Same Trajectory, Contradictory Rewards (ROBORMBENCH): Paraphrase Fragility in Vision Language Reward Models

Same Trajectory, Contradictory Rewards (ROBORMBENCH): Paraphrase Fragility in Vision Language Reward Models

|Authors: Wonje Jeung, Sangyeon Yoon, Hyesoo Hong, Yoonjun Cho, Dongjae Jeon, Bumjun Kim, Jean Oh, Youngjae Yu, Albert No

|Venue: arXiv preprint | Year: 2026

|arXiv: 2609.05401v1

Abstract

Vision-language models are increasingly used as reward functions for robotic learning, but this role requires paraphrase invariance: the same trajectory should receive the same reward under semantically equivalent goal descriptions. We show that current VLM reward models often violate this property. Paraphrasing the instruction alone can substantially change predicted progress scores, and can even flip identical robot behavior between failure and success. To measure this failure mode, we introduce ROBORMBENCH, a benchmark with 2,390 real-robot trajectories, ground-truth progress labels, and 21,673 verified paraphrases spanning lexical, syntactic, and action-goal rewrites. Across proprietary and open-source VLMs, paraphrase-induced instability is widespread and severe, grows under more divergent rewrites, and is not reliably reduced by scale or explicit reasoning. Dedicated reward models trained with trajectory-grounded supervision are substantially more stable. These results show that paraphrase robustness is a core requirement for reliable VLM-based reward modeling in robotics.

Key Contributions

• Vision-language models are increasingly used as reward functions for robotic learning, but this r…
• We show that current VLM reward models often violate this property
• Paraphrasing the instruction alone can substantially change predicted progress scores, and can ev…
• To measure this failure mode, we introduce ROBORMBENCH, a benchmark with 2,390 real-robot traject…
• Across proprietary and open-source VLMs, paraphrase-induced instability is widespread and severe,…

Topics

• vision
• vla
• benchmark

Code & Data

No code repository linked in paper metadata.

BibTeX

@article{Jeung2026_260905401v1,
  title     = {Same Trajectory, Contradictory Rewards (ROBORMBENCH): Paraphrase Fragility in Vision Language Reward Models},
  author    = {Wonje Jeung and Sangyeon Yoon and Hyesoo Hong and Yoonjun Cho and Dongjae Jeon and Bumjun Kim and Jean Oh and Youngjae Yu and Albert No},
  year      = {2026},
  eprint    = {2609.05401v1},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url       = {https://arxiv.org/abs/2609.05401v1}
}