Pith. sign in

REVIEW 2 major objections 6 minor 46 references

When Memory Lies: An Empirical Study of Spatial Memory Staleness in VLM Agents

T0 review · 2 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Stale memory can make a VLM navigator deadlier than no memory.

desk verdict A carefully controlled empirical study that deserves peer review, though the headline safety tax should be re-framed as a maximal-coverage result until a memory-coverage ablation is added. read the letter →

arxiv 2608.04574 v1 pith:WAGFGJOO submitted 2026-08-05 cs.CL

Yushi Sun , Yanjie Zhang This is my paper
classification cs.CL
keywords spatialmemorystalenessVLMagentsFrozenLakememory-observationconflictvisualgroundingread-timefilteringnavigationsafetymemory-augmented
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Memory-augmented VLM agents often act on persistent spatial claims that the world has silently made false. This paper tries to establish that spatial-memory staleness is a distinct, measurable safety failure mode: in its primary GPT-4o text setting, an agent that trusts raw stale memory died in 74.4% of episodes versus 28.0% for the same agent with no memory at all. It also argues that text solvability does not imply visual grounding, since models that flag stale entries from text range from 0.887 to 0.067 vision F1 on identical grids. A transparent read-time filter removes much of the text-mode safety tax, yet oracle stale labels bring no further significant gain, locating the remaining bottleneck in action selection over filtered memory rather than in stale-label accuracy. If these findings hold, memory-augmented agents need an explicit consistency gate on every spatial memory read, especially when observations arrive as images.

What carries the argument

The central machinery is SpatialSTALE, a dynamic FrozenLake testbed that pairs a staleness-detection task with a navigation task under three change regimes: L1 scattered pregame changes, L2 clustered pregame changes, and L3 online changes that arrive during the episode. The memory store holds one textual claim per cell, and the evaluation defines stale by ground-truth type comparison, so SAFE is stale exactly when the current cell is a hole. The four navigation strategies (NoMemory, NoFilter, SelfVerify, and OMCD) share the same per-step LLM navigator and differ only in how the memory view is presented; OMCD is a transparent batched read-time filter that removes detector-flagged entries and re-audits after realized events under L3. The Oracle ablation feeds ground-truth stale labels through the same filtering interface, and outcome-conditioned trajectory lengths distinguish 'rush-into-hole' stale-memory deaths from memoryless exploration failures.

What would settle it

Run the same NoMemory/NoFilter/OMCD contrast under partial observability with memory entries produced by the agent's own exploration rather than by a generator; if the NoFilter death rate no longer exceeds the NoMemory rate, or if filtering stops helping, the central safety tax is an artifact of the clean, fully observable setting. A more direct check: on a 16x16 dynamic grid, if oracle-label filtering still yields no success gain over learned filtering, the label-saturating claim holds; if it does yield a gain, the action-selection bottleneck is specific to 8x8.

Watch

Extended reading notes

Core claim

The paper claims that when an agent's stored map says SAFE at a cell the current observation shows as a hole, current VLMs often fail to resolve the conflict before acting, and the failure is safety-relevant. Across a dynamic 8x8 FrozenLake testbed with a synthetic one-entry-per-cell memory store, models that solve the staleness-detection task from text at near-ceiling accuracy collapse when the same grids are presented as images: vision F1 spans 0.887 (Qwen) down to 0.067 (GLM), with the weakest model issuing fluent, confident decisions that ignore the rendered image. On the downstream navigation task, consuming raw stale memory is deadlier than having no memory: GPT-4o L2 death rises from 28.0% to 74.4%. A controlled read-time filter (OMCD) reduces that death rate to 31.6%, but the improvement is conditional on reliable perception; when the visual audit is unreliable, filtering yields no consistent benefit. The authors' stated conclusion is that spatial-memory staleness is a safety failure mode, with reliable visual grounding and action selection under memory-observation conflict as the open challenges.

Load-bearing premise

The general conclusions assume that a fully observable 8x8 grid with a synthetically completed one-entry-per-cell textual memory store is a representative proxy for real memory-augmented agents; the authors state their numbers are lower bounds on a clean setting, not field estimates.

Editorial extensions

If this is right

  • Unfiltered persistent spatial memory should be treated as a liability: in the primary GPT-4o setting the NoFilter strategy is 2.7 times deadlier than NoMemory (74.4% vs 28.0% death on L2).
  • A text-level consistency gate can recover most of the safety tax, with OMCD reducing GPT-4o L2 death from 74.4% to 31.6% while raising success from 14.4% to 32.8%.
  • Text-mode staleness detection is near ceiling for capable models, but vision-mode auditing varies by over an order of magnitude on identical grids, so multimodal memory-augmented systems cannot assume transfer.
  • Oracle stale labels produce no detectable navigation gain over the learned text filter in the 8x8 setting, and per-seed detection F1 is uncorrelated with OMCD success, pointing to action selection over filtered memory as the residual bottleneck.
  • When visual auditing is unreliable, filtering provides no consistent downstream benefit, so the value of any read-time audit is conditional on perception quality.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the safety tax generalizes beyond the clean FrozenLake setting, memory-augmented deployments should pair every read of a persistent spatial claim with a cheap contradiction check; OMCD is a proof-of-concept gate, not the only possible design.
  • The paper's lower-bound framing implies that in real environments with partial observability, perception noise, or non-synthetic memory, the measured death-rate reversal could be larger or smaller; a direct test would run the same strategy contrast under partial observability.
  • Because detection F1 did not predict navigation success, evaluation suites for memory-augmented agents should report downstream outcome metrics, not just auditing accuracy, when they want to capture safety.
  • The vision collapse pattern in which a model verbally describes the image as matching the memory suggests models may resolve memory-observation conflict by trusting the textual prior; this is testable by varying the rendering format or inserting distractor memory entries, extensions the paper lists as planned.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. This paper introduces SpatialSTALE, a dynamic 8x8 FrozenLake benchmark in which a VLM agent holds a textual spatial memory of cell types that can go stale when the environment changes. The authors run two paired tasks: an entry-level staleness detection task (with matched text and vision observations) and a downstream navigation task comparing NoMemory, NoFilter, SelfVerify, and OMCD (a batched filtering intervention), plus an Oracle label ablation. On 1,800 detection runs and 12,000 text-mode navigation episodes with 50 shared seeds, they report three findings: text detection is near ceiling for capable models but vision transfer varies widely (Qwen 0.887 vs GLM 0.067 F1); consuming unfiltered stale memory is a safety liability (GPT-4o L2 death 74.4% vs 28.0% with no memory); and reliable text filtering removes much of the tax, with Oracle labels providing no detectable further gain, localizing residual failures to action selection.

Significance. The paper's contribution is a clean, reproducible measurement of a previously under-explored failure mode: persistent spatial memory that silently becomes invalid. The experimental design is strong in internal validity: shared seeds, paired Wilcoxon tests, outcome-conditioned trajectory analysis, verbatim failure traces, released code and data, and explicit exploratory labeling for the small vision-navigation sample. The GLM visual-audit collapse with confident but memory-dominated reasoning is a particularly compelling qualitative finding. The main gap is external validity: the synthetic completion of the memory store to one entry per cell makes the headline safety tax conditional on a worst-case exposure assumption, and the paper's 'lower bounds' framing is not fully justified. Still, as a mechanism-isolation study, the work is valuable and likely to influence follow-up work on memory auditing in agents.

major comments (2)
  1. [§3.2 and Appendix B] The memory store is synthetically completed to exactly 64 entries, so the NoFilter condition always contains a stale claim for every changed cell. The headline result in the Abstract and Table 3 (L2 GPT-4o: NoFilter death 74.4% vs NoMemory death 28.0%) is therefore measured against a memory store that maximizes exposure to stale claims. Real memory-augmented agents normally maintain partial, unevenly sampled episodic memories; with sparse coverage, many changed cells would carry no claim and the agent would fall back to the fully observable current grid, likely reducing the death gap. No coverage ablation is reported, and the Section 6 statement that the numbers are 'lower bounds on a clean setting' is not obviously the right direction: complete coverage is more naturally an upper bound on exposure. Please add an ablation that varies the fraction of cells with memory entries (e.g., the entries produced by the 20 random walks alone, or random subsets of the 64 entries) and show how the NoFilter-vs-NoMemory gap changes, or explicitly restrict the claimed safety tax to the complete-store condition.
  2. [§5.3 and Table 3] The NoFilter condition adds 64 memory lines on top of the full current-grid observation, while NoMemory has no memory block at all. The safety tax could therefore be inflated by prompt length and distraction rather than by the semantic content of stale claims. The SelfVerify and OMCD conditions remove flagged entries, so they change both content and the number of memory lines, and the design does not separate the two. A control that keeps the same 64 entries but marks them as 'possibly outdated' (or shuffles their order) would isolate the effect of a confident but wrong claim from the effect of reading 64 extra text lines. The trajectory analysis in Figure 6 supports the stale-claim mechanism, but the magnitude of the 2.7x death-rate increase is not fully identified without such a control.
minor comments (6)
  1. [Tables 2 and 3] The model is called 'Qwen3.6' in Tables 2 and 3 but 'Qwen3.6-Plus' in Section 5.1 and Figure 3; please make the naming consistent.
  2. [Section 4] The acronym OMCD is never expanded; please define it at first use.
  3. [Appendix K] The text refers to 'Finding 8', but no findings are numbered anywhere in the paper; either number the findings or remove the reference.
  4. [Table 1] In the L3 row, '2/event' under 'Requested Stale%' is cryptic; write '2 cells per event' for readability.
  5. [Section 5.2] The statement that the best and worst vision models differ by 'a factor of thirteen' would be more precise as 'thirteenfold' or with the exact ratio, since 0.887/0.067 is approximately 13.2.
  6. [Figure 3] Printing the per-model mean F1 values on the bars would help readers verify the reported delta-F1 values without reconstructing them from the figure.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the study is a controlled empirical benchmark; the only self-citation (Chao et al. 2026) is non-load-bearing.

full rationale

This paper is a controlled empirical study rather than a derivation chain. The headline quantities—text/vision detection F1, navigation success/death rates, the NoFilter vs NoMemory safety gap, and the OMCD filter effect—are measured from model outputs on 50 shared seeds, not algebraically forced by the benchmark definitions. Staleness itself is defined independently in Eq. (1) as a ground-truth type mismatch between the stored entry and the current grid cell, so the valid/stale annotation is not imported as an unexamined result from the self-cited STALE paper (Chao et al. 2026, an author-overlapping reference): the definition is self-contained and externally checkable. The memory-construction procedure (20 random walks plus synthetic completion of unvisited cells to N=64) is an explicit design choice to remove coverage variance, and the paper itself states 'This controlled completion is not a model prediction' (Section 3.2) and 'removes coverage variance as a confound' (Appendix B). This choice may make the measured safety tax an upper bound on stale-memory exposure rather than a field estimate, but that is a scope limitation the authors acknowledge in Section 6 ('lower bounds on a clean setting, not field estimates'), not a circular reduction: the death-rate gap is not equivalent by construction to the presence of stale entries, since navigators can ignore memory and Table 3 shows they often do. The vision-navigation results are flagged as a 10-seed exploratory preview. No fitted parameter is renamed as a prediction, no uniqueness theorem is invoked from prior work, and no result reduces to the self-citation. Therefore the circularity score is minimal, reflecting only the minor author-overlapping citation.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or theoretical entities; SpatialSTALE is a benchmark and OMCD is a filtering intervention, both are methodological artifacts rather than invented entities in the 'graviton' sense.

free parameters (2)
  • OMCD batch size B = 10
    Selected via ablation (Appendix E) as a cost-quality tradeoff: 9x fewer LLM calls than per-entry queries with only a 2.7pp F1 drop. It is a hand-chosen design parameter that affects the filtering intervention's performance, but it was not tuned to maximize navigation success.
  • Initial hole fraction = 0.25
    Benchmark construction choice; fixed across all instances to set task difficulty. Not fitted to outcomes, but affects the base difficulty of navigation and detection.
assumptions (5)
  • domain assumption Binary ground-truth staleness per entry as defined by Eq. (1) is the correct target; semantic ambiguity in 'safe'/'danger' claims is ignored.
    The evaluation reduces staleness to a comparison of stored vs current cell type; this is clean but may not capture graded or contextual staleness in real spatial claims.
  • domain assumption The current grid observation is always complete and truthful in both modalities.
    Full observability eliminates perception error as a confound, but real agents may have incomplete or noisy observations; the authors acknowledge this is a deliberate simplification (Appendix B).
  • domain assumption The 384x384 rendered image with coordinate labels is a fair visual condition for comparing VLMs.
    Qwen's high performance is used to argue the renderings are readable, but the format may still disadvantage models with weaker OCR of small coordinate labels; the encoding of coordinates inside each cell is a design choice that could interact with model capability.
  • standard math Standard statistical tools (paired Wilcoxon tests, Pearson/Spearman correlations) are appropriate for the 50-seed design.
    No unusual statistical assumptions; but multiple comparisons are not explicitly corrected, which is a minor concern.
  • domain assumption FrozenLake dynamics with absorbing hole states provide a valid operationalization of safety.
    Death by stepping into a hole is used as the safety metric; this is appropriate for the gridworld but may not capture continuous-domain safety costs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Memory Lies: An Empirical Study of Spatial Memory Staleness in VLM Agents." pith.science (2026). https://pith.science/paper/WAGFGJOO

@misc{pith2026260804574,
  author       = {Pith},
  title        = {Pith review of: When Memory Lies: An Empirical Study of Spatial Memory Staleness in VLM Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WAGFGJOO}},
  note         = {Machine review of arXiv:2608.04574}
}
read the original abstract

Memory-augmented VLM agents act on persistent spatial knowledge, yet that knowledge silently goes stale as the environment changes. We ask what happens when an agent must reconcile a confident memory claim with a contradicting observation, and whether current models can catch the conflict before it becomes a safety-relevant mistake. Using a dynamic FrozenLake testbed, we pair a staleness-detection task with a downstream navigation task across three closed-source models and three open-weight VLMs under both text and image inputs (1,800 detection runs, and 12,000 text-mode navigation episodes over four LLM navigators at a shared 50-seed scale). Three findings emerge. First, text solvability does not imply visual grounding: models that flag stale entries reliably from text nonetheless span vision F1 from 0.887 down to 0.067 on the identical grids, and the weakest keeps making fluent, confident decisions that ignore the image. Second, consuming stale memory without an audit is a safety liability: in our primary GPT-4o setting, an agent that trusts raw memory dies more than twice as often as the same agent given no memory at all. Third, auditing helps but does not close the gap: a transparent read-time filter removes much of the safety cost in text mode, yet even oracle stale labels bring no further significant gain on the current grid size, and when visual auditing is unreliable, filtering yields no consistent benefit. Together these results frame spatial-memory staleness as a safety failure mode and isolate reliable visual grounding and action selection under memory--observation conflict as the central open challenges for memory-augmented agents.

Figures

Figures reproduced from arXiv: 2608.04574 by the authors.

Figure 1
Figure 1. Schematic of the failure mechanism.Top: a per-step action policy, reasoning over stale memory, marches through (5, 6) after it has silently become a hole. Bottom: filtering the contradictory entry from the memory view enables a safe detour. The summary rates are from the GPT-4o navigation experiment over 50 L2 seeds, where OMCD reduces death from 74.4% to 31.6%. hierarchical trajectory memory; Reflexion (Shinn et al… view at source ↗
Figure 2
Figure 2. Illustrative regimes. L1 and L2 are observed once [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Rate at which each model flags a memory entry as [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Filtering intervention analysis (GPT-4o, text, suc [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Outcome-conditioned trajectory length across [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Per-run mean precision and recall for every (model, [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Per-entry recall on truly stale entries, split by change [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Per-seed detection F1 versus OMCD navigation [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 17 canonical work pages

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

Show all 46 references
  1. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  2. [10]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  3. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  4. [12]

    arXiv preprint arXiv:2305.16291 , year=

    Voyager: An Open-Ended Embodied Agent with Large Language Models , author=. arXiv preprint arXiv:2305.16291 , year=

    arXiv

  5. [13]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Reflexion: Language Agents with Verbal Reinforcement Learning , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  6. [14]

    arXiv preprint arXiv:2305.17144 , year=

    Ghost in the Minecraft: Generally Capable Agents for Open-World Environments via Large Language Models with Text-based Knowledge and Memory , author=. arXiv preprint arXiv:2305.17144 , year=

    arXiv

  7. [15]

    arXiv preprint arXiv:2310.08560 , year=

    MemGPT: Towards LLMs as Operating Systems , author=. arXiv preprint arXiv:2310.08560 , year=

    arXiv

  8. [16]

    arXiv preprint arXiv:2605.06527 , year=

    STALE: Can LLM Agents Know When Their Memories Are No Longer Valid? , author=. arXiv preprint arXiv:2605.06527 , year=

    arXiv

  9. [17]

    arXiv preprint arXiv:2310.03214 , year=

    FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation , author=. arXiv preprint arXiv:2310.03214 , year=

    arXiv

  10. [18]

    arXiv preprint arXiv:2406.13642 , year=

    SpatialBot: Precise Spatial Understanding with Vision Language Models , author=. arXiv preprint arXiv:2406.13642 , year=

    arXiv

  11. [19]

    Transactions of the Association for Computational Linguistics , volume=

    Visual Spatial Reasoning , author=. Transactions of the Association for Computational Linguistics , volume=

  12. [20]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    What's ``Up'' with Vision-Language Models? Investigating Their Struggle with Spatial Reasoning , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    2023

  13. [21]

    ISPRS Journal of Photogrammetry and Remote Sensing , volume=

    Change Detection from Remotely Sensed Images: From Pixel-Based to Object-Based Approaches , author=. ISPRS Journal of Photogrammetry and Remote Sensing , volume=

  14. [22]

    Guan, Tianrui and Liu, Fuxiao and Wu, Xiyang and Xian, Ruiqi and Li, Zongxia and Liu, Xiaoyu and Wang, Xijun and Chen, Lichang and Huang, Furong and Yacoob, Yaser and Manocha, Dinesh and Zhou, Tianyi , booktitle=

  15. [23]

    International Conference on Learning Representations (ICLR) , year=

    Measuring Massive Multitask Language Understanding , author=. International Conference on Learning Representations (ICLR) , year=

  16. [24]

    arXiv preprint arXiv:2211.09110 , year=

    Holistic Evaluation of Language Models , author=. arXiv preprint arXiv:2211.09110 , year=

    arXiv

  17. [25]

    Laird, John E. , year=. The

  18. [26]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Cognitive Mapping and Planning for Visual Navigation , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  19. [27]

    Savva, Manolis and Kadian, Abhishek and Maksymets, Oleksandr and Zhao, Yili and Wijmans, Erik and Jain, Bhavana and Straub, Julian and Liu, Jia and Koltun, Vladlen and Malik, Jitendra and Parikh, Devi and Batra, Dhruv , booktitle=

  20. [28]

    Minimalistic Gridworld Environment for

    Chevalier-Boisvert, Maxime and Willems, Lucas and Pal, Suman , year=. Minimalistic Gridworld Environment for

  21. [29]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Is a Picture Worth a Thousand Words? Delving into Spatial Reasoning for Vision Language Models , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  22. [30]

    Chen, Boyuan and Xu, Zhuo and Kirmani, Sean and Ichter, Brian and Driess, Danny and Florence, Pete and Sadigh, Dorsa and Guibas, Leonidas and Xia, Fei , journal=

  23. [31]

    Cheng, An-Chieh and Yin, Hongxu and Fu, Yang and Guo, Qiushan and Yang, Ruihan and Kautz, Jan and Wang, Xiaolong and Liu, Sifei , booktitle=

  24. [32]

    Wu, Qiucheng and Zhao, Handong and Saxon, Michael and Bui, Trung and Wang, William Yang and Zhang, Yang and Chang, Shiyu , journal=

  25. [33]

    and Ma, Wei-Chiu and Krishna, Ranjay , booktitle=

    Fu, Xingyu and Hu, Yushi and Li, Bangzheng and Feng, Yu and Wang, Haoyu and Lin, Xudong and Roth, Dan and Smith, Noah A. and Ma, Wei-Chiu and Krishna, Ranjay , booktitle=

  26. [34]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    Evaluating Object Hallucination in Large Vision-Language Models , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    2023

  27. [35]

    Mousavi, Seyed Mahed and Alghisi, Simone and Riccardi, Giuseppe , booktitle=

  28. [36]

    Locating and Editing Factual Associations in

    Meng, Kevin and Bau, David and Andonian, Alex and Belinkov, Yonatan , booktitle=. Locating and Editing Factual Associations in

  29. [37]

    International Conference on Learning Representations (ICLR) , year=

    Mass-Editing Memory in a Transformer , author=. International Conference on Learning Representations (ICLR) , year=

  30. [38]

    Ahn, Michael and Brohan, Anthony and Brown, Noah and Chebotar, Yevgen and Cortes, Omar and David, Byron and Finn, Chelsea and Fu, Chuyuan and Gopalakrishnan, Keerthana and Hausman, Karol and others , booktitle=. Do As

  31. [39]

    and Chao, Wei-Lun and Su, Yu , booktitle=

    Song, Chan Hee and Wu, Jiaman and Washington, Clayton and Sadler, Brian M. and Chao, Wei-Lun and Su, Yu , booktitle=

  32. [40]

    Conference on Robot Learning (CoRL) , year=

    Inner Monologue: Embodied Reasoning through Planning with Language Models , author=. Conference on Robot Learning (CoRL) , year=

  33. [41]

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle=

  34. [42]

    Madaan, Aman and Tandon, Niket and Gupta, Prakhar and Hallinan, Skyler and Gao, Luyu and Wiegreffe, Sarah and Alon, Uri and Dziri, Nouha and Prabhumoye, Shrimai and Yang, Yiming and others , booktitle=

  35. [43]

    arXiv preprint arXiv:2203.12667 , year=

    Vision-and-Language Navigation: A Survey of Tasks, Methods, and Future Directions , author=. arXiv preprint arXiv:2203.12667 , year=

    arXiv

  36. [44]

    Chevalier-Boisvert, Maxime and Bahdanau, Dzmitry and Lahlou, Salem and Willems, Lucas and Saharia, Chitwan and Nguyen, Thien Huu and Bengio, Yoshua , booktitle=

  37. [45]

    arXiv preprint arXiv:2407.17032 , year=

    Gymnasium: A Standard Interface for Reinforcement Learning Environments , author=. arXiv preprint arXiv:2407.17032 , year=

    arXiv

  38. [46]

    and Everitt, Tom and Lefrancq, Andrew and Orseau, Laurent and Legg, Shane , journal=

    Leike, Jan and Martic, Miljan and Krakovna, Victoria and Ortega, Pedro A. and Everitt, Tom and Lefrancq, Andrew and Orseau, Laurent and Legg, Shane , journal=

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.