mlsanityData to 7 Oct 2026

TRAM: Test-Time Risk Adaptation with Mixture of Agents

arXiv 2408.08812 ↗First posted 16 Aug 2024cs.LGcited by 1

On the map: LLM agents and tool use → Agent threats, guardrails and governance

Abstract (arXiv)

Deployed reinforcement learning agents often face safety requirements that are specified only after training, such as new hazard maps, revised risk thresholds, or behavioral alignment constraints. We study zero-update deployment-time adaptation, where a fixed library of risk-neutral source policies is reused under a newly specified reward-risk tradeoff. We propose TRAM (Test-Time Risk Adaptation via Mixture of Agents), a source-scored composition rule that evaluates each source policy under the target reward and an occupancy-based deployment risk, then selects actions using risk-adjusted source scores. Unlike training-time risk-sensitive methods tied to a fixed surrogate such as return variance, TRAM supports spatial barrier exposure, divergence from a reference behavior, and local volatility risks specified at test time. We explicitly characterize TRAM as a surrogate method: it does not solve the full occupancy-control problem of the stitched policy, but admits a measurable source-hull mismatch term connecting source-scored risk to realized risk. Experiments in gridworlds, MuJoCo Reacher, Safety-Gymnasium, and an LLM alignment setting show that TRAM reduces deployment risk while preserving reward, without requiring any parameter updates at test time.

Key facts. Cited by 1 paper in the map (data through 7 Oct 2026). Placed in Agent threats, guardrails and governance, one of 16 directions in LLM agents and tool use. Builds on 27 of its 41 references that are in the map. First posted on arXiv 16 Aug 2024.

Builds on

27 of 41 references.

References · 27

  1. 2024RewardBench: Evaluating Reward Models for Language Modelingcited by 480
  2. 2016Successor Features for Transfer in Reinforcement Learningcited by 331
  3. 2017DARLA: Improving Zero-Shot Transfer in Reinforcement Learningcited by 147
  4. 2023Safety-Gymnasium: A Unified Safe Reinforcement Learning Benchmarkcited by 135
  5. 2023Controlled Decoding from Language Modelscited by 133
  6. 2013Risk-sensitive Reinforcement Learningcited by 129
  7. 2020Learning to be Safe: Deep RL with a Safety Criticcited by 109
  8. 2017Zero-Shot Task Generalization with Multi-Task Deep Reinforcement Learningcited by 97
  9. 2019Transfer in Deep Reinforcement Learning Using Successor Features and Generalised Policy Improvementcited by 88
  10. 2022Does Zero-Shot Reinforcement Learning Exist?cited by 79
  11. 2020Safe Reinforcement Learning via Curriculum Inductioncited by 73
  12. 2020Risk-Sensitive Reinforcement Learning: Near-Optimal Risk-Sample Tradeoff in Regretcited by 57
15 more
  1. 2019Risk-Averse Trust Region Optimization for Reward-Volatility Reductioncited by 53
  2. 2023Dual RL: Unification and New Methods for Reinforcement and Imitation Learningcited by 48
  3. 2018Variance Reduction for Reinforcement Learning in Input-Driven Environmentscited by 30
  4. 2020Mean-Variance Policy Iteration for Risk-Averse Reinforcement Learningcited by 28
  5. 2019Entropic Risk Measure in Policy Searchcited by 19
  6. 2020Cautious Reinforcement Learning via Distributional Risk in the Dual Domaincited by 18
  7. 2022Hypernetworks for Zero-shot Transfer in Reinforcement Learningcited by 18
  8. 2018Learning Robust Optionscited by 17
  9. 2023Constraint-Conditioned Policy Optimization for Versatile Safe Reinforcement Learningcited by 17
  10. 2021Risk-Aware Transfer in Reinforcement Learning using Successor Featurescited by 12
  11. 2018Safe Option-Critic: Learning Safety in the Option-Critic Architecturecited by 10
  12. 2021Variance Penalized On-Policy and Off-Policy Actor-Criticcited by 10
  13. 2017Probabilistically Safe Policy Transfercited by 6
  14. 2024SF-DQN: Provable Knowledge Transfer using Successor Feature for Deep Reinforcement Learningcited by 5
  15. 2016Situational Awareness by Risk-Conscious Skillscited by 2

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026Why AI systems don't learn and what to do about it: Lessons on autonomous learning from cognitive sciencecited by 8

Similar, not cited

Close in wording, no citation either way

  1. 2026SafeEvolve: Harness-Policy Co-Evolution from Agent Experience for Safety Alignment
  2. 2026SafeCoEvo: Co-Evolving Safety Harnesses and Guards for LLM Agents at Test-Time
  3. 2025Safe Planning and Policy Optimization via World Model Learning
  4. 2024Latent Safety-Constrained Policy Approach for Safe Offline Reinforcement Learning
  5. 2026Imagine to Ensure Safety in Hierarchical Reinforcement Learning
  6. 2025DIAL: Distribution-Informed Adaptive Learning of Multi-Task Constraints for Safety-Critical Systems
  7. 2026Lyapunov-Guided Self-Alignment: Test-Time Adaptation for Offline Safe Reinforcement Learning
  8. 2026Decoupled Guidance Diffusion for Adaptive Offline Safe Reinforcement Learning