Pith. sign in

REVIEW 3 major objections 6 minor 59 references

Conditional Multi-Stage Failure Recovery for Embodied Agents

T0 review · 3 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 This paper claims that a conditional four-stage chain-prompting recovery layer raises task success on the TEACH TfD benchmark by 11.5 points over the same planner and executor, and that the recovery structure itself, not the plan…

desk verdict Solid controlled study of a structured failure-recovery chain for TEACH TfD; the gain over the no-recovery baseline is real, but the SOTA framing and the scene-representation blind spot need attention. read the letter →

arxiv 2507.06016 v1 pith:A4JZ5ABD submitted 2025-07-08 cs.CL

Youmna Farag , Svetlana Stoyanchev , Mohan Li , Simon Keizer , Rama Doddipatla This is my paper
classification cs.CL
keywords failurerecoveryembodiedagentschainpromptingzero-shotLLMplanningTEACHTfDbenchmarkobjectsearchscenerepresentation
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

Embodied agents that plan with large language models often fail mid-task, and this paper's claim is that a structured, conditional recovery process—not a better initial plan—is what converts failures into completed tasks. The authors wrap an existing GPT-4o planner and executor in a four-stage error-handling layer on the TEACH TfD benchmark of long-horizon household tasks: skip unimportant failing subgoals, generate missing preconditions, propose workaround actions, and reflect on the whole task after execution. With the same initial plan and same executor, the full system reaches 36.46% success on the seen split versus 24.86% without recovery, and 31.20% versus 22.05% on unseen episodes, which the paper presents as state of the art. Ablations show the precondition stage and the post-execution reflection drive most of the gain. If the claim is right, failure recovery is a generalizable LLM capability that can be bolted onto any embodied planner without retraining.

What carries the argument

The central object is the Conditional Multi-stage Failure Recovery (CMFR) chain: four zero-shot prompts chained so that each stage activates only if the previous stage says the failure matters. Stage 1 asks whether the failing subgoal is important, and skips it if not; stage 2 asks what preconditions are missing and generates corrective subgoals; stage 3 asks for a workaround that substitutes an alternative action or object; stage 4, after execution, asks what the plan missed at task level. The companion mechanism is the pruned scene representation, built from the agent's object memory by keeping only plan-relevant objects, their task-relevant properties, parent/child containment relations, and the object currently in hand. This representation is what lets each prompt inspect the environment within a small prompt budget.

What would settle it

Run two conditions on the seen split with the same GPT-4o initial plans, the same executor, and the same pruned scene representation: the four-stage CMFR chain, and a single zero-shot prompt that gets the task, plan, failing subgoal, and scene representation and is asked to "diagnose why this action failed, fix any missing preconditions, or propose a workaround." If the single prompt matches CMFR's 36.46% success rate, the staged conditional decomposition is not the cause of the gain. A complementary check: filter episodes whose failures are dominated by positioning or obstacles, and test whether recovery succeeds when the scene representation deliberately omits any spatial or positional information; the paper's own 10-of-30 stage-2 sample predicts those episodes should remain unrecoverable.

Watch

Extended reading notes

Core claim

The central discovery is that when an embodied agent fails, an LLM can recover more effectively if the question "what went wrong?" is replaced by a conditional chain of narrower questions. CMFR routes each failing subgoal through successive stages: is the subgoal important? If yes, are preconditions missing? If no precondition is evident, is there an alternative action that achieves the same effect? A fourth stage runs after the plan is finished and looks at the whole scene for steps the plan omitted, such as cleaning a mug before filling it. The paper's controlled evidence for this discovery is the comparison that holds the initial GPT-4o plan and the executor fixed, so the 11.5-point seen-split improvement must come from the recovery layer itself. The framework also includes an LLM-based object search that uses object locations extracted from the dialogue, reducing failures caused by objects not yet found.

Load-bearing premise

The recovery chain assumes the pruned scene representation—plan-relevant objects with their task-relevant properties, containment, and held object—is enough for the LLM to see why an action failed, so failures caused by the agent's position or by unseen obstacles are outside what the chain can diagnose.

Editorial extensions

If this is right

  • Failure recovery becomes a drop-in layer: any LLM planner can be wrapped with CMFR and gain success without changing the plan or executor.
  • Hard-coded precondition checkers can be replaced by zero-shot LLM reasoning, removing domain-specific engineering from embodied agents.
  • Post-execution reflection catches errors that online recovery cannot, implying agents should budget a final verification pass even after all planned steps succeed.
  • Filtering unimportant subgoals is not just an efficiency win: it reduces the fraction of games lost to the 30-failure limit from 8% to 6%.
  • Small fixed example sets (3–5 shots) push seen-split success to 38.12%, so the staged prompts tolerate light supervision.

Reading between the lines

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

  • The paper's own failure analysis isolates a class of failures—about a third of the sampled stage-2 cases—where the LLM identified correct preconditions but execution still failed for positioning or obstacle reasons invisible to the scene representation; enriching that representation with agent pose, reachability, and obstacle information is the natural next test.
  • The conditional gating pattern should transfer to other instruction-following benchmarks and physical robots, where the same "skip, repair, substitute, verify" hierarchy could be evaluated against monolithic replanning.
  • Because object-not-found is the most frequent failure reason (41.1%), the dialogue-grounded object search module is arguably as valuable as any single recovery stage; making location extraction more robust may yield larger gains than additional recovery stages.
  • The wide disagreement between LLMs on what counts as an important subgoal (GPT-4o marks 66% important, Qwen2.5 only 25%) suggests the framework's benefit depends on the model's calibration, so a cheap importance-scoring method could make small models competitive.
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

3 major / 6 minor

Summary. The paper introduces CMFR, a zero-shot chain-prompting framework for failure recovery in embodied agents operating on the TEACH TfD benchmark. The framework has four stages: subgoal importance filtering, precondition generation, workaround generation, and a post-execution reflection stage. The authors evaluate CMFR using GPT-4o, o3-mini, Qwen2.5-7B, and Llama-3.1-8B, all sharing the same initial GPT-4o plan and the same executor as a no-recovery baseline. They report a seen-split success rate of 36.46% for CMFR-GPT-4o versus 24.86% for the no-recovery baseline, an unseen-split rate of 31.20% versus 22.05%, and an ablation showing that removing stages 2 and 4 together reduces seen-split success to 25.96%, close to baseline. The paper claims state-of-the-art results on TEACH TfD.

Significance. If the results are robust, the paper makes a useful empirical contribution: it provides a modular, zero-shot recovery framework that does not require hard-coded preconditions or example memory, and it evaluates the framework under a controlled comparison (same plan, same executor, with and without recovery). The public code link, per-task results, multi-LLM evaluation, and manual failure analysis are strengths. The ablation design is informative, and the comparison against a CoT-based recovery baseline helps isolate the effect of the staged prompt structure. However, the main quantitative claims currently lack statistical support, and the scope of the scene representation places an acknowledged but important ceiling on the fraction of failures the method can recover, which should be reflected in the state-of-the-art framing.

major comments (3)
  1. [Section 5, Tables 1 and 2] All headline numbers are reported as single values without variance, number of runs, confidence intervals, or significance tests. On the seen split (181 episodes per Appendix A), differences such as 32.04% versus 33.70% in the ablation and 25.96% versus 24.86% for the w/o-stage-2-and-4 condition versus baseline may be within run-to-run noise. The claim that stages 2 and 4 jointly account for essentially all of the improvement over baseline is load-bearing and needs at least several repeated runs with reported mean and standard deviation, or bootstrap confidence intervals and a paired significance test over episodes.
  2. [Section 4.4 and Section 6.2] The scene representation stores only plan-relevant objects, their task-relevant properties, parent/child containment, and the held object; it omits the agent's position, object coordinates, and obstacles. The paper's own Section 6.2 analysis of 30 subgoals where stage 2 generated correct preconditions but execution still failed finds 10 cases with 'no apparent reason in scene representation', attributed to positioning or obstacles, and Table 3 attributes 32.2% of all subgoal failures to positioning. Because stages 2 and 3 can only reason over the provided representation, a substantial failure class is structurally outside the recovery mechanism. This is partly acknowledged in the Limitations section, but the claim that CMFR 'analyse[s] execution challenges within their environmental context' and the state-of-the-art framing should be qualified to state that the gain is demonstrated only for failures representable in the pruned scene, and the paper should discuss whether enriching the representation with spatial information would extend recovery.
  3. [Section 6, Table 1] The state-of-the-art comparison is confounded by executor and positioning-heuristic differences: the paper acknowledges that HELPER* and DANLI* use different heuristics, and the margin over HELPER* (36.46% versus 17.12%) is therefore not a clean measure of the recovery framework alone. The controlled baseline (same GPT-4o plan and same executor) is the appropriate primary evidence for the method's effect. In addition, the abstract's 'surpassing the strongest existing model by 19%' should be stated as an absolute difference in percentage points (19.34 percentage points over the reported HELPER* value), not as a relative percentage improvement, and the same clarification should be made for the 11.5% versus baseline (36.46% - 24.86% = 11.60 percentage points).
minor comments (6)
  1. [Abstract and Section 6] The numbers in the abstract (11.5% and 19%) are inconsistent with Table 1 (11.6 and 19.34 percentage points); please use consistent values and label them as percentage points.
  2. [Listings 3 and 5, Appendix G] There are typos in the prompts: 'impotant' should be 'important' in Listing 3, and 'aorkaround' should be 'workaround' in Listing 5.
  3. [Table 5, Appendix E] The scene representation example contains 'Bow_1' where the object should be 'Bowl_1'.
  4. [Section 3 and Appendix A] The metric is defined as Path Length Weighted (PLW) in Section 3 and Appendix A, but the tables use the header 'PWL'; please make the notation consistent. The PLW formula in Appendix A, 'Pm = m * L*/max(L* ˆL)', is missing the intended denominator notation and should read max(L*, L_hat) or equivalent.
  5. [Section 2] The statement that chain prompting 'has not been explored in the context of embodied AI' is too strong given existing memory-augmented and multi-step prompting systems; consider softening it to 'is underexplored' or narrowing the claim to the specific conditional staged structure.
  6. [Figure 2] The text gives the percentage of subgoals reaching stage 1 (19.5%) and stage 3 (1%), but not the percentages for stages 2 and 4; adding those values to the figure or caption would make the stage-by-stage analysis easier to verify.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the CMFR gain is established by a controlled comparison against a no-recovery baseline using identical plans and executor on the external TEACH benchmark.

full rationale

The paper makes no formal derivation; its central claim is an empirical evaluation on the TEACH TfD benchmark. The load-bearing comparison is the Table 1 row 'No Failure Recovery' (24.86% seen, 22.05% unseen) versus CMFR-GPT-4o (36.46%, 31.20%), and Section 5 states that 'All experiments use the same initial plans generated by GPT-4o ... allowing a focused assessment of error recovery.' Because the planner and executor are held fixed, the improvement is attributable to the CMFR stages, not to a fitted parameter later relabeled as a prediction. The ablations in Table 2 test each stage independently, and the comparisons to HELPER*, DANLI*, and MSI use published numbers or disclosed ground-truth-perception replications, so no self-citation chain carries the result. There are no relevant self-citations to the present authors' prior results that would constitute load-bearing support. The paper's own Section 6.2 admission that 10 of 30 stage-2-persistent failures show 'no apparent reason in scene representation' is a scope limitation on the method's ceiling, not a circular step: it does not make the reported success rate equivalent to its inputs. Similarly, the Limitations section's acknowledgment of simulator positioning issues qualifies generality but does not reduce any prediction to a fit. No step in the paper matches self-definition, fitted-input-as-prediction, self-citation load-bearing, imported uniqueness, ansatz-by-citation, or renaming of a known result. The evaluation is self-contained against an external benchmark, so the appropriate finding is no significant circularity with score 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

This empirical paper makes no formal derivation; the ledger records hand-chosen design parameters and the modeling assumptions on which the success-rate gain depends.

free parameters (3)
  • Number of few-shot demonstrations per CMFR stage = 3 (stage 1), 4 (stage 2), 5 (stage 3), 4 (stage 4)
    Fixed by hand; the ablation only compares zero-shot vs few-shot and does not test sensitivity to the number of examples.
  • Planning demonstration pool and retrieval count = 24 hand-written demonstrations; 3 retrieved by Sentence-BERT
    Hand-selected from training data; the retrieval count of 3 is fixed with no sensitivity analysis.
  • Positioning heuristic sequence = Six ordered adjustments (Rotate Left, Rotate Right, Look Up, Look Down, Forward, Backward)
    Hand-crafted action sequence for interaction failures; the paper reports positioning failures as a dominant failure mode, so the result depends on this heuristic.
assumptions (3)
  • domain assumption Ground-truth perception from the AI2-THOR simulator is used for object properties, positions, and containment (Section 4.2).
    Prompts assume accurate oracle state of relevant objects; real perception would introduce noise and likely reduce performance, as the authors acknowledge in Limitations.
  • domain assumption The textual scene representation, a pruned memory of relevant objects, properties, parent/child containment, and the held object, is sufficient for the LLM to diagnose and recover from execution failures (Section 4.4).
    This is load-bearing for stages 2 and 3; the paper's own analysis reports 10 of 30 persistent failures after stage 2 with 'no apparent reason in scene representation' (Section 6.2), so the premise is only partially satisfied.
  • domain assumption The LLM's judgments of subgoal importance, missing preconditions, and workarounds are sufficiently accurate to improve success (Section 4.5).
    The whole CMFR stack depends on the reasoning quality of GPT-4o/o3-mini; smaller LLMs show smaller gains, so capability is a load-bearing variable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Conditional Multi-Stage Failure Recovery for Embodied Agents." pith.science (2026). https://pith.science/paper/A4JZ5ABD

@misc{pith2026250706016,
  author       = {Pith},
  title        = {Pith review of: Conditional Multi-Stage Failure Recovery for Embodied Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A4JZ5ABD}},
  note         = {Machine review of arXiv:2507.06016}
}
read the original abstract

Embodied agents performing complex tasks are susceptible to execution failures, motivating the need for effective failure recovery mechanisms. In this work, we introduce a conditional multistage failure recovery framework that employs zero-shot chain prompting. The framework is structured into four error-handling stages, with three operating during task execution and one functioning as a post-execution reflection phase. Our approach utilises the reasoning capabilities of LLMs to analyse execution challenges within their environmental context and devise strategic solutions. We evaluate our method on the TfD benchmark of the TEACH dataset and achieve state-of-the-art performance, outperforming a baseline without error recovery by 11.5% and surpassing the strongest existing model by 19%.

Figures

Figures reproduced from arXiv: 2507.06016 by the authors.

Figure 1
Figure 1. System components: planning, the four CMFR stages, execution, object search and scene representation. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Subgoals that reach each stage of CMFR, during execution, from ‘All subgoals’ generated in the initial plan. Failure Reason Frequency Object not found 41.1% Positioning 32.2% pick-up an object while already holding another object 4.7% place an object while robot hand is empty 3.4% slice an object while not holding a knife 2.1% Others 16.5% [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Examples of RGB images from the agent’s view when it is failing to execute a subgoal. The failures are [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Subgoals that reach each stage of CMFR, during execution, from ‘All subgoals’ generated in the initial [PITH_FULL_IMAGE:figures/full_fig_p028_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 27 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Josh Abramson, Arun Ahuja, Federico Carnevale, Petko Georgiev, Alex Goldin, Alden Hung, Jessica Landon, Jirka Lhotka, Timothy Lillicrap, Alistair Muldal, et al. 2022. Improving multimodal interactive agents with reinforcement learning from human feedback. arXiv preprint arXiv:2211.11602

    arXiv 2022

  4. [4]

    Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Kua...

    arXiv 2022

  5. [5]

    Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko S \"u nderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. 2018. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3674--3683

    2018

  6. [6]

    Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias M \"u ller. 2023. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288

    arXiv 2023

  7. [7]

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alex Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuh...

    arXiv 2023

  8. [8]

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. 2022. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817

    arXiv 2022

Show all 59 references
  1. [9]

    Bin Dong, Fangao Zeng, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. 2021. Solq: Segmenting objects by learning queries. Advances in Neural Information Processing Systems, 34:21898--21909

    2021

  2. [10]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. https://openreview.net/forum?id=YicbFdNTTy An image is worth 1...

    2021

  3. [11]

    Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint...

    2023 arXiv

  4. [12]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  5. [13]

    Mauzama Firdaus, Gopendra Singh, Asif Ekbal, and Pushpak Bhattacharyya. 2023. Multi-step prompting for few-shot emotion-grounded conversations. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 3886--3891

    2023

  6. [14]

    Dayuan Fu, Biqing Qi, Yihuai Gao, Che Jiang, Guanting Dong, and Bowen Zhou. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.38 MSI -agent: Incorporating multi-scale insight into embodied agents for superior planning and decision-making . In Proceedings of the 2024 Conference...

    2024 doi

  7. [15]

    Xiaofeng Gao, Qiaozi Gao, Ran Gong, Kaixiang Lin, Govind Thattai, and Gaurav S Sukhatme. 2022. Dialfred: Dialogue-enabled agents for embodied instruction following. arXiv preprint arXiv:2202.13330

    2022 arXiv

  8. [16]

    Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. 2022 a . Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning, pages 9118--9147. PMLR

    2022

  9. [17]

    Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Noah Brown, Tomas Jackson, Linda Luu, Sergey Levine, Karol Hausman, and Brian Ichter. 2022 b . Inner monologue: Embodied rea...

    2022 arXiv

  10. [18]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    2024 arXiv

  11. [19]

    Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J. Davison. 2020. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters

    2020

  12. [20]

    Byeonghwi Kim, Jinyeon Kim, Yuyeong Kim, Cheolhong Min, and Jonghyun Choi. 2023. Context-aware planning and environment-aware memory for instruction following embodied agents. In ICCV

    2023

  13. [21]

    Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, et al. 2017. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474

    2017 arXiv

  14. [22]

    Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.356 Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding . In Proceedings of the 2020 Conference on Empirica...

    2020 doi

  15. [23]

    Alice Kwak, Clayton Morrison, Derek Bambauer, and Mihai Surdeanu. 2024. https://doi.org/10.18653/v1/2024.nllp-1.25 Classify first, and then extract: Prompt chaining technique for information extraction . In Proceedings of the Natural Legal Language Processing Workshop 2024, pa...

    2024 doi

  16. [24]

    Leonardo Lamanna, Luciano Serafini, Alessandro Saetti, Alfonso Gerevini, and Paolo Traverso. 2021. Online grounding of pddl domains by acting and sensing in unknown environments. arXiv preprint arXiv:2112.10007

    2021 arXiv

  17. [25]

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. https://doi.org/10.18653/v1/2020.acl-main.703 BART : Denoising sequence-to-sequence pre-training for natural language generation, translatio...

    2020 doi

  18. [26]

    Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Mart \' n-Mart \' n, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, et al. 2023 a . Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In...

    2023

  19. [27]

    Hao Li, Dingwen Zhang, Yalun Dai, Nian Liu, Lechao Cheng, Jingfeng Li, Jingdong Wang, and Junwei Han. 2024. Gp-nerf: Generalized perception nerf for context-aware 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page...

    2024

  20. [28]

    Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. 2023 b . Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pa...

    2023

  21. [29]

    Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. 2023. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493--9500. IEEE

    2023

  22. [30]

    Kevin Lin, Christopher Agia, Toki Migimatsu, Marco Pavone, and Jeannette Bohg. 2023. https://doi.org/10.1007/s10514-023-10131-7 Text2motion: from natural language instructions to feasible plans . Autonomous Robots

    2023 doi

  23. [31]

    Bo Liu, Yuqian Jiang, Xiaohan Zhang, Qiang Liu, Shiqi Zhang, Joydeep Biswas, and Peter Stone. 2023 a . Llm+p: Empowering large language models with optimal planning proficiency. arXiv preprint arXiv:2304.11477

    2023 arXiv

  24. [32]

    Zeyi Liu, Arpit Bahety, and Shuran Song. 2023 b . Reflect: Summarizing robot experiences for failure explanation and correction. arXiv preprint arXiv:2306.15724

    2023 arXiv

  25. [33]

    Subhankar Maity, Aniket Deroy, and Sudeshna Sarkar. 2024. A novel multi-stage prompting approach for language agnostic mcq generation using gpt. In European Conference on Information Retrieval, pages 268--277. Springer

    2024

  26. [34]

    Manolis Savva* , Abhishek Kadian* , Oleksandr Maksymets* , Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. 2019. Habitat: A P latform for E mbodied AI R esearch. In Proceedings of the IEEE/CVF Interna...

    2019

  27. [35]

    So Yeon Min, Devendra Singh Chaplot, Pradeep Ravikumar, Yonatan Bisk, and Ruslan Salakhutdinov. 2021. Film: Following instructions in language with modular methods. arXiv preprint arXiv:2110.07342

    2021 arXiv

  28. [36]

    Aishwarya Padmakumar, Mert Inan, Spandana Gella, Patrick Lange, and Dilek Hakkani-Tur. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.374 Multimodal embodied plan prediction augmented with synthetic embodied dialogue . In Proceedings of the 2023 Conference on Empirical Meth...

    2023 doi

  29. [37]

    Aishwarya Padmakumar, Jesse Thomason, Ayush Shrivastava, Patrick Lange, Anjali Narayan-Chen, Spandana Gella, Robinson Piramuthu, Gokhan Tur, and Dilek Hakkani-Tur. 2022. Teach: Task-driven embodied agents that chat. In Proceedings of the AAAI Conference on Artificial Intellige...

    2022

  30. [38]

    Alexander Pashevich, Cordelia Schmid, and Chen Sun. 2021. Episodic transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15942--15952

    2021

  31. [39]

    Daniel Philipov, Vardhan Dongre, gokhan tur, and Dilek Hakkani Tur. 2024. https://openreview.net/forum?id=xcok2UNQvB Simulating user agents for embodied conversational AI . In NeurIPS 2024 Workshop on Open-World Agents

    2024

  32. [40]

    Nils Reimers and Iryna Gurevych. 2019. https://arxiv.org/abs/1908.10084 Sentence-bert: Sentence embeddings using siamese bert-networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics

    2019 arXiv

  33. [41]

    Gabriel Sarch, Sahil Somani, Raghav Kapoor, Michael J Tarr, and Katerina Fragkiadaki. 2024. Helper-x: A unified instructable embodied agent to tackle four interactive vision-language domains with memory-augmented language models. In ICLR 2024 LLMAgents Workshop

    2024

  34. [42]

    Gabriel Sarch, Yue Wu, Michael Tarr, and Katerina Fragkiadaki. 2023. Open-ended instructable embodied agents with memory-augmented large language models. In Findings of the Association for Computational Linguistics: EMNLP 2023

    2023

  35. [43]

    Mohit Shridhar, Lucas Manuelli, and Dieter Fox. 2022. Perceiver-actor: A multi-task transformer for robotic manipulation. In Proceedings of the 6th Conference on Robot Learning (CoRL)

    2022

  36. [44]

    Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. 2020. https://arxiv.org/abs/1912.01734 ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks . In The IEEE Conference on Compute...

    2020 arXiv

  37. [45]

    Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. 2023. https://doi.org/10.1109/ICRA48891.2023.10161317 Progprompt: Generating situated robot task plans using large language models . In 2023 ...

    2023

  38. [46]

    Kunal Pratap Singh, Luca Weihs, Alvaro Herrasti, Jonghyun Choi, Aniruddha Kembhavi, and Roozbeh Mottaghi. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/file/674ad201bc8fa74b3c9979230aa0c63b-Paper-Conference.pdf Ask4help: Learning to leverage an expert for embodie...

    2022

  39. [47]

    Sadler, Wei-Lun Chao, and Yu Su

    Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M. Sadler, Wei-Lun Chao, and Yu Su. 2023. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    2023

  40. [48]

    Shichao Sun, Ruifeng Yuan, Ziqiang Cao, Wenjie Li, and Pengfei Liu. 2024. https://doi.org/10.18653/v1/2024.findings-acl.449 Prompt chaining or stepwise prompt? refinement in text summarization . In Findings of the Association for Computational Linguistics: ACL 2024, pages 7551...

    2024 doi

  41. [49]

    Dietrich Trautmann. 2023. Large language model prompt chaining for long legal document classification. arXiv preprint arXiv:2308.04138

    2023 arXiv

  42. [50]

    Zihao Wang, Shaofei Cai, Anji Liu, Yonggang Jin, Jinbing Hou, Bowei Zhang, Haowei Lin, Zhaofeng He, Zilong Zheng, Yaodong Yang, Xiaojian Ma, and Yitao Liang. 2023 a . Jarvis-1: Open-world multi-task agents with memory-augmented multimodal language models. arXiv preprint arXiv:...

    2023 arXiv

  43. [51]

    Zihao Wang, Shaofei Cai, Anji Liu, Xiaojian Ma, and Yitao Liang. 2023 b . Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents. arXiv preprint arXiv:2302.01560

    2023 arXiv

  44. [52]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824--24837

    2022

  45. [53]

    Tongshuang Wu, Ellen Jiang, Aaron Donsbach, Jeff Gray, Alejandra Molina, Michael Terry, and Carrie J Cai. 2022. Promptchainer: Chaining large language model prompts through visual programming. In CHI Conference on Human Factors in Computing Systems Extended Abstracts, pages 1--10

    2022

  46. [54]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Me...

    2024 arXiv

  47. [55]

    Haopeng Zhang, Xiao Liu, and Jiawei Zhang. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.714 S umm I t: Iterative text summarization via C hat GPT . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 10644--10657, Singapore. Association for...

    2023 doi

  48. [56]

    Yichi Zhang, Jianing Yang, Jiayi Pan, Shane Storks, Nikhil Devraj, Ziqiao Ma, Keunwoo Yu, Yuwei Bao, and Joyce Chai. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.83 DANLI : Deliberative agent for following natural language instructions . In Proceedings of the 2022 Confere...

    2022 doi

  49. [57]

    Zirui Zhao, Wee Sun Lee, and David Hsu. 2023. https://openreview.net/forum?id=Wjp1AYB8lH Large language models as commonsense knowledge for large-scale task planning . In Thirty-seventh Conference on Neural Information Processing Systems

    2023

  50. [58]

    Kaizhi Zheng, Xiaotong Chen, Odest Jenkins, and Xin Eric Wang. 2022 a . https://openreview.net/forum?id=NAYoSV3tk9 VLM bench: A compositional benchmark for vision-and-language manipulation . In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Bench...

    2022

  51. [59]

    Kaizhi Zheng, Kaiwen Zhou, Jing Gu, Yue Fan, Jialu Wang, Zonglin Di, Xuehai He, and Xin Eric Wang. 2022 b . Jarvis: A neuro-symbolic commonsense reasoning framework for conversational embodied agents. arXiv preprint arXiv:2208.13266

    2022 arXiv

Pith tools

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