Pith. sign in

REVIEW 5 major objections 6 minor 44 references

Hindsight Planner: A Closed-Loop Few-Shot Planner for Embodied Instruction Following

T0 review · 5 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a few-shot LLM planner using closed-loop hindsight-corrected examples can match or surpass the performance of a fully supervised planner on the ALFRED embodied instruction following benchmark.

desk verdict The hindsight relabeling idea is genuinely new and the few-shot results are competitive, but the 'surpasses full-shot' claim rests on an unreplicated 0.4-point margin that does not survive contact with stronger baselines. read the letter →

arxiv 2412.19562 v1 pith:5NEMSZAI submitted 2024-12-27 cs.AI cs.RO

Yuxiao Yang , Shenao Zhang , Zhihan Liu , Huaxiu Yao , Zhaoran Wang This is my paper · ORCID
classification cs.AIcs.RO
keywords LargeLanguageModelsIn-ContextLearningEmbodiedInstructionFollowingPOMDPHindsightActor-CriticALFREDFew-Shot
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

This paper tries to show that an embodied instruction-following agent can plan at a level comparable to fully supervised systems using only about 140 training examples. It reframes the problem as a partially observable Markov decision process and builds a closed-loop planner that, at each step, rolls out future sub-goals with an actor-critic beam search and selects the best one. The key novelty is a 'hindsight' method that collects suboptimal trajectories from the agent itself and uses an LLM to relabel them into successful completions, so that the planner can learn to recover from its own mistakes. On ALFRED, the method reports 25.51% success on seen test tasks and 18.77% on unseen test tasks, the first few-shot agent to approach or exceed the best full-shot supervised agents.

What carries the argument

The load-bearing components are the hindsight relabeling procedure and the actor-critic beam-search planner. The hindsight procedure collects suboptimal trajectories from the agent running on 140 training tasks, then prompts an LLM to reflect on each incomplete rollout and append recovery actions using the ground truth rollout as a guide, producing corrected examples stored in a 'hindsight pool' used to prompt the hindsight actor and the critic. The planner uses two actors (ground truth and hindsight) to propose candidate sub-goals, a next-state generator, and a critic to score rollouts, selecting the best first action; this converts open-loop imitation into closed-loop planning. The adaptation module separately predicts invisible PDDL arguments (target object, parent receptacle, toggle target, and others) at each step and appends them to the observed object list.

What would settle it

Measure how often the LLM's appended completions in the hindsight pool match the ground truth continuation on the 140 collected suboptimal rollouts; a low match rate would undercut the assumption that relabeled examples teach genuine recovery.

Watch

Extended reading notes

Core claim

The central claim is that retrospective correction—feeding the planner in-context examples of its own failed rollouts repaired into successful ones—lets a few-shot LLM planner recover from suboptimal actions and match full-shot supervised planners. The paper frames this as finding a policy that optimizes the task objective while keeping the distribution of the goal variable unshifted, unlike standard hindsight replay methods that relabel trajectories to achievable goals and thereby change the task distribution. Concretely, the method prompts two actors, one with ground truth examples and one with hindsight-repaired examples, evaluates their proposed rollouts with a critic, and returns the best first action. This closed-loop procedure, together with an adaptation module that predicts latent PDDL arguments from egocentric observations, yields the reported success rates.

Load-bearing premise

The central assumption is that an LLM can, from a suboptimal rollout and the ground truth trajectory, synthesize correct recovery actions, and that these repaired examples teach the planner to fix similar mistakes in novel situations rather than just memorizing the 140 training tasks. This assumption is not directly evaluated in the paper.

Editorial extensions

If this is right

  • A few-shot LLM planner with hindsight-corrected examples can rival fully supervised planners on ALFRED, meaning large-scale human-annotated trajectories may not be necessary for this class of embodied tasks.
  • Closed-loop planning with beam search over future sub-goals is a viable alternative to open-loop imitation, reducing compounding errors from single-step prediction.
  • The hindsight method's distribution-preserving property, if correct, suggests a general recipe for using failed rollouts in LLM in-context learning without shifting the task distribution.
  • The adaptation module that predicts latent PDDL state at each step shows that explicit state estimation can substitute partially for full observability in a few-shot setting.
  • Since the gains are measured against baselines with a shared low-level controller, the reported improvements are attributable to planning rather than to perception or control.

Reading between the lines

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

  • One can test whether the hindsight relabeling transfers beyond ALFRED, for example to web navigation or tool-use tasks where failed rollouts are cheap to collect; the paper does not try these domains.
  • A natural measurement is to check how often the LLM's appended completions in the hindsight pool match the ground truth continuation on the 140 collected suboptimal rollouts; a low match rate would undercut the claimed recovery benefit.
  • Because the few-shot comparison shares the low-level controller with full-shot baselines, the practical gap depends on that controller's coverage; a better controller might widen the gap, a weaker one might erase it.
  • The framework leaves open whether adding supervised fine-tuning later would still improve the few-shot planner, so the method could be a starting point rather than a replacement.
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 / 6 minor

Summary. This paper proposes Hindsight Planner, a few-shot high-level planner for the ALFRED embodied instruction-following benchmark. The task is framed as a POMDP: an LLM-based adaptation module predicts latent PDDL arguments from the current observation, and a beam-search planner combines two LLM actors (Actorgt prompted with ground-truth trajectories and Actorhind prompted with hindsight-relabeled trajectories) with an LLM critic to select the best sub-goal rollout, which is executed by the frozen HLSM low-level controller. The hindsight relabeling (Algorithm 3) collects suboptimal rollouts from a 140-trajectory pool and asks the LLM to append corrective actions, with the stated aim of preserving the goal distribution z = I(τ*) while providing recovery examples. Experiments report Test Seen SR 25.51 and Test Unseen SR 18.77, and the abstract claims this is the first few-shot planner to approach or surpass a full-shot supervised agent. Section 4.3 argues the hindsight method is theoretically superior to prior hindsight methods, and ablations attribute parts of the gain to the adaptation module and the hindsight prompts.

Significance. Embodied instruction following with few-shot LLM planners is a timely problem, and the paper makes a plausible design contribution: combining a POMDP-style adaptation module that predicts latent PDDL arguments at each step, a closed-loop two-actor plus critic beam search, and a hindsight relabeling step that converts suboptimal rollouts into corrective in-context examples. If the reported results survive robustness testing, the method would be a meaningful step toward few-shot planners that recover from suboptimal actions, and the same-controller comparison in Table 2 is the right experimental design for isolating the planner's contribution. The appendix's detailed prompt templates and algorithm pseudocode are a genuine strength and would facilitate replication. However, the central claimed result, that few-shot performance surpasses a full-shot supervised planner, is not established by the current evidence: the margins over HLSM are within noise on the seen splits and negative on Test Unseen, and stronger full-shot baselines in the paper's own Table 1 (FILM) still outperform the method. The theoretical claim in Section 4.3 is circular in its current form.

major comments (5)
  1. [§5.2, Tables 1–2; Abstract] The headline claim that the few-shot agent 'approaches and even surpasses that of the full-shot supervised agent' is not supported by the reported numbers. In the same-controller comparison (Table 2), Hindsight-Planner beats HLSM full-shot on Test Seen SR (25.51 vs 25.11, +0.40) and Valid Unseen SR (19.00 vs 18.28, +0.72) but loses on Test Unseen SR (18.77 vs 20.27, -1.50), which is the primary held-out generalization split. The positive margins are on the order of a few episodes, yet no seeds, repeated sample-pool draws, confidence intervals, or per-split task counts are reported, and temperature 0 does not eliminate API nondeterminism or low-level controller stochasticity. Moreover, Table 1 shows that the stronger full-shot baselines in the paper's own table (FILM, 28.83/27.80; MCR-Agent, 30.13 on Test Seen) exceed the reported few-shot numbers, so 'the full-shot supervised agent' in the abstract is in fact only HLSM, the weakest cited full-shot method. The central claim should be either re-qualified to 'approaches HLSM full-shot on seen splits' or backed by error bars and a comparison with the stronger full-shot baselines.
  2. [§4.3, Eqs. (4.1)–(4.3)] The theoretical argument does not establish the claimed superiority over previous hindsight methods. Equation (4.2) asserts that the relabeled policy bπ is worse than the oracle π* under p(z) 'as the distribution of z is shifted from p to q', but this inequality is a restatement of the fact that π* is the oracle (the optimal policy under p(z)); it would hold for any suboptimal bπ, including one trained with q = p from finite samples, so it does not isolate the effect of the distribution shift. The key claim that the proposed relabeling keeps the goal distribution unshifted is true by construction: the relabeled trajectory bτ is defined to satisfy I(bτ) = I(τ*) = z, so the KL divergence in Eq. (4.3) is zero by definition whenever the relabeling succeeds. No concrete information statistic I(τ) is ever specified for the ALFRED setting, and the guarantee silently assumes perfect LLM relabeling, which the paper nowhere verifies. I recommend rewriting this section as a design rationale, with the distribution-preservation property stated as a conditional, construction-level observation, or providing a concrete I(τ) and an analysis of imperfect relabeling.
  3. [§5.1, Algorithm 3; Appendix B.2] The core mechanism of the paper, hindsight relabeling, is never validated. Algorithm 3 assumes that an LLM can take a suboptimal rollout and the ground-truth rollout and append correct recovery actions, and that the resulting prompts generalize to novel suboptimal states during deployment. The appendix provides the prompts, but no quantitative measure of relabeling quality is given: no fraction of relabeled trajectories that are syntactically valid, executable in the simulator, or actually complete the goal; no manual inspection statistics; no analysis of whether the appended actions are consistent with the agent's current holding state. Without this, the ablation attribution (Table 4 shows SR drops by 2.08 without hindsight on Valid Seen) cannot distinguish 'corrective in-context learning' from prompt-format artifacts, and the generalization of the hindsight actor to states outside the 140 collected trajectories remains an unverified assumption.
  4. [§5.1–5.2, Table 1] The few-shot comparison does not control for the LLM backbone. The paper states that GPT-4 Turbo is the target LLM but does not report which LLM was used for the two few-shot baselines (LLM-Planner and Socratic-Planner). Since the claimed few-shot gains over LLM-Planner (10.18 SR on Test Seen) are much larger than the full-shot margins, the improvement could partly reflect a stronger base model rather than the proposed planner components. Please report the baselines' backbone models and, if they differ from GPT-4 Turbo, add a baseline running LLM-Planner or Socratic-Planner with the same backbone, or explicitly discuss this confound.
  5. [§4.5, Algorithm 2] The planning procedure is underspecified at a load-bearing point: the generator ψ that produces next states s_{u+1} for candidate actions in the beam search is never defined. It is not stated whether ψ is the ALFRED simulator, an LLM-based predictor, or the HLSM low-level controller, and the state representation x_t over which the search operates (object list plus predicted PDDL arguments) is described only informally. Since the entire actor-critic search, the critic's value estimates, and the final rollout selection depend on ψ's outputs, this must be specified, including whether the search uses privileged environment information, for the method to be replicable and for the closed-loop claims in Section 4.1 to be assessable.
minor comments (6)
  1. [§5.3, Table 3] The statement that hindsight prompting is 'especially crucial' in relatively long-horizon tasks is only partially supported by Table 3: on Clean tasks (average sub-goal length 7.25) the base method scores 15.18 versus 16.96 without hindsight, and on Pick Two the numbers are 30.65 versus 29.84, so the component's benefit is not consistently aligned with horizon length; please qualify this claim and report per-task-type episode counts so the reader can judge the size of these differences.
  2. [Appendix B.2] There are typos in the provided prompts ('in orde' in the critic sample, 'mrcrowave' in the Actor_Think prompt), and the second Actor_Complete example ends with 'CloseObject : Fridge' without the terminating 'Stop : NIL' that the first example includes; please specify whether relabeled rollouts are required to end with Stop.
  3. [§5.2 and Introduction] The sentence 'our method achieves a 10.18 and 5.36 higher success rate' is ungrammatical, and the 60% and 39% improvement figures in the introduction do not identify the reference baseline (presumably LLM-Planner with high-level instructions); please state the reference explicitly.
  4. [§1, Contributions] The claim that this is 'for the first time' a study of ALFRED from a POMDP perspective is not supported by a literature search or citation; please soften or remove this novelty statement.
  5. [§4.3, Eqs. (3.3) and (4.1)] The notation is loose: bπ is used without explicit definition, 'Assume the distribution of I(τ) ∼ q' should specify that the distribution is over τ ~ P(·|π, z) and whether q depends on π, and KL(I(τ), z) in Eq. (3.3) is written as a KL divergence over a statistic rather than over the full trajectory distribution.
  6. [Reproducibility] No code, sample pool, or collected hindsight pool D_hind is released; since the hindsight pool is the learned artifact of the method, releasing it together with the kNN retrieval code would be important for independent replication.

Circularity Check

1 steps flagged · score 4.0 of 10

The theoretical claim of an unshifted goal distribution is true by construction; the empirical performance claims remain externally benchmarked.

  1. self definitional [Section 4.3 (Hindsight Method), immediately after Eq. (4.3)]
    "Our method utilizes LLMs to relabel bτ = τT + {aT , xT +1, yT +1, aT +2, . . .} in such a way that I(bτ ) = I(τ ∗) = z. Thus, we minimize the divergence in Equation (4.3) while keeping the distribution of z unshifted."

    The claimed benefit (the distribution of z remains unshifted) is exactly the defining condition imposed on the relabeling: the sentence says the trajectory is relabeled 'in such a way that I(bτ) = I(τ*) = z', and the next sentence presents 'keeping the distribution of z unshifted' as a consequence. The equality is therefore an assumption built into the construction, not a derived result. Since I(·) is never specified and relabeling quality is never evaluated, the theoretical superiority over prior hindsight methods reduces to a definitional tautology. This does not taint the empirical benchmark comparisons, which stand or fall on external baselines.

full rationale

The one clear circular step is in Section 4.3: the theoretical argument that the hindsight method avoids the distribution shift of earlier relabeling methods is true by construction, because the relabeled trajectory bτ is defined to make I(bτ) = I(τ*). The paper then reads 'Thus ... keeping the distribution of z unshifted' as if this were a derived guarantee. This is a self-definitional reduction of the theoretical claim. I also examined the other candidate circularities. The actor-critic framework is borrowed from Liu et al. (2024), which includes overlapping authors, but it is used as a design choice and is not the vehicle of the paper's main claim; no uniqueness theorem is imported to forbid alternatives. The empirical claims—25.51 SR on Test Seen, 18.77 SR on Test Unseen, and the ablations in Tables 3–4—are measured against external baselines such as HLSM, LLM-Planner, Socratic-Planner, and full-shot supervised models, so the headline performance claim has independent grounding. Its fragility (no seeds, no error bars, sub-episode margins, and a 1.50 SR deficit on Test Unseen versus HLSM full) is a correctness and reproducibility concern, not circularity. Overall score 4: one partial circularity in the theory, while the central empirical comparison is not built from self-defined quantities.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The empirical result depends on the low-level controller, the object list perception, and the LLM's in-context behavior; the theoretical argument depends on an information statistic that is assumed to be matchable by relabeling. None of these are independently verified in the paper.

free parameters (4)
  • in-context example counts K = K=5 for Adapter; K=2 for Actors and Critic
    Chosen by hand in Section 5.1; no sensitivity analysis.
  • beam search depth and width = depth=2, width=2
    Set in Section 5.1; no ablation on these hyperparameters.
  • training trajectories per task type = 20 per type (140 total)
    Chosen to define the few-shot regime; no justification or sensitivity analysis.
  • LLM sampling temperature = 0
    Set in Section 5.1; reasonable for reproducibility but not varied.
assumptions (5)
  • domain assumption The EIF task can be modeled as a POMDP with state approximated by the observed object list plus predicted PDDL arguments.
    Section 3.1 and 4.4; the planner only receives an object list, not raw images.
  • ad hoc to paper The information statistic I(τ) exists and can be set equal to the goal variable z by relabeling bτ.
    Section 4.3, Eq 4.3; this is the core of the theoretical claim and is assumed by construction.
  • domain assumption The low-level controller HLSM executes high-level sub-goals reliably enough that planner improvements translate to task success.
    Section 5.1; the planner output is a sub-goal, and Low-PL is a pretrained black box.
  • domain assumption The frozen BERT kNN retriever selects in-context examples that are informative for the current task.
    Section 4.2 and 5.1; no analysis of retrieval quality is provided.
  • domain assumption GPT-4 Turbo's in-context learning is stable enough at temperature 0 for the reported results.
    Section 5.1; no multi-run variance is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hindsight Planner: A Closed-Loop Few-Shot Planner for Embodied Instruction Following." pith.science (2026). https://pith.science/paper/5NEMSZAI

@misc{pith2026241219562,
  author       = {Pith},
  title        = {Pith review of: Hindsight Planner: A Closed-Loop Few-Shot Planner for Embodied Instruction Following},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5NEMSZAI}},
  note         = {Machine review of arXiv:2412.19562}
}
read the original abstract

This work focuses on building a task planner for Embodied Instruction Following (EIF) using Large Language Models (LLMs). Previous works typically train a planner to imitate expert trajectories, treating this as a supervised task. While these methods achieve competitive performance, they often lack sufficient robustness. When a suboptimal action is taken, the planner may encounter an out-of-distribution state, which can lead to task failure. In contrast, we frame the task as a Partially Observable Markov Decision Process (POMDP) and aim to develop a robust planner under a few-shot assumption. Thus, we propose a closed-loop planner with an adaptation module and a novel hindsight method, aiming to use as much information as possible to assist the planner. Our experiments on the ALFRED dataset indicate that our planner achieves competitive performance under a few-shot assumption. For the first time, our few-shot agent's performance approaches and even surpasses that of the full-shot supervised agent.

Figures

Figures reproduced from arXiv: 2412.19562 by the authors.

Figure 1
Figure 1. Left: The illustration of the Hindsight Planner: at each time step t, the planner receives a partial observation y t from the environment. The adaptation module estimates the latent variable and concatenates it with y t to produce the complete state xt. Actorhind and Actorgt are prompted with different samples and make decisions. The Critic is utilized to evaluate the actions. The best rollout (xt, a∗ t , x∗ t+1, a∗… view at source ↗
Figure 2
Figure 2. A comparison of Hindsight Planner and previous supervised methods when taking a suboptimal action. The agent initially picks up the incorrect object (“Basketball”). In the supervised method, the planner fails to handle this situation, which leads to task failure. In contrast, the Hindsight Planner can adjust after the incorrect action and successfully complete the task. At time step t, the planner receives an observ… view at source ↗
Figure 3
Figure 3. The entire process of the Hindsight Planner is as follows: At the start of the task, which is to “Place a plate with a ladle on it in a cabinet,” the Adapter mistakenly identifies the task as picking up a plate and placing it into a cabinet. Actorhind and Actorgt make decisions separately. Critic then selects the best action as its output. Upon further exploration, the agent detects more objects, and the Adapter adj… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 9 canonical work pages

  1. [1]

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

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sentence := ...

  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]

    , Brohan, A

    Ahn, M. , Brohan, A. , Brown, N. , Chebotar, Y. , Cortes, O. , David, B. , Finn, C. , Fu, C. , Gopalakrishnan, K. , Hausman, K. , Herzog, A. , Ho, D. , Hsu, J. , Ibarz, J. , Ichter, B. , Irpan, A. , Jang, E. , Ruano, R. J. , Jeffrey, K. , Jesmonth, S. , Joshi, N. J. , Julian, R. , Kalashnikov, D. , Kuang, Y. , Lee, K.-H. , Levine, S. , Lu, Y. , Luu, L. , ...

    arXiv 2022

  4. [4]

    , Wolski, F

    Andrychowicz, M. , Wolski, F. , Ray, A. , Schneider, J. , Fong, R. , Welinder, P. , McGrew, B. , Tobin, J. , Abbeel, P. and Zaremba, W. (2018). Hindsight experience replay. ://arxiv.org/abs/1707.01495

    arXiv 2018

  5. [5]

    Multi-Level Compositional Reasoning for Interactive Instruction Following

    Bhambri, S. , Kim, B. and Choi, J. (2024). Multi-level compositional reasoning for interactive instruction following. ://arxiv.org/abs/2308.09387

    work page Pith review arXiv 2024

  6. [6]

    , Paxton, C

    Blukis, V. , Paxton, C. , Fox, D. , Garg, A. and Artzi, Y. (2021). A persistent spatial semantic representation for high-level natural language instruction execution. Cornell University - arXiv,Cornell University - arXiv

    work page 2021

  7. [7]

    Brown, T. B. , Mann, B. , Ryder, N. , Subbiah, M. , Kaplan, J. , Dhariwal, P. , Neelakantan, A. , Shyam, P. , Sastry, G. , Askell, A. , Agarwal, S. , Herbert-Voss, A. , Krueger, G. , Henighan, T. , Child, R. , Ramesh, A. , Ziegler, D. M. , Wu, J. , Winter, C. , Hesse, C. , Chen, M. , Sigler, E. , Litwin, M. , Gray, S. , Chess, B. , Clark, J. , Berner, C. ...

    arXiv 2020

  8. [8]

    Chapman, D. (1987). Planning for conjunctive goals. Artif. Intell., 32 333--377. ://api.semanticscholar.org/CorpusID:1525549

    work page 1987

Show all 44 references
  1. [9]

    , Sun, Y

    Dai, D. , Sun, Y. , Dong, L. , Hao, Y. , Ma, S. , Sui, Z. and Wei, F. (2023). Why can gpt learn in-context? language models implicitly perform gradient descent as meta-optimizers. ://arxiv.org/abs/2212.10559

    2023 arXiv

  2. [10]

    , Chang, M.-W

    Devlin, J. , Chang, M.-W. , Lee, K. and Toutanova, K. (2019). Bert: Pre-training of deep bidirectional transformers for language understanding. ://arxiv.org/abs/1810.04805

    2019 arXiv

  3. [11]

    Dong, Q. , Li, L. , Dai, D. , Zheng, C. , Ma, J. , Li, R. , Xia, H. , Xu, J. , Wu, Z. , Chang, B. , Sun, X. , Li, L. and Sui, Z. (2024). A survey on in-context learning. ://arxiv.org/abs/2301.00234

    2024 arXiv

  4. [12]

    , Geng, X

    Eysenbach, B. , Geng, X. , Levine, S. and Salakhutdinov, R. (2020). Rewriting history with inverse rl: Hindsight inference for policy improvement. ://arxiv.org/abs/2002.11089

    2020 arXiv

  5. [13]

    , Matsuo, Y

    Furuta, H. , Matsuo, Y. and Gu, S. S. (2022). Generalized decision transformer for offline hindsight information matching. ://arxiv.org/abs/2111.10364

    2022 arXiv

  6. [14]

    , Gupta, A

    Ghosh, D. , Gupta, A. , Fu, J. , Reddy, A. , Devin, C. , Eysenbach, B. and Levine, S. (2019). Learning to reach goals without reinforcement learning. ArXiv, abs/1912.06088

    2019 arXiv

  7. [15]

    , Zhang, R

    Guo, J. , Zhang, R. , Zhang, X. , Peng, S. , Yi, Q. , Du, Z. , Hu, X. , Guo, Q. and Chen, Y. (2021). Hindsight value function for variance reduction in stochastic dynamic environment. ://arxiv.org/abs/2107.12216

    2021 arXiv

  8. [16]

    , Kakade, S

    Hazan, E. , Kakade, S. M. , Singh, K. and Soest, A. V. (2019). Provably efficient maximum entropy exploration. ://arxiv.org/abs/1812.02690

    2019 arXiv

  9. [17]

    , Kim, J

    Kim, B. , Kim, J. , Kim, Y. , Min, C. and Choi, J. (2024). Context-aware planning and environment-aware memory for instruction following embodied agents. ://arxiv.org/abs/2308.07241

    2024 arXiv

  10. [18]

    Kumar, A. , Fu, Z. , Pathak, D. and Malik, J. (2021). Rma: Rapid motor adaptation for legged robots. ://arxiv.org/abs/2107.04034

    2021 arXiv

  11. [19]

    Lee, J. N. , Agarwal, A. , Dann, C. and Zhang, T. (2023). Learning in pomdps is sample-efficient with hindsight observability. ://arxiv.org/abs/2301.13857

    2023 arXiv

  12. [20]

    , Eysenbach, B

    Lee, L. , Eysenbach, B. , Parisotto, E. , Xing, E. , Levine, S. and Salakhutdinov, R. (2020). Efficient exploration via state marginal matching. ://arxiv.org/abs/1906.05274

    2020 arXiv

  13. [21]

    Li, A. C. , Pinto, L. and Abbeel, P. (2020). Generalized hindsight for reinforcement learning. ://arxiv.org/abs/2002.11708

    2020 arXiv

  14. [22]

    Liu, Z. , Hu, H. , Zhang, S. , Guo, H. , Ke, S. , Liu, B. and Wang, Z. (2024). Reason for future, act for now: A principled framework for autonomous llm agents with provable sample efficiency. ://arxiv.org/abs/2309.17382

    2024 arXiv

  15. [23]

    , Chaplot, D

    Min, S. , Chaplot, D. , Ravikumar, P. , Bisk, Y. and Salakhutdinov, R. (2021). Film: Following instructions in language with modular methods. Learning,Learning

    2021

  16. [24]

    , Adler, S

    OpenAI , Achiam, J. , Adler, S. , Agarwal, S. , Ahmad, L. , Akkaya, I. , Aleman, F. L. , Almeida, D. , Altenschmidt, J. , Altman, S. , Anadkat, S. , Avila, R. , Babuschkin, I. , Balaji, S. , Balcom, V. , Baltescu, P. , Bao, H. , Bavarian, M. , Belgum, J. , Bello, I. , Berdine,...

    2024 arXiv

  17. [25]

    Peng, X. B. , Coumans, E. , Zhang, T. , Lee, T.-W. , Tan, J. and Levine, S. (2020). Learning agile robotic locomotion skills by imitating animals. ://arxiv.org/abs/2004.00784

    2020 arXiv

  18. [26]

    Pong, V. , Gu, S. , Dalal, M. and Levine, S. (2020). Temporal difference models: Model-free deep rl for model-based control. ://arxiv.org/abs/1802.09081

    2020 arXiv

  19. [27]

    , jeon, S

    Shin, S. , jeon, S. , Kim, J. , Kang, G.-C. and Zhang, B.-T. (2024). Socratic planner: Inquiry-based zero-shot planning for embodied instruction following. ://arxiv.org/abs/2404.15190

    2024 arXiv

  20. [28]

    , Thomason, J

    Shridhar, M. , Thomason, J. , Gordon, D. , Bisk, Y. , Han, W. , Mottaghi, R. , Zettlemoyer, L. and Fox, D. (2020). Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). ://d...

    2020

  21. [29]

    , Dan, S

    Silver, T. , Dan, S. , Srinivas, K. , Tenenbaum, J. B. , Kaelbling, L. P. and Katz, M. (2023). Generalized planning in pddl domains with pretrained large language models. ://arxiv.org/abs/2305.11014

    2023 arXiv

  22. [30]

    Song, C. H. , Wu, J. , Washington, C. , Sadler, B. M. , Chao, W.-L. and Su, Y. (2023). Llm-planner: Few-shot grounded planning for embodied agents with large language models. ://arxiv.org/abs/2212.04088

    2023 arXiv

  23. [31]

    , Lavril, T

    Touvron, H. , Lavril, T. , Izacard, G. , Martinet, X. , Lachaux, M.-A. , Lacroix, T. , Rozière, B. , Goyal, N. , Hambro, E. , Azhar, F. , Rodriguez, A. , Joulin, A. , Grave, E. and Lample, G. (2023). Llama: Open and efficient foundation language models. ://arxiv.org/abs/2302.13971

    2023 arXiv

  24. [32]

    , Shazeer, N

    Vaswani, A. , Shazeer, N. , Parmar, N. , Uszkoreit, J. , Jones, L. , Gomez, A. N. , Kaiser, L. and Polosukhin, I. (2023). Attention is all you need. ://arxiv.org/abs/1706.03762

    2023 arXiv

  25. [33]

    , Tay, Y

    Wei, J. , Tay, Y. , Bommasani, R. , Raffel, C. , Zoph, B. , Borgeaud, S. , Yogatama, D. , Bosma, M. , Zhou, D. , Metzler, D. , Chi, E. H. , Hashimoto, T. , Vinyals, O. , Liang, P. , Dean, J. and Fedus, W. (2022). Emergent abilities of large language models. ://arxiv.org/abs/2206.07682

    2022 arXiv

  26. [34]

    , Wang, X

    Wei, J. , Wang, X. , Schuurmans, D. , Bosma, M. , Ichter, B. , Xia, F. , Chi, E. , Le, Q. and Zhou, D. (2023). Chain-of-thought prompting elicits reasoning in large language models. ://arxiv.org/abs/2201.11903

    2023 arXiv

  27. [35]

    , Debut, L

    Wolf, T. , Debut, L. , Sanh, V. , Chaumond, J. , Delangue, C. , Moi, A. , Cistac, P. , Rault, T. , Louf, R. , Funtowicz, M. , Davison, J. , Shleifer, S. , von Platen, P. , Ma, C. , Jernite, Y. , Plu, J. , Xu, C. , Scao, T. L. , Gugger, S. , Drame, M. , Lhoest, Q. and Rush, A. ...

    2020 arXiv

  28. [36]

    Xie, S. M. , Raghunathan, A. , Liang, P. and Ma, T. (2022). An explanation of in-context learning as implicit bayesian inference. ://arxiv.org/abs/2111.02080

    2022 arXiv

  29. [37]

    Yao, S. , Yu, D. , Zhao, J. , Shafran, I. , Griffiths, T. L. , Cao, Y. and Narasimhan, K. (2023). Tree of thoughts: Deliberate problem solving with large language models. ://arxiv.org/abs/2305.10601

    2023 arXiv

  30. [38]

    Yao, Y. , Li, Z. and Zhao, H. (2024). Beyond chain-of-thought, effective graph-of-thought reasoning in language models. ://arxiv.org/abs/2305.16582

    2024 arXiv

  31. [39]

    and Chai, J

    Zhang, Y. and Chai, J. (2021). Hierarchical task learning from language instructions with unified transformers and self-monitoring. ://arxiv.org/abs/2106.03427

    2021 arXiv

  32. [40]

    Zhang, Y. , Li, Y. , Cui, L. , Cai, D. , Liu, L. , Fu, T. , Huang, X. , Zhao, E. , Zhang, Y. , Chen, Y. , Wang, L. , Luu, A. T. , Bi, W. , Shi, F. and Shi, S. (2023). Siren's song in the ai ocean: A survey on hallucination in large language models. ://arxiv.org/abs/2309.01219

    2023 arXiv

  33. [41]

    , Pinto, L

    Zhou, W. , Pinto, L. and Gupta, A. (2019). Environment probing interaction policies. Publisher Copyright: 7th International Conference on Learning Representations, ICLR 2019. All Rights Reserved.; 7th International Conference on Learning Representations, ICLR 2019 ; Conference...

    2019

  34. [42]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  35. [43]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  36. [44]

    Put a chilled potato on the small black table

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

    2007

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.