Pith. sign in

REVIEW 5 major objections 4 minor 25 references

SE-VLN: A Self-Evolving Vision-Language Navigation Framework Based on Multimodal Large Language Models

T0 review · 5 major / 4 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 SE-VLN is a training-free vision-language navigation framework that claims the first self-evolving multimodal-LLM agent, lifting success rates on unseen environments to 57% (R2R) and 35.2% (REVERIE) by storing, retrieving, and correcting…

desk verdict A clearly built memory/RAG/reflection VLN agent whose headline self-evolution claim is undercut by a missing protocol and an overclaimed novelty. read the letter →

arxiv 2507.13152 v3 pith:FNJF4RWH submitted 2025-07-17 cs.CV cs.AIcs.RO

Xiangyu Dong , Haoran Zhao , Jiang Gao , Haozhou Li , Xiaoguang Ma , Yaoming Zhou , Fuhai Chen , Juan Liu This is my paper
classification cs.CVcs.AIcs.RO
keywords vision-languagenavigationself-evolutionmultimodallargelanguagemodelsretrieval-augmentedgenerationchain-of-thoughtreasoningreflectionmoduleexperiencerepositoryverbaltopologicalmap
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

The paper sets out to show that a vision-language navigation agent can improve its own performance continuously during testing, without any network training, by converting successful and failed routes into reusable experience. It claims this is the first multimodal-LLM-powered self-evolving framework for VLN, and reports that it reaches 57% success on unseen R2R environments and 35.2% on unseen REVERIE environments, outperforming the prior state-of-the-art MapGPT by 23.9% and 15.0% respectively. The central design is a closed loop: the agent records its observations and decisions in a verbal topological map, retrieves similar past experience to guide multi-step reasoning, then reflects on task outcomes to correct its mistakes and store the corrected decisions for future tasks. A sympathetic reader would care because the framework points to a way around the standard data-hungry, fine-tuning approach to embodied AI: an agent that learns from its own experience at inference time.

What carries the argument

The load-bearing mechanism is a three-module closed loop. A hierarchical memory module maintains a verbal topological map as short-term memory—a dynamically updated graph whose nodes are annotated with textual scene descriptions and the agent's thinking, planning, and executing decisions—alongside a long-term experience repository of quadruples (landmark features, scene descriptions, decision processes, revised decisions). A retrieval-augmented thought-based reasoning module extracts landmark features from the instruction, encodes them with Sentence-BERT, retrieves the most similar experiences as few-shot prompts, and uses chain-of-thought prompting to produce step-by-step decisions. A reflection module computes outcome metrics (navigation error, success rate, SPL, oracle success rate) from simulator ground truth, then has the MLLM identify the first unreasonable decision and produce a corrected decision process, which is stored back into the repository. That closed loop is what turns past successes and failures into future decision guidance.

What would settle it

Take the validation-unseen split, run the agent with an empty repository, then refill the repository with the exact trajectories of the episodes being scored (or with randomly shuffled versions of those episodes), and measure success rate; if performance rises with repository size even when the stored experiences are the same episodes being tested, the claimed self-evolution is an artifact of information leakage rather than transferable skill acquisition.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that coupling a hierarchical memory (a dynamically annotated verbal topological map plus a long-term experience repository), retrieval-augmented chain-of-thought reasoning, and a reflection module enables a multimodal large language model to navigate unseen indoor environments better than any prior training-free LLM-based VLN method, and to keep improving as the experience repository grows from 0 to 50 entries. The reported numbers on the R2R validation-unseen split (SR 57%, SPL 50%) and the REVERIE validation-unseen split (SR 35.2%, SPL 24.8%) exceed the previous best training-free method, and the ablation shows that removing the chain-of-thought decider or the outcome evaluator drops performance substantially. The paper interprets the monotonic improvement with repository size as evidence of continual self-evolution without any gradient updates.

Load-bearing premise

The self-evolution result rests on the assumption that the experience entries used in the 0/10/30/50-entry experiment come from episodes separate from the ones being scored; if the repository is populated from the same validation episodes that are measured, the improvement would reflect memorization rather than transferable learning.

Editorial extensions

If this is right

  • If the self-evolution claim holds, embodied agents can improve at deployment time from their own experience, removing the need for large-scale annotated navigation data and fine-tuning.
  • The framework suggests that a fixed, frozen multimodal LLM can serve as a continual learner when paired with an external memory and a reflection signal.
  • The reported gains over MapGPT indicate that experience reuse, not just map-guided prompting, is a major driver of LLM-based VLN performance.
  • The ablation shows that chain-of-thought reasoning and the ground-truth-based outcome evaluator are both necessary components, so future work should treat them as essential rather than optional.

Reading between the lines

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

  • The paper leaves open whether the same self-evolution would hold if the outcome evaluator used only the agent's own observations or human feedback instead of simulator ground truth; a natural extension would replace the ground-truth metrics with a learned or interactive evaluator and test whether the loop still improves.
  • The framework's reliance on a fixed landmark vocabulary for retrieval may limit transfer to novel object types; a testable extension would replace the predefined vocabulary with an open-vocabulary landmark extractor and measure whether retrieval quality and navigation success are maintained.
  • The flattening of the evolution curve at 50 entries hints that the bottleneck is the LLM's context window and reasoning capacity, not the memory; an extension would be to compress or abstract stored experiences (e.g., summarizing similar experiences) before retrieval.
  • If the self-evolution result is real, it suggests that the key to long-lived agents is not bigger models but a closed loop of recording, retrieval, and correction; this could transfer to other embodied tasks such as manipulation or indoor exploration.
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

5 major / 4 minor

Summary. SE-VLN is a training-free vision-language navigation framework that couples a hierarchical memory module (a verbal topological map plus a long-term experience repository stored in Chroma), a retrieval-augmented thought-based reasoning module for multi-step decision-making, and a reflection module that computes success metrics from ground-truth trajectories and uses an MLLM to generate corrected decisions that are added back to the experience repository. The authors report that SE-VLN with GPT-4o reaches 57% SR and 50% SPL on the R2R validation-unseen set and 35.2% SR on REVERIE, improving over MapGPT (GPT-4o), and that performance rises as the repository grows from 0 to 50 entries, which they take as evidence of continual self-evolution during testing.

Significance. The proposed architecture is clearly described and the idea of closing the loop through retrieval, reflection, and repository growth is timely. The paper includes detailed prompts and pipeline figures that make the framework easy to understand and replicate at a high level. The strongest positive aspects are the clean modular decomposition and the explicit attempt to turn failure cases into reusable corrective experience, which is a promising direction for LLM-based VLN. However, the evaluation as presented does not yet substantiate the central claims: the self-evolution result is under-specified about the provenance of experience entries, no comparison is made to the closely related EvolveNav work that is cited in the bibliography, and all results are single-run numbers without variance or significance tests. If the framework is confirmed with a proper disjoint-split protocol and a head-to-head comparison, it would be a useful contribution to the training-free VLN literature, but at present the empirical evidence falls short of the stated claims.

major comments (5)
  1. [§Evaluation of Self-Evolution (Fig. 4)] The paper does not specify whether the 0/10/30/50 experience entries used in Fig. 4 are generated from the same validation episodes on which SR and SPL are computed. Since the reflection module (Eqs. 11–13) stores corrected decisions derived from the ground-truth trajectory for each task, and the retriever (Eqs. 5–8) selects experiences by semantic similarity to the current instruction, entries originating from the scored episodes could inject the correct outcome for those very episodes into the prompt. Under that protocol, the observed improvement with repository size would reflect test-set memorization rather than transferable self-evolution. The authors should provide a disjoint-split protocol (e.g., experience entries built only from episodes outside the evaluation set), release episode IDs, and report a control that removes same-episode entries; without this, the central self-evolution claim is unverified.
  2. [Related Work; Contributions] EvolveNav (Lin et al. 2025, arXiv:2506.01551) is listed in the references but is neither discussed in the Related Work section nor compared experimentally, despite proposing self-improving embodied reasoning for LLM-based VLN and appearing before this submission. The manuscript's claim to be 'the first time that an multimodal LLM-powered self-evolving VLN framework was proposed' is therefore not supportable as stated. The authors should revise the novelty claim and add a direct comparison or at least a substantive discussion explaining the differences from EvolveNav.
  3. [Tables 2–4; Implementation Details] All reported results are single-run numbers for one backbone (GPT-4o), and no standard deviations, confidence intervals, or statistical tests are given. MLLM-based inference is stochastic (sampling temperature, prompt order), so the differences of a few percentage points in Table 3 (e.g., SR 46→57 vs. SPL 34→50) cannot be assessed without repeated runs. Please report the number of runs, random seeds, decoding parameters, and error bars, and state whether the improvements over MapGPT are statistically significant.
  4. [Ablation Study; Fig. 2; Tables 2–3] The main experiments do not state the number of retrieved experiences used in the final evaluation. Figure 2 shows that 0-, 2-, and 5-shot retrieval are compared, with 2-shot best, but the repository-size experiment in Fig. 4 is reported without specifying which shot setting is active for each point. In addition, the text describes Table 3 as a 'larger validation set' than the 72-scene ablation, but Table 3 uses fewer scenes (11) and more trajectories (783); these are different test distributions, so the sentence comparing the two sets is misleading. Please specify the shot count and clarify the relationship between the two evaluation sets.
  5. [A Reflection Module (Eq. 11)] The outcome evaluator computes metrics from the ground-truth trajectory tau_gt provided by the simulator, meaning the reflection process uses privileged information that would not be available in a deployed system. The paper should explicitly state this as a limitation of the self-evolution experiments and, if the framework is meant to be training-free, explain how the 'real world' human-expert feedback mentioned in the text would be obtained and whether any experiments avoid ground-truth supervision.
minor comments (4)
  1. [Abstract] 'REVERSE' should be 'REVERIE'; the phrase 'absolute performance improvements of 23.9% and 15.0%' is inaccurate because these are relative improvements (11.0 and 4.6 percentage points absolute).
  2. [Methodology, Eq. (4)] The notation 'H(Pfull) ∝ O(N · d)' is non-standard; the right-hand side should be O(N · d) directly, and the limit statement should be reworded.
  3. [Related Work] EvolveNav is in the bibliography but is never mentioned in the text; at minimum, add a sentence positioning SE-VLN relative to it.
  4. [Throughout] The manuscript contains several grammatical errors, including 'an multimodal' in the abstract and 'This paper designed' in the Introduction; a proofreading pass is recommended.

Circularity Check

1 steps flagged · score 6.0 of 10

The self-evolution evidence is partly circular: the stored 'experience' is generated from ground-truth metrics of the very navigation tasks being scored, and Fig. 4 never excludes same-episode entries, so the improvement can reduce to test-set memorization by construction.

  1. self definitional [Methodology, A Reflection Module (Eqs. 11-13) and Evaluation of Self-Evolution (Fig. 4)]
    "we utilized ground truth data from the MatterPort3D simulator to accurately compute the outcomes of the current navigation tasks. ... Emetric = [NE, SR, SPL, OSR] = feval(τnav, τgt) ... ⟨T ′, P ′, A′⟩ = MLLM (Sref, Ψ(Mt), Emetric) ... EDB ← EDB ∪ enew ... As shown in Fig.4, we examined the evolution ability of the VLN framework by adjusting the number of experience entries (set to 0, 10, 30, and 50)."

    The experience entry stored in EDB is defined via Eq. 11-13 from the ground-truth outcome of 'the current navigation tasks' (τgt). At inference, Eqs. 5-8 retrieve top-N experiences by landmark similarity and inject them into the prompt. The paper never states that the 0/10/30/50 entries in Fig. 4 come from episodes disjoint from the scored validation episodes. If an episode's own corrected decision is in EDB, the prompt for that episode contains the correct outcome and revised actions derived from that episode's ground truth, so the measured SR/SPL improvement as EDB grows is partly the re-injection of the target answer, not self-evolution. Table 3's own explanation that the validation gain 'might be attributed to ...

full rationale

The paper's central 'self-evolving' claim rests on Fig. 4, which shows SR/SPL improving as the experience repository grows from 0 to 50 entries. The repository is populated by a reflection module that uses ground-truth simulator metrics (Eq. 11) to produce corrected decisions (Eq. 12) and stores them as experience (Eq. 13). The retriever then injects those entries into the prompt (Eq. 8). Because the paper does not specify whether the entries are drawn from the same validation episodes being scored, the improvement curve is consistent with test-set memorization: the agent is shown the correct decision for the very task it is about to solve. This is a by-construction reduction for overlapping episodes, not a demonstrated transfer of knowledge. The framework's other components (verbal topological map, RAG+CoT reasoning) are self-contained and not circular, and there are no load-bearing self-citations. The R2R/REVERIE SOTA comparison would be valid if the repository were built from a disjoint source, but the paper does not report such a protocol, and its own Table 3 discussion suggests repository overlap with the validation scenes. Score 6 reflects partial circularity in the central evolution evidence.

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

The framework contributes an architecture rather than a derivation. The central evaluation rests on three hand-selected parameters, the choice of GPT-4o, and assumptions about transferability of experience, availability of ground-truth feedback, and fidelity of verbalized observations.

free parameters (3)
  • top_k_retrieved_experiences = 2
    The ablation in Fig.2 shows 2-shot retrieval is best and 5-shot hurts; this k is chosen on the 72-scene R2R subset and used in main results.
  • experience_repository_size_schedule = 0, 10, 30, 50
    Self-evolution evaluation in Fig.4 varies the number of seeded experience entries; the values and the source of entries are not justified.
  • backbone_MLLM = GPT-4o
    Five MLLMs were tested and GPT-4o selected as best; only GPT-4o results are reported in Tables 2 through 4, which is a post hoc model selection.
assumptions (3)
  • domain assumption Historical experience retrieved by landmark similarity transfers to new, unseen indoor environments.
    The retrieval module (Eq.5 through Eq.8) assumes decision sequences under the same landmark features are reusable across different buildings; no experiment isolates cross-environment transfer.
  • domain assumption Ground-truth navigation metrics from the simulator are available to drive reflection at test time.
    The outcome evaluator uses NE, SR, SPL, and OSR computed with ground-truth paths (Eq.11); the paper notes real deployment would need human experts, so the experimental loop is not fully self-contained.
  • domain assumption Verbal scene descriptions produced by the MLLM preserve enough visual detail for reliable navigation decisions.
    The verbal topological map replaces raw visual observations with text (Eq.1 and Eq.2); no analysis shows that information loss is acceptable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SE-VLN: A Self-Evolving Vision-Language Navigation Framework Based on Multimodal Large Language Models." pith.science (2026). https://pith.science/paper/FNJF4RWH

@misc{pith2026250713152,
  author       = {Pith},
  title        = {Pith review of: SE-VLN: A Self-Evolving Vision-Language Navigation Framework Based on Multimodal Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FNJF4RWH}},
  note         = {Machine review of arXiv:2507.13152}
}
read the original abstract

Recent advances in vision-language navigation (VLN) were mainly attributed to emerging large language models (LLMs). These methods exhibited excellent generalization capabilities in instruction understanding and task reasoning. However, they were constrained by the fixed knowledge bases and reasoning abilities of LLMs, preventing fully incorporating experiential knowledge and thus resulting in a lack of efficient evolutionary capacity. To address this, we drew inspiration from the evolution capabilities of natural agents, and proposed a self-evolving VLN framework (SE-VLN) to endow VLN agents with the ability to continuously evolve during testing. To the best of our knowledge, it was the first time that an multimodal LLM-powered self-evolving VLN framework was proposed. Specifically, SE-VLN comprised three core modules, i.e., a hierarchical memory module to transfer successful and failure cases into reusable knowledge, a retrieval-augmented thought-based reasoning module to retrieve experience and enable multi-step decision-making, and a reflection module to realize continual evolution. Comprehensive tests illustrated that the SE-VLN achieved navigation success rates of 57% and 35.2% in unseen environments, representing absolute performance improvements of 23.9% and 15.0% over current state-of-the-art methods on R2R and REVERSE datasets, respectively. Moreover, the SE-VLN showed performance improvement with increasing experience repository, elucidating its great potential as a self-evolving agent framework for VLN.

Figures

Figures reproduced from arXiv: 2507.13152 by the authors.

Figure 1
Figure 1. The workflow of SE-VLN, which was driven by multimodal large language models (MLLMs), encompassed three [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The impact of different shots experience. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Impact of outcome evaluator on navigation perfor [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The evaluation of self-evolution ability. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Performance of Different MLLMs on the R2R Sampling Subsets. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: The pipeline of the verbal topological map. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: The pipeline of the retrieval-augmented thought-based reasoning module. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: The reflection module workflow. The agent computed the navigation performance of current task, i.e., score list, [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: The prompts for the retrieval-augmented thought-based reasoning module consisted of two parts: the prefix prompt [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: A successful case of SE-VLN which by referencing past experience, effectively avoided potential collisions (red [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 10 canonical work pages

  1. [1]

    Anderson, P.; Wu, Q.; Teney, D.; Bruce, J.; Johnson, M.; S \"u nderhauf, N.; Reid, I.; Gould, S.; and Van Den Hengel, A. 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, 3674--3683

    2018

  2. [2]

    Chen, J.; Lin, B.; Xu, R.; Chai, Z.; Liang, X.; and Wong, K.-Y. 2024 a . Mapgpt: Map-guided prompting with adaptive path planning for vision-and-language navigation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 9796--9810

    work page 2024

  3. [3]

    Chen, Q.; Pitawela, D.; Zhao, C.; Zhou, G.; Chen, H.-T.; and Wu, Q. 2024 b . Webvln: Vision-and-language navigation on websites. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 1165--1173

    work page 2024

  4. [4]

    Chen, S.; Guhur, P.-L.; Tapaswi, M.; Schmid, C.; and Laptev, I. 2022. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 16537--16547

    work page 2022

  5. [5]

    Fried, D.; Hu, R.; Cirik, V.; Rohrbach, A.; Andreas, J.; Morency, L.-P.; Berg-Kirkpatrick, T.; Saenko, K.; Klein, D.; and Darrell, T. 2018. Speaker-follower models for vision-and-language navigation. Advances in neural information processing systems, 31

    work page 2018

  6. [6]

    V.; Bell, P.; and Storkey, A

    Han, D.; McInroe, T.; Jelley, A.; Albrecht, S. V.; Bell, P.; and Storkey, A. 2024. LLM-Personalize: Aligning LLM Planners with Human Preferences via Reinforced Self-Training for Housekeeping Robots. arXiv preprint arXiv:2404.14285

    arXiv 2024

  7. [7]

    Hong, Y.; Wu, Q.; Qi, Y.; Rodriguez-Opazo, C.; and Gould, S. 2021. Vln bert: A recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, 1643--1653

    2021

  8. [8]

    Lan, Y.; Hu, Z.; Wang, L.; Wang, Y.; Ye, D.; Zhao, P.; Lim, E.-P.; Xiong, H.; and Wang, H. 2023. Llm-based agent society investigation: Collaboration and confrontation in avalon gameplay. arXiv preprint arXiv:2310.14985

    arXiv 2023

Show all 25 references
  1. [9]

    Li, D.; Chen, W.; and Lin, X. 2024. TINA: Think, Interaction, and Action Framework for Zero-Shot Vision Language Navigation. arXiv preprint arXiv:2403.08833

    2024 arXiv

  2. [10]

    Li, J.; Padmakumar, A.; Sukhatme, G.; and Bansal, M. 2024 a . Vln-video: Utilizing driving videos for outdoor vision-and-language navigation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 18517--18526

    2024

  3. [11]

    Li, Z.; Lv, Y.; Tu, Z.; Shang, D.; and Qiao, H. 2024 b . Vision-language navigation with continual learning. arXiv preprint arXiv:2409.02561

    2024 arXiv

  4. [12]

    Lin, B.; Nie, Y.; Wei, Z.; Chen, J.; Ma, S.; Han, J.; Xu, H.; Chang, X.; and Liang, X. 2024. NavCoT: Boosting LLM-Based Vision-and-Language Navigation via Learning Disentangled Reasoning. arXiv preprint arXiv:2403.07376

    2024 arXiv

  5. [13]

    L.; Wei, Z.; Han, M.; Xu, R.; Niu, M.; Han, J.; Lin, L.; Lu, C.; et al

    Lin, B.; Nie, Y.; Zai, K. L.; Wei, Z.; Han, M.; Xu, R.; Niu, M.; Han, J.; Lin, L.; Lu, C.; et al. 2025. EvolveNav: Self-Improving Embodied Reasoning for LLM-Based Vision-Language Navigation. arXiv preprint arXiv:2506.01551

    2025

  6. [14]

    Long, Y.; Li, X.; Cai, W.; and Dong, H. 2024. Discuss before moving: Visual language navigation via multi-expert discussions. In 2024 IEEE International Conference on Robotics and Automation (ICRA), 17380--17387. IEEE

    2024

  7. [15]

    Y.; Shen, C.; and Hengel, A

    Qi, Y.; Wu, Q.; Anderson, P.; Wang, X.; Wang, W. Y.; Shen, C.; and Hengel, A. v. d. 2020. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9982--9991

    2020

  8. [16]

    Reimers, N. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. arXiv preprint arXiv:1908.10084

    2019 arXiv

  9. [17]

    Schumann, R.; Zhu, W.; Feng, W.; Fu, T.-J.; Riezler, S.; and Wang, W. Y. 2024. Velma: Verbalization embodiment of llm agents for vision and language navigation in street view. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 18924--18933

    2024

  10. [18]

    Shinn, N.; Cassano, F.; Gopinath, A.; Narasimhan, K.; and Yao, S. 2024. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36

    2024

  11. [19]

    Tan, H.; Yu, L.; and Bansal, M. 2019. Learning to navigate unseen environments: Back translation with environmental dropout. arXiv preprint arXiv:1904.04195

    2019 arXiv

  12. [20]

    Wang, Z.; Li, J.; Hong, Y.; Wang, Y.; Wu, Q.; Bansal, M.; Gould, S.; Tan, H.; and Qiao, Y. 2023. Scaling data generation in vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 12009--12020

    2023

  13. [21]

    Zhan, Z.; Yu, L.; Yu, S.; and Tan, G. 2024. Mc-gpt: Empowering vision-and-language navigation with memory map and reasoning chains. arXiv preprint arXiv:2405.10620

    2024 arXiv

  14. [22]

    Zhang, W.; Tang, K.; Wu, H.; Wang, M.; Shen, Y.; Hou, G.; Tan, Z.; Li, P.; Zhuang, Y.; and Lu, W. 2024. Agent-pro: Learning to evolve via policy-level reflection and optimization. arXiv preprint arXiv:2402.17574

    2024 arXiv

  15. [23]

    Zhou, G.; Hong, Y.; and Wu, Q. 2024. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 7641--7649

    2024

  16. [24]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  17. [25]

    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 gl...

Pith tools

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