Elamur Research⌕
☰
home / papers / 2026 / Seeing Isn't Believing: Mitigating Belief Inertia via Active Intervention in Embodied Agents

Seeing Isn't Believing: Mitigating Belief Inertia via Active Intervention in Embodied Agents

Paper · 2026 · Hanlin Wang, Chak Tou Leong, Jian Wang et al.

Summary

LLM-driven embodied agents often see relevant feedback but then act as if they had not, for example searching for an item already in view. The authors call this "observational neglect." On ALFWorld it is a dominant failure pattern and appears whether the agent is prompted, SFT-tuned or RL-tuned. To find its cause, they probe the agents' internal states with yes/no questions and read the first-token logits. The probes show that, after acting, agents form a confident but wrong expectation, and new observations barely shift it. The authors name this effect "belief inertia." Their remedy is Estimate-Verify-Update (EVU), which separates belief management from action selection by producing an explicit textual belief state. The agent predicts what its action should produce, checks that prediction against the actual observation, and revises its belief accordingly. EVU works with both prompting-based and training-based agents. Across three embodied benchmarks it consistently raises task success, and further analysis confirms that it reduces belief inertia. This points toward embodied agents that stay grounded in what they actually observe.

Key points

  • Scale of the failure: On ALFWorld, observational neglect occurs far more often in failed trajectories than in successful ones, by a factor of up to roughly 11×. It shows up in vanilla, SFT and RL agents.
  • Probing method:
    • The authors define a debiased "True Belief Value" from yes/no logit differences, subtracting a question-only baseline.
    • They apply it to 100 neglect cases from SFT and RL agents, measuring belief at three stages: current, estimated and subsequent.
    • Beliefs stay on the wrong side of the boundary even after the corrective observation arrives.
  • Oracle check: When the agent's belief is corrected externally (oracle belief intervention), success improves. This suggests the agents can reason well enough once the faulty belief is fixed.
  • Benchmarks and models:
    • EVU is evaluated on ALFWorld, ScienceWorld and VirtualHome (v2.3.0).
    • Prompting experiments mainly use DeepSeek V3.2; fine-tuning uses open-source LLMs.
  • Comparison with reflection:
    • EVU beats a Multiple Reflection baseline while using fewer reasoning tokens, so the gains are not just extra test-time compute.
    • The authors attribute this to EVU's predictive, mismatch-driven design.
    • Code is available at github.com/WangHanLinHenry/EVU.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Hong Kong Polytechnic UniversityHong Kong · China84 works · 4,956 citations🎓Sichuan UniversityChengdu · China14 works · 1,293 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksEngineering · 133 worksControl and Systems Engineering · 45 works

Cite

Cite as Hanlin Wang, Chak Tou Leong, Jian Wang and Wenjie Li. 2026. Seeing Isn't Believing: Mitigating Belief Inertia via Active Intervention in Embodied Agents. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1884

@inproceedings{wang2026seeing,
  title = {{Seeing Isn't Believing: Mitigating Belief Inertia via Active Intervention in Embodied Agents}},
  author = {Hanlin Wang, Chak Tou Leong, Jian Wang and Wenjie Li},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1884},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1884}
}

References

40 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Baian Chen, Chang Shu, Ehsan Shareghi, Nigel Collier, Karthik Narasimhan, and Shunyu Yao. 2023. Fireact: Toward language agent fine-tuning. arXiv preprint arXiv:2310.05915.
  2. Hanyang Chen, Mark Zhao, Rui Yang, Qinwei Ma, Ke Yang, Jiarui Yao, Kangrui Wang, Hao Bai, Zhenhailong Wang, Rui Pan, and 1 others. 2025. Era: Transforming vlms into embodied agents via embodied prior learning and online reinforcement learning. arXiv preprint arXiv:2510.12693. — [arXiv:2510.12693](https://arxiv.org/abs/2510.12693) find
  3. Xianzhe Fan, Xuhui Zhou, Chuanyang Jin, Kolby Nottingham, Hao Zhu, and Maarten Sap. 2025. Somi-tom: Evaluating multi-perspective theory of mind in embodied social interactions. arXiv preprint arXiv:2506.23046. — [arXiv:2506.23046](https://arxiv.org/abs/2506.23046) find
  4. Lang Feng, Zhenghai Xue, Tingcong Liu, and Bo An. 2025. Group-in-group policy optimization for llm agent training. arXiv preprint arXiv:2505.10978. find
  5. Pascale Fung, Yoram Bachrach, Asli Celikyilmaz, Kamalika Chaudhuri, Delong Chen, Willy Chung, Emmanuel Dupoux, Hongyu Gong, Hervé Jégou, Alessandro Lazaric, and 1 others. 2025. Embodied ai agents: Modeling the world. arXiv preprint arXiv:2506.22355. — [arXiv:2506.22355](https://arxiv.org/abs/2506.22355) find
  6. Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. 2022. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning, pages 9118–9147. PMLR.
  7. Wenlong Huang, Fei Xia, Dhruv Shah, Danny Driess, Andy Zeng, Yao Lu, Pete Florence, Igor Mordatch, Sergey Levine, Karol Hausman, and 1 others. 2023. Grounded decoding: Guiding text generation with grounded models for embodied agents. Advances in Neural Information Processing Systems, 36:59636– 59661… find
  8. Jeonghye Kim, Sojeong Rhee, Minbeom Kim, Dohyung Kim, Sangmook Lee, Youngchul Sung, and Kyomin Jung. 2025. Reflact: World-grounded decision making in llm agents via goal-state reflection. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 33421–33453. find
  9. Manling Li, Yunzhu Li, Jiayuan Mao, and Wenlong Huang. 2025. Foundation models meet embodied agents. In Proceedings of the 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 5: Tutorial Abstracts), pages… find
  10. Manling Li, Shiyu Zhao, Qineng Wang, Kangrui Wang, Yu Zhou, Sanjana Srivastava, Cem Gokmen, Tony Lee, Erran Li Li, Ruohan Zhang, and 1 others. 2024. Embodied agent interface: Benchmarking llms for embodied decision making. Advances in Neural Information Processing Systems, 37:100428–100534. find
  11. Yi Liao, Yu Gu, Yuan Sui, Zining Zhu, Yifan Lu, Guohua Tang, Zhongqian Sun, and Wei Yang. 2025. Think in games: Learning to reason in games via reinforcement learning with large language models. arXiv preprint arXiv:2508.21365. — [arXiv:2508.21365](https://arxiv.org/abs/2508.21365) find
  12. Mircea Lică, Ojas Shirekar, Baptiste Colle, and Chirag Raman. 2024. Mindforge: Empowering embodied agents with theory of mind for lifelong collaborative learning. arXiv preprint arXiv:2411.12977. — [arXiv:2411.12977](https://arxiv.org/abs/2411.12977) find
  13. Aly Lidayan, Jakob Bjorner, Satvik Golechha, Kartik Goyal, and Alane Suhr. 2025. Abbel: Llm agents acting through belief bottlenecks expressed in language. arXiv preprint arXiv:2512.20111. — [arXiv:2512.20111](https://arxiv.org/abs/2512.20111) find
  14. Xingyao Lin, Xinghao Zhu, Tianyi Lu, Sicheng Xie, Hui Zhang, Xipeng Qiu, Zuxuan Wu, and Yu-Gang Jiang. 2025. Ask-to-clarify: Resolving instruction ambiguity through multi-turn dialogue. arXiv preprint arXiv:2509.15061. — [arXiv:2509.15061](https://arxiv.org/abs/2509.15061) find
  15. Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, and 1 others. 2025. Deepseek-v3. 2: Pushing the frontier of open large language models. arXiv preprint arXiv:2512.02556.
  16. Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. 2025. Reasoning models can be effective without thinking. arXiv preprint arXiv:2504.09858. find
  17. Xavier Puig, Kevin Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. 2018. Virtualhome: Simulating household activities via programs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8494–8502. find
  18. Shuofei Qiao, Runnan Fang, Ningyu Zhang, Yuqi Zhu, Xiang Chen, Shumin Deng, Yong Jiang, Pengjun Xie, Fei Huang, and Huajun Chen. 2024. Agent planning with world knowledge model. Advances in Neural Information Processing Systems, 37:114843–114871. find
  19. Ram Ramrakhya, Matthew Chang, Xavier Puig, Ruta Desai, Zsolt Kira, and Roozbeh Mottaghi. 2025. Grounding multimodal llms to embodied agents that ask for help with reinforcement learning. arXiv preprint arXiv:2504.00907. — [arXiv:2504.00907](https://arxiv.org/abs/2504.00907) find
  20. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347. find
  21. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, and 1 others. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.
  22. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36:8634–8652.
  23. Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. 2020. Alfworld: Aligning text and embodied environments for interactive learning. arXiv preprint arXiv:2010.03768.
  24. Yifan Song, Da Yin, Xiang Yue, Jie Huang, Sujian Li, and Bill Yuchen Lin. 2024. Trial and error: Exploration-based trajectory optimization of llm agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7584–7600.
  25. Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291. find
  26. Hanlin Wang, Chak Tou Leong, Jian Wang, and Wenjie Li. 2024. E2cl: exploration-based error correction learning for embodied agents. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 7626–7639. find
  27. Hanlin Wang, Chak Tou Leong, Jiashuo Wang, Jian Wang, and Wenjie Li. 2025a. Spa-rl: Reinforcing llm agents via stepwise progress attribution. arXiv preprint arXiv:2505.20732. — [arXiv:2505.20732](https://arxiv.org/abs/2505.20732) find
  28. Hanlin Wang, Jian Wang, Chak Tou Leong, and Wenjie Li. 2025b. Steca: Step-level trajectory calibration for llm agent learning. In Findings of the Association for Computational Linguistics: ACL 2025, pages 11597– 11614. find
  29. Jiashuo Wang, Jiawen Duan, Jian Wang, Kaitao Song, Chunpu Xu, Johnny K. W. Ho, Fenggang Yu, Wenjie Li, and Johan F. Hoorn. 2026. Foresight optimization for strategic reasoning in large language models. Preprint, arXiv:2604.13592.
  30. Kangrui Wang, Pingyue Zhang, Zihan Wang, Yaning Gao, Linjie Li, Qineng Wang, Hanyang Chen, Chi Wan, Yiping Lu, Zhengyuan Yang, and 1 others. 2025c. Vagen: Reinforcing world model reasoning for multi-turn vlm agents. arXiv preprint arXiv:2510.16907. find
  31. Zijian Zhou, Ao Qu, Zhaoxuan Wu, Sunghwan Kim, Alok Prakash, Daniela Rus, Jinhua Zhao, Bryan Kian Hsiang Low, and Paul Pu Liang. 2025. Mem1: Learning to synergize memory and reasoning for efficient long-horizon agents. arXiv preprint arXiv:2506.15841. — [arXiv:2506.15841](https://arxiv.org/abs/2506.15841) find
  32. Ruoyao Wang, Peter Jansen, Marc-Alexandre Côté, and Prithviraj Ammanabrolu. 2022. Scienceworld: Is your agent smarter than a 5th grader? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11279–11298. find
  33. Di Wu, Jiaxin Fan, Junzhe Zang, Guanbo Wang, Wei Yin, Wenhao Li, and Bo Jin. 2025. Reinforced reasoning for embodied planning. arXiv preprint arXiv:2505.22050. — [arXiv:2505.22050](https://arxiv.org/abs/2505.22050) find
  34. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025a. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  35. Rui Yang, Hanyang Chen, Junyu Zhang, Mark Zhao, Cheng Qian, Kangrui Wang, Qineng Wang, Teja Venkat Koripella, Marziyeh Movahedi, Manling Li, and 1 others. 2025b. Embodiedbench: Comprehensive benchmarking multi-modal large language models for vision-driven embodied agents. In International Conference… find
  36. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809–11822. find
  37. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. In The eleventh international conference on learning representations.
  38. Wenqi Zhang, Ke Tang, Hai Wu, Mengna Wang, Yongliang Shen, Guiyang Hou, Zeqi Tan, Peng Li, Yueting Zhuang, and Weiming Lu. 2024. Agent-pro: Learning to evolve via policy-level reflection and optimization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Vol…
  39. Wenqi Zhang, Mengna Wang, Gangao Liu, Xu Huixin, Yiwei Jiang, Yongliang Shen, Guiyang Hou, Zhe Zheng, Hang Zhang, Xin Li, and 1 others. 2025. Embodied-reasoner: Synergizing visual search, reasoning, and action for embodied interactive tasks. arXiv preprint arXiv:2503.21696.
  40. Zirui Zhao, Wee Sun Lee, and David Hsu. 2023. Large language models as commonsense knowledge for large-scale task planning. Advances in neural information processing systems, 36:31967–31987. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2026-seeing-isn-t-believing-mitigating/