Samuel K Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. arXiv preprint arXiv:2209.04836, 2022

    work page arXiv 2022

  3. [3]

    An end-to-end curriculum learning approach for autonomous driving scenarios

    Luca Anzalone, Paola Barra, Silvio Barra, Aniello Castiglione, and Michele Nappi. An end-to-end curriculum learning approach for autonomous driving scenarios. IEEE Transactions on Intelligent Transportation Systems, 23 0 (10): 0 19817--19826, 2022

    work page 2022

  4. [4]

    Constitutional AI: Harmlessness from AI Feedback

    Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  5. [5]

    Advdo: Realistic adversarial attacks for trajectory prediction

    Yulong Cao, Chaowei Xiao, Anima Anandkumar, Danfei Xu, and Marco Pavone. Advdo: Realistic adversarial attacks for trajectory prediction. In European Conference on Computer Vision, pp.\ 36--52. Springer, 2022

    work page 2022

  6. [6]

    Reinforcement learning with human feedback for realistic traffic simulation

    Yulong Cao, Boris Ivanovic, Chaowei Xiao, and Marco Pavone. Reinforcement learning with human feedback for realistic traffic simulation. In 2024 IEEE international conference on robotics and automation (ICRA), pp.\ 14428--14434. IEEE, 2024

    work page 2024

  7. [7]

    Safe-sim: Safety-critical closed-loop traffic simulation with diffusion-controllable adversaries

    Wei-Jer Chang, Francesco Pittaluga, Masayoshi Tomizuka, Wei Zhan, and Manmohan Chandraker. Safe-sim: Safety-critical closed-loop traffic simulation with diffusion-controllable adversaries. In European Conference on Computer Vision, pp.\ 242--258. Springer, 2024

    work page 2024

  8. [8]

    Frea: Feasibility-guided generation of safety-critical scenarios with reasonable adversariality

    Keyu Chen, Yuheng Lei, Hao Cheng, Haoran Wu, Wenchao Sun, and Sifa Zheng. Frea: Feasibility-guided generation of safety-critical scenarios with reasonable adversariality. arXiv preprint arXiv:2406.02983, 2024

    work page arXiv 2024

  9. [9]

    Rift: Closed-loop rl fine-tuning for realistic and controllable traffic simulation

    Keyu Chen, Wenchao Sun, Hao Cheng, and Sifa Zheng. Rift: Closed-loop rl fine-tuning for realistic and controllable traffic simulation. arXiv preprint arXiv:2505.03344, 2025

    work page arXiv 2025

  10. [10]

    Multi-objective optimization

    Kalyanmoy Deb, Karthik Sindhya, and Jussi Hakanen. Multi-objective optimization. In Decision sciences, pp.\ 161--200. CRC Press, 2016

    work page 2016

  11. [11]

    Learning to collide: An adaptive safety-critical scenarios generating method

    Wenhao Ding, Baiming Chen, Minjun Xu, and Ding Zhao. Learning to collide: An adaptive safety-critical scenarios generating method. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp.\ 2243--2250. IEEE, 2020

    work page 2020

  12. [12]

    A survey on safety-critical driving scenario generation—a methodological perspective

    Wenhao Ding, Chejian Xu, Mansur Arief, Haohong Lin, Bo Li, and Ding Zhao. A survey on safety-critical driving scenario generation—a methodological perspective. IEEE Transactions on Intelligent Transportation Systems, 24 0 (7): 0 6971--6988, 2023

    work page 2023

  13. [13]

    Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset

    Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 9710--9719, 2021

    work page 2021

  14. [14]

    Intelligent driving intelligence test for autonomous vehicles with naturalistic and adversarial environment

    Shuo Feng, Xintao Yan, Haowei Sun, Yiheng Feng, and Henry X Liu. Intelligent driving intelligence test for autonomous vehicles with naturalistic and adversarial environment. Nature communications, 12 0 (1): 0 748, 2021

    work page 2021

  15. [15]

    Dense reinforcement learning for safety validation of autonomous vehicles

    Shuo Feng, Haowei Sun, Xintao Yan, Haojie Zhu, Zhengxia Zou, Shengyin Shen, and Henry X Liu. Dense reinforcement learning for safety validation of autonomous vehicles. Nature, 615 0 (7953): 0 620--627, 2023

    work page 2023

  16. [16]

    Linear mode connectivity and the lottery ticket hypothesis

    Jonathan Frankle, Gintare Karolina Dziugaite, Daniel Roy, and Michael Carbin. Linear mode connectivity and the lottery ticket hypothesis. In International Conference on Machine Learning, pp.\ 3259--3269. PMLR, 2020

    work page 2020

  17. [17]

    Addressing function approximation error in actor-critic methods

    Scott Fujimoto, Herke Hoof, and David Meger. Addressing function approximation error in actor-critic methods. In International conference on machine learning, pp.\ 1587--1596. PMLR, 2018

    work page 2018

  18. [18]

    Densetnt: End-to-end trajectory prediction from dense goal sets

    Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 15303--15312, 2021

    work page 2021

  19. [19]

    King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients

    Niklas Hanselmann, Katrin Renz, Kashyap Chitta, Apratim Bhattacharyya, and Andreas Geiger. King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients. In European Conference on Computer Vision, pp.\ 335--352. Springer, 2022

    work page 2022

  20. [20]

    Editing Models with Task Arithmetic

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  21. [21]

    Deep learning without poor local minima

    Kenji Kawaguchi. Deep learning without poor local minima. Advances in neural information processing systems, 29, 2016

    work page 2016

  22. [22]

    Training adversarial agents to exploit weaknesses in deep control policies

    Sampo Kuutti, Saber Fallah, and Richard Bowden. Training adversarial agents to exploit weaknesses in deep control policies. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pp.\ 108--114. IEEE, 2020

    work page 2020

  23. [23]

    Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning

    Quanyi Li, Zhenghao Peng, Lan Feng, Qihang Zhang, Zhenghai Xue, and Bolei Zhou. Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning. IEEE transactions on pattern analysis and machine intelligence, 45 0 (3): 0 3461--3475, 2022

    work page 2022

  24. [24]

    Curse of rarity for autonomous vehicles

    Henry X Liu and Shuo Feng. Curse of rarity for autonomous vehicles. nature communications, 15 0 (1): 0 4808, 2024

    work page 2024

  25. [25]

    Peo: Improving bi-factorial preference alignment with post-training policy extrapolation

    Yuxuan Liu. Peo: Improving bi-factorial preference alignment with post-training policy extrapolation. arXiv preprint arXiv:2503.01233, 2025

    work page arXiv 2025

  26. [26]

    Improved robustness and safety for autonomous vehicle control with adversarial reinforcement learning

    Xiaobai Ma, Katherine Driggs-Campbell, and Mykel J Kochenderfer. Improved robustness and safety for autonomous vehicle control with adversarial reinforcement learning. In 2018 IEEE Intelligent Vehicles Symposium (IV), pp.\ 1665--1671. IEEE, 2018

    work page 2018

  27. [27]

    Bayesian fault injection safety testing for highly automated vehicles with uncertainty

    Yuewen Mei, Tong Nie, Jian Sun, and Ye Tian. Bayesian fault injection safety testing for highly automated vehicles with uncertainty. IEEE Transactions on Intelligent Vehicles, 2024

    work page 2024

  28. [28]

    Llm-attacker: Enhancing closed-loop adversarial scenario generation for autonomous driving with large language models

    Yuewen Mei, Tong Nie, Jian Sun, and Ye Tian. Llm-attacker: Enhancing closed-loop adversarial scenario generation for autonomous driving with large language models. IEEE Transactions on Intelligent Transportation Systems, 2025 a

    work page 2025

  29. [29]

    Seeking to collide: Online safety-critical scenario generation for autonomous driving with retrieval augmented large language models

    Yuewen Mei, Tong Nie, Jian Sun, and Ye Tian. Seeking to collide: Online safety-critical scenario generation for autonomous driving with retrieval augmented large language models. arXiv preprint arXiv:2505.00972, 2025 b

    work page arXiv 2025

  30. [30]

    Exploring the roles of large language models in reshaping transportation systems: A survey, framework, and roadmap

    Tong Nie, Jian Sun, and Wei Ma. Exploring the roles of large language models in reshaping transportation systems: A survey, framework, and roadmap. Artificial Intelligence for Transportation, 1: 0 100003, 2025

    work page 2025

  31. [31]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 0 27730--27744, 2022

    work page 2022

  32. [32]

    Aed: Automatic discovery of effective and diverse vulnerabilities for autonomous driving policy with large language models

    Le Qiu, Zelai Xu, Qixin Tan, Wenhao Tang, Chao Yu, and Yu Wang. Aed: Automatic discovery of effective and diverse vulnerabilities for autonomous driving policy with large language models. arXiv preprint arXiv:2503.20804, 2025

    work page arXiv 2025

  33. [33]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36: 0 53728--53741, 2023

    work page 2023

  34. [34]

    Rewarded soups: towards pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards

    Alexandre Rame, Guillaume Couairon, Corentin Dancette, Jean-Baptiste Gaya, Mustafa Shukor, Laure Soulier, and Matthieu Cord. Rewarded soups: towards pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards. Advances in Neural Information Processing Systems, 36: 0 71095--71134, 2023

    work page 2023

  35. [35]

    Goose: Goal-conditioned reinforcement learning for safety-critical scenario generation

    Joshua Ransiek, Johannes Plaum, Jacob Langner, and Eric Sax. Goose: Goal-conditioned reinforcement learning for safety-critical scenario generation. In 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), pp.\ 2651--2658. IEEE, 2024 a

    work page 2024

  36. [36]

    Adversarial and reactive traffic agents for realistic driving simulation

    Joshua Ransiek, Philipp Reis, and Eric Sax. Adversarial and reactive traffic agents for realistic driving simulation. arXiv e-prints, pp.\ arXiv--2409, 2024 b

    work page 2024

  37. [37]

    Generating useful accident-prone driving scenarios via a learned traffic prior

    Davis Rempe, Jonah Philion, Leonidas J Guibas, Sanja Fidler, and Or Litany. Generating useful accident-prone driving scenarios via a learned traffic prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 17305--17315, 2022

    work page 2022

  38. [38]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  39. [39]

    Decoding-time language model alignment with multiple objectives

    Ruizhe Shi, Yifang Chen, Yushi Hu, Alisa Liu, Hanna Hajishirzi, Noah A Smith, and Simon S Du. Decoding-time language model alignment with multiple objectives. Advances in Neural Information Processing Systems, 37: 0 48875--48920, 2024

    work page 2024

  40. [40]

    Seal: Towards safe autonomous driving via skill-enabled adversary learning for closed-loop scenario generation

    Benjamin Stoler, Ingrid Navarro, Jonathan Francis, and Jean Oh. Seal: Towards safe autonomous driving via skill-enabled adversary learning for closed-loop scenario generation. IEEE Robotics and Automation Letters, 0 (99): 0 1--8, 2025

    work page 2025

  41. [41]

    Failure-Scenario Maker for Rule-Based Agent using Multi-agent Adversarial Reinforcement Learning and its Application to Autonomous Driving

    Akifumi Wachi. Failure-scenario maker for rule-based agent using multi-agent adversarial reinforcement learning and its application to autonomous driving. arXiv preprint arXiv:1903.10654, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1903

  42. [42]

    Advsim: Generating safety-critical scenarios for self-driving vehicles

    Jingkang Wang, Ava Pun, James Tu, Sivabalan Manivasagam, Abbas Sadat, Sergio Casas, Mengye Ren, and Raquel Urtasun. Advsim: Generating safety-critical scenarios for self-driving vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 9909--9918, 2021

    work page 2021

  43. [43]

    Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time

    Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, pp.\ 23965--23998. ...

    work page 2022

  44. [44]

    Bone soups: A seek-and-soup model merging approach for controllable multi-objective generation

    Guofu Xie, Xiao Zhang, Ting Yao, and Yunsheng Shi. Bone soups: A seek-and-soup model merging approach for controllable multi-objective generation. arXiv preprint arXiv:2502.10762, 2025

    work page arXiv 2025

  45. [45]

    Advdiffuser: Generating adversarial safety-critical driving scenarios via guided diffusion

    Yuting Xie, Xianda Guo, Cong Wang, Kunhua Liu, and Long Chen. Advdiffuser: Generating adversarial safety-critical driving scenarios via guided diffusion. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp.\ 9983--9989. IEEE, 2024

    work page 2024

  46. [46]

    Safebench: A benchmarking platform for safety evaluation of autonomous vehicles

    Chejian Xu, Wenhao Ding, Weijie Lyu, Zuxin Liu, Shuai Wang, Yihan He, Hanjiang Hu, Ding Zhao, and Bo Li. Safebench: A benchmarking platform for safety evaluation of autonomous vehicles. Advances in Neural Information Processing Systems, 35: 0 25667--25682, 2022

    work page 2022

  47. [47]

    Diffscene: Diffusion-based safety-critical scenario generation for autonomous vehicles

    Chejian Xu, Aleksandr Petiushko, Ding Zhao, and Bo Li. Diffscene: Diffusion-based safety-critical scenario generation for autonomous vehicles. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 8797--8805, 2025

    work page 2025

  48. [49]

    Cat: Closed-loop adversarial training for safe end-to-end driving

    Linrui Zhang, Zhenghao Peng, Quanyi Li, and Bolei Zhou. Cat: Closed-loop adversarial training for safe end-to-end driving. In Conference on Robot Learning, pp.\ 2357--2372. PMLR, 2023

    work page 2023

  49. [50]

    On adversarial robustness of trajectory prediction for autonomous vehicles

    Qingzhao Zhang, Shengtuo Hu, Jiachen Sun, Qi Alfred Chen, and Z Morley Mao. On adversarial robustness of trajectory prediction for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 15159--15168, 2022

    work page 2022

  50. [51]

    Language-guided traffic simulation via scene-level diffusion

    Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, and Baishakhi Ray. Language-guided traffic simulation via scene-level diffusion. In Conference on robot learning, pp.\ 144--177. PMLR, 2023

    work page 2023

  51. [52]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

    work page

  52. [53]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

    work page

  53. [54]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

    work page arXiv 2023