Large Language Model-enhanced Reinforcement Learning for Low-Altitude Economy Networking

Changyuan Zhao; Dusit Niyato; Jiawen Kang; Lingyi Cai; Ruichen Zhang; Tao Jiang; Xuemin Shen; Yu Zhang

REVIEW 4 major objections 5 minor 5 cited by

Large Language Model-enhanced Reinforcement Learning for Low-Altitude Economy Networking

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read An LLM-written reward function beats a handcrafted one in simulated UAV networking, cutting energy use by up to 7.2 percent.

desk verdict Useful tutorial, but the case study's headline result is confounded by an added position term and lacks the seeds/ablations that would attribute the gain to LLM reward design. read the letter →

arxiv 2505.21045 v1 pith:VQLSQGH3 submitted 2025-05-27 cs.AI

Lingyi Cai , Ruichen Zhang , Changyuan Zhao , Yu Zhang , Jiawen Kang , Dusit Niyato , Tao Jiang , Xuemin Shen This is my paper · ORCID

classification cs.AI

keywords low-altitudeeconomynetworkinglargelanguagemodelsreinforcementlearningrewardfunctiondesignUAV-assistedIoTenergyconsumptionoptimizationLLM-enhancedRL

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 argues that large language models can be woven into reinforcement learning for low-altitude economy networking, with the model acting as information processor, reward designer, decision-maker, and generator. The illustrative claim is concrete: when prompted with a role definition and a task description, the LLM writes a Python reward function for a UAV-assisted IoT data-collection task, and RL agents trained with that reward consume less energy than agents trained with a manually crafted reward from prior work. In the simulated marine IoT zone, TD3 with the LLM-designed reward reaches up to 7.2 percent lower final energy consumption, and the advantage holds across tested packet sizes. A sympathetic reader would take the paper as evidence that LLM-based reward design can automate part of reward engineering in constrained aerial networking tasks.

What carries the argument

The load-bearing mechanism is the prompt-and-filter loop that turns task language into executable reward code. A structured prompt with role definition, constraints, and a JSON output format asks the model to reason step by step about which observable states are good or bad, then produce a Python function. Multiple candidate functions are generated and filtered by logical consistency and constraint checks, a practice the paper credits to reflection-style candidate evaluation, before the surviving reward function is inserted into the RL training loop. What carries the improvement is the added position term returned by that process.

What would settle it

Repeat the case study with the manual reward augmented by the identical position term, $w \times \text{energy} \times \text{Penalty} + w_2 \times \text{position}$, using multiple random seeds for both DDPG and TD3. If the augmented manual reward matches or beats the LLM reward's energy consumption, the paper's attribution to LLM reasoning collapses; if the LLM reward still wins across seeds, the claim survives.

Watch

Extended reading notes

Core claim

The central claim, on the paper's own terms, is that reward design, a known bottleneck in reinforcement learning, can be delegated to an LLM. The LLM-generated reward function $(w_1 \times \text{energy} + w_2 \times \text{position}) \times \text{Penalty}$ differs from the manual baseline $w \times \text{energy} \times \text{Penalty}$ by adding a position factor that pushes the UAV toward the center of the sensor distribution, reducing flight distance, travel time, and hovering overhead. Across 200 training episodes, DDPG and TD3 with the LLM-designed reward consistently report lower total energy consumption; the largest reported gap is 7.2 percent for TD3 in final energy, and 6.2 percent at the 2.0-Mbits packet size. The paper attributes these gains to the LLM's ability to incorporate high-level reasoning and task-specific context when generating reward functions.

Load-bearing premise

The whole measured advantage rests on the assumption that the manually designed reward from earlier work is a fair baseline, so the improvement can be credited to the LLM's reward design rather than to the extra position term alone.

Editorial extensions

If this is right

If RL agents genuinely learn faster with LLM-designed rewards, reward engineering in LAENet can shift from manual tuning to prompt-based specification.
The same four-role framework predicts that LLM-generated state representations, action candidates, and simulated trajectories could each improve RL, and each role can be evaluated separately.
Because the LLM reward adds a position factor absent from the baseline, the result implies that even simple task-relevant reward factors overlooked by manual design can yield measurable energy savings.
The prompt structure, role definition plus task description, offers a reusable template for other constrained networking tasks, not just UAV energy minimization.

Reading between the lines

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

The paper does not test whether adding the same position term to the manual reward closes the gap; if it does, the contribution is the reward factor rather than the LLM's reasoning. This is an editorial inference.
A natural extension is to vary the LLM, the prompt template, and the random seeds; if gains persist across those variations, the mechanism is robust, and if not, the effect may be idiosyncratic to one model and one run.
The framework's other roles, LLM as simulator and LLM as decision-maker, could be tested on the same UAV task by measuring whether generated trajectories or action candidates improve sample efficiency.

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.

Desk Editor's Note

read the letter

Read this if you want a quick map of how LLMs are being bolted onto RL for aerial networking. The tutorial half is genuinely useful: the four-role framework (information processor, reward designer, decision-maker, generator) is clearly laid out, and the workflow example ties it to a concrete UAV task. The survey of prior work is competent and the figure summarizing LLM-enhanced RL is a nice orientation tool. Credit where it's due: this is a reasonable synthesis, and the framework is a sensible organizing device.

The empirical part is where I'd push back. The case study shows GPT-4o generating a reward function that adds a position term, and the paper reports 7.2% (TD3) and 6.2% energy savings over a manual baseline. But the comparison is between a reward with a position term and one without it. The paper itself says the gain comes from that position factor, which is a standard UAV-placement trick. Without an ablation that adds the same position term to the manual reward, without multi-seed runs or error bars, and without any report of the reward weights or penalty terms, the 7.2% number is suggestive but not established. The stress-test note got this right. I don't think the authors are being deceptive—they're just one careful experiment away from a credible claim.

The tutorial and framework are not affected by this weakness. For a reader coming into the area, the paper is a decent entry point. For a reviewer, the case study needs a substantial revision: run multiple seeds, show variance, ablate the position term on both sides, and release the code and prompts. If that's done, the result would be worth publishing on its own. As written, the empirical claim is a pilot experiment, not a demonstration.

My recommendation: send it to review, but with the expectation of major revision. The tutorial half deserves to be in the literature, and the case study is worth referee time to sort out. I would not cite the empirical numbers in my own work yet, but I might cite the framework. This is an honest paper with a clear structure; it just oversells its case study.

Referee Report

4 major / 5 minor

Summary. The paper presents a tutorial on enhancing reinforcement learning (RL) with large language models (LLMs), proposes a framework in which the LLM serves as information processor, reward designer, decision-maker, and generator, and reports a case study on using GPT-4o to design a reward function for a UAV-assisted IoT network in the low-altitude economy. The case study claims that the LLM-designed reward yields up to 7.2% lower final energy consumption (TD3, Fig. 4) and up to 6.2% lower energy across packet sizes (Fig. 5) compared with a manually designed reward taken from reference [3].

Significance. If the case-study claim were established, the work would provide a useful, falsifiable demonstration of LLM-based reward design in a concrete networking scenario. The tutorial and the four-role framework are clearly structured and offer a reasonable organizing principle for researchers entering this area. The paper does not ship machine-checked proofs or reproducible code, and the empirical core is a single illustrative experiment; nevertheless, the central claim is clearly stated and testable. The main value at present is the synthesis and framework, with the case study as an illustrative, but not yet fully convincing, demonstration.

major comments (4)

[Section IV-C, Figs. 4 and 5] The central quantitative claim (up to 7.2% and 6.2% lower energy) rests on what appear to be single training runs. RL training is highly variable across random seeds; without multiple seeds, error bars, or a significance test, the observed gap may be within run-to-run noise. Please report means and standard deviations (or confidence intervals) over at least five independent seeds and, ideally, a paired test comparing reward designs.
[Section IV-B, Fig. 3 and 'Exploration of Reward Factors'] The comparison confounds the reward-design method with reward content. The manually designed baseline is reward = w * energy * Penalty, while the LLM-designed reward is (w1 * energy + w2 * position) * Penalty. The paper itself states that the added position factor is what reduces propulsion energy by keeping the UAV near the sensor-distribution center. The data are therefore equally consistent with 'LLMs added a useful conventional position term' and 'LLMs are better reward designers.' Add an ablation that augments the manual reward with the same position term, and/or evaluate an LLM-designed reward without the position term, so that the method of reward design is the only variable.
[Section IV-B, 'LLM-designed Reward Evaluation' and Section IV-C] The reward weights (w, w1, w2), the penalty terms, and the GPT-4o generation settings (temperature, number of candidates, selection criterion) are not reported. Without these details the experiment is not reproducible and the reader cannot tell whether the comparison was fair. Please provide the full reward functions, the exact prompt, and the sampling/selection protocol.
[Section IV-B, 'LLM-designed Reward Evaluation'] The acceptance criteria for generated candidates are only syntactic (valid JSON, correct return type). This does not constitute an evaluation of reward quality, so the phrase 'LLM-designed Reward Evaluation' is misleading. If the paper claims a multi-candidate selection mechanism, it must specify how candidates are ranked; otherwise, describe the check as a validity filter.

minor comments (5)

[Fig. 3 caption] The caption contains a typo: 'funtioon' should be 'function'.
[Fig. 3, JSON example] In the JSON example, 'thee IoT network' should be 'the IoT network'.
[Fig. 4] The inset showing the first 20 episodes is too small to read; consider enlarging it or using a separate panel.
[Section II-B] The citation numbering around [11]-[13] should be checked; reference [12] is cited for multi-round dialogue reward shaping, but the flow of references in the text is not always transparent.
[Introduction] The term 'UA Vs' appears with inconsistent spacing; use 'UAVs' consistently throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the case study is an external evaluation of an LLM-generated reward against an external manual baseline, and the paper's self-citations are not load-bearing.

full rationale

The paper's central empirical claim is Contribution 3: LLM-designed reward functions improve RL performance in the LAENet case study. This claim is evaluated by training DDPG and TD3 agents with (a) a manually designed reward taken from prior external work [3] and (b) a reward function generated by GPT-4o from a task-description prompt before training. The reward weights are not fitted to the reported energy-consumption numbers, and the reported 7.2% and 6.2% improvements are measured outcomes rather than quantities constructed from the inputs. The absence of an ablation that adds the position term to the manual baseline is a genuine threat to causal attribution, because the LLM reward differs from the manual reward by including a position-related term; however, this is a confound or validity concern, not circularity, since the LLM's output is not defined in terms of the measured outcome and the baseline is external rather than derived from the LLM reward. The paper's self-citations [5] and [14] support background statements about LLM capabilities and do not enter the case-study equations. No invariance or uniqueness theorem is imported, and no known result is merely renamed as a framework. The tutorial and framework portions are organizational and do not claim a derivation that reduces to its own inputs. Therefore, no specific circular step can be exhibited, and the circularity score is 0.

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

The central claim rests mainly on the single case study. The ledger counts the reward weights and penalty terms as free parameters because their values are not reported, and the GPT-4o sampling configuration as a free parameter because it governs the stochastic reward generation. Domain assumptions include the MDP modeling of the task, the Rician channel and deployment model, the qualitative energy model, and the convergence of DDPG/TD3 in 200 episodes. The syntactic validation gate is an ad hoc step specific to this pipeline. No invented physical entities appear.

free parameters (4)

w1, w2 (LLM-designed reward weights) = not reported in the paper
The LLM reward is reward = (w1*energy + w2*position)*Penalty (Section IV-B), but the weight values, whether they come from the prompt, the LLM, or hand tuning, are not given. The reported energy curves depend on them.
w (manual baseline reward weight) = taken from prior work [3], not restated
The manual baseline reward = w*energy*Penalty (Fig. 3) uses a weight inherited from reference [3]; any unfairness in that weight choice directly affects the claimed 7.2% and 6.2% gains.
Penalty terms in both rewards = not specified in the text
Penalty terms appear in both reward functions (Fig. 3) but are never defined; they influence the learned policy and the absolute energy-consumption values.
GPT-4o candidate generation settings = not reported
The authors generate candidate reward functions and select one that passes syntactic checks (Section IV-B), but do not report the number of candidates, sampling temperature, or selection procedure across independent runs; this affects the stability of the reported comparison.

assumptions (5)

domain assumption The UAV decision problem is modeled as a Markov Decision Process with states capturing spatial, energy, and communication conditions (Section III-B, Step 1).
The entire RL formulation, including the four-role framework and the case study, assumes that LAENet decision-making is adequately represented as such an MDP; if the state space misses relevant factors, the learned policies and the measured energy savings would change.
domain assumption Rician fading channel between the UAV and terminals, with a 300m x 300m area and 10 randomly deployed terminals (Section IV-C).
The simulation results in Fig. 4 and Fig. 5 depend on this channel and deployment model; other channel models or terminal densities could change the magnitude or sign of the reported differences.
domain assumption Total energy consumption is the sum of terminal transmission energy and UAV propulsion and communication energy (Section IV-A), taken as given without closed-form equations.
The reward functions target this energy measure, but the text does not derive or state the energy equations, so the objective is an assumption inherited from the system description.
domain assumption DDPG and TD3 policy-gradient updates converge to a representative policy within 200 training episodes (Section IV-C).
The comparison in Fig. 4 assumes both algorithms with both reward designs reach a stable, representative performance within 200 episodes; different hyperparameters or run lengths could change the ordering of the curves.
ad hoc to paper Syntactic checks (valid JSON, correct return type) are sufficient to accept an LLM-generated reward function (Section IV-B, 'LLM-designed Reward Evaluation').
The paper's evaluation gate verifies only code validity, not whether the reward semantics match the objective; this gate is specific to the paper's pipeline and is inspired by the reflexion-style candidate generation in [11].

how reviews work

0 comments

Cite this review

Pith. "Pith review of Large Language Model-enhanced Reinforcement Learning for Low-Altitude Economy Networking." pith.science (2026). https://pith.science/paper/VQLSQGH3

@misc{pith2026250521045,
  author       = {Pith},
  title        = {Pith review of: Large Language Model-enhanced Reinforcement Learning for Low-Altitude Economy Networking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VQLSQGH3}},
  note         = {Machine review of arXiv:2505.21045}
}

read the original abstract

Low-Altitude Economic Networking (LAENet) aims to support diverse flying applications below 1,000 meters by deploying various aerial vehicles for flexible and cost-effective aerial networking. However, complex decision-making, resource constraints, and environmental uncertainty pose significant challenges to the development of the LAENet. Reinforcement learning (RL) offers a potential solution in response to these challenges but has limitations in generalization, reward design, and model stability. The emergence of large language models (LLMs) offers new opportunities for RL to mitigate these limitations. In this paper, we first present a tutorial about integrating LLMs into RL by using the capacities of generation, contextual understanding, and structured reasoning of LLMs. We then propose an LLM-enhanced RL framework for the LAENet in terms of serving the LLM as information processor, reward designer, decision-maker, and generator. Moreover, we conduct a case study by using LLMs to design a reward function to improve the learning performance of RL in the LAENet. Finally, we provide a conclusion and discuss future work.

Figures

Figures reproduced from arXiv: 2505.21045 by the authors.

**Figure 1.** An illustration of RL, LLM, and applications of LLM-enhanced RL. The number of peer-reviewed publications regarding RL, LLM, and LLM-enhanced [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** An overview of the LLM’s multiple roles in reinforcement learning, including information processor, reward designer, decision-maker, and generator, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** UAV-assisted IoT network with LLM-designed reward funtioon for RL in the LAENet. The UAV agent interacts with the environment by selecting [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Energy consumption over episodes of different algorithm with [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Effect of packet size on energy consumption using different reward [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Efficient Onboard Vision-Language Inference in UAV-Enabled Low-Altitude Economy Networks via LLM-Enhanced Optimization
cs.LG 2025-10 conditional novelty 5.0 of 10

A hierarchical ARPO+LLaRA framework that jointly sets image resolution, transmit power, and UAV trajectory reduces simulated latency for onboard VLM inference in low-altitude economy networks.
Large Language Models for Next-Generation Wireless Network Management: A Survey and Tutorial
cs.NI 2025-09 conditional novelty 4.0 of 10

A survey and tutorial that organizes LLM-enabled wireless network optimization into formulation, solution, and verification stages, with case studies drawn from the authors' own prior papers.
Toward Edge General Intelligence with Agentic AI and Agentification: Concepts, Technologies, and Future Directions
cs.NI 2025-08 conditional novelty 4.0 of 10

A survey that organizes agentic AI for 6G edge networks into four pillars, compactness, efficiency, knowledge and reasoning, and migration, and illustrates them with prior case studies.
Toward Edge General Intelligence with Multiple-Large Language Model (Multi-LLM): Architecture, Trust, and Orchestration
cs.NI 2025-07 conditional novelty 4.0 of 10

A survey of multi-LLM systems in edge computing, covering architectures, enabling technologies, trust mechanisms, applications, and open datasets for edge general intelligence.
HybridRAG-based LLM Agents for Low-Carbon Optimization in Low-Altitude Economy Networks
cs.NI 2025-06 reject novelty 3.0 of 10

HybridRAG merges keyword, vector, and graph retrieval to let an LLM formulate carbon-emission optimization problems for multi-UAV MEC networks, and R2DSAC solves them with a diffusion-regularized SAC plus neuron pruni...

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages · cited by 5 Pith papers

[3]

Synchronizing UA V teams for timely data collection and energy transfer by deep reinforcement learning,

O. S. Oubbatiet al., “Synchronizing UA V teams for timely data collection and energy transfer by deep reinforcement learning,”IEEE Trans. Veh. Technol., vol. 71, no. 6, pp. 6682–6697, 2022

work page 2022
[1]

Unauthorized UA V countermeasure for low-altitude econ- omy: Joint communications and jamming based on MIMO cellular systems,

Z. Liet al., “Unauthorized UA V countermeasure for low-altitude econ- omy: Joint communications and jamming based on MIMO cellular systems,”IEEE Internet Things J., vol. 12, no. 6, pp. 6659–6672, 2025

work page 2025
[2]

Multi-UA V-enabled energy-efficient data delivery for low-altitude economy: Joint coded caching, user grouping, and UA V deployment,

Q. Weiet al., “Multi-UA V-enabled energy-efficient data delivery for low-altitude economy: Joint coded caching, user grouping, and UA V deployment,”IEEE Internet Things J., pp. 1–1, 2025

work page 2025
[4]

Survey on large language model-enhanced reinforcement learning: Concept, taxonomy, and methods,

Y . Caoet al., “Survey on large language model-enhanced reinforcement learning: Concept, taxonomy, and methods,”IEEE Trans. Neural Netw. Learn. Syst., pp. 1–21, 2024

work page 2024
[5]

Generative AI agents with large language model for satellite networks via a mixture of experts transmission,

R. Zhanget al., “Generative AI agents with large language model for satellite networks via a mixture of experts transmission,”IEEE J. Sel. Areas Commun., vol. 42, no. 12, pp. 3581–3596, 2024

work page 2024
[6]

Chain-of-thought prompting elicits reasoning in large language models,

J. Weiet al., “Chain-of-thought prompting elicits reasoning in large language models,” inProc. NeurIPS, vol. 35, 2022, pp. 24 824–24 837

work page 2022
[7]

Reward design with language models,

M. Kwonet al., “Reward design with language models,” inProc. ICLR, 2023

work page 2023
[8]

History compression via language models in rein- forcement learning,

F. Paischeret al., “History compression via language models in rein- forcement learning,” inProc. ICML, 2022, pp. 17 156–17 185

work page 2022

Show all 15 references

[9]

Plan-seq-learn: Language model guided RL for solving long horizon robotics tasks,

M. Dalalet al., “Plan-seq-learn: Language model guided RL for solving long horizon robotics tasks,” inProc. ICLR, 2024

2024
[10]

Unleashing the power of pre-trained language models for offline reinforcement learning,

R. Shiet al., “Unleashing the power of pre-trained language models for offline reinforcement learning,” inProc. ICLR, 2024

2024
[11]

Reflexion: language agents with verbal reinforcement learning,

N. Shinnet al., “Reflexion: language agents with verbal reinforcement learning,” inProc. NeurIPS, vol. 36, 2023, pp. 8634–8652

2023
[12]

Rl2: Reinforce large language model to assist safe reinforcement learning for energy management of active distribution networks,

X. Yanget al., “Rl2: Reinforce large language model to assist safe reinforcement learning for energy management of active distribution networks,”arXiv preprint arXiv:2412.01303, 2024

2024 arXiv
[13]

Rlingua: Improving reinforcement learning sample efficiency in robotic manipulations with large language models,

L. Chenet al., “Rlingua: Improving reinforcement learning sample efficiency in robotic manipulations with large language models,”IEEE Robot. Autom. Lett., vol. 9, no. 7, pp. 6075–6082, 2024

2024
[14]

Large language model guided reinforcement learning based six-degree-of-freedom flight control,

Y . Hanet al., “Large language model guided reinforcement learning based six-degree-of-freedom flight control,”IEEE Access, vol. 12, pp. 89 479–89 492, 2024

2024
[15]

TELeR: A general taxonomy of LLM prompts for benchmarking complex tasks,

K. Santuet al., “TELeR: A general taxonomy of LLM prompts for benchmarking complex tasks,” inFindings Assoc. Comput. Linguist.: EMNLP, 2023, pp. 14 197–14 203

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [3]

Synchronizing UA V teams for timely data collection and energy transfer by deep reinforcement learning,

O. S. Oubbatiet al., “Synchronizing UA V teams for timely data collection and energy transfer by deep reinforcement learning,”IEEE Trans. Veh. Technol., vol. 71, no. 6, pp. 6682–6697, 2022

work page 2022

[2] [1]

Unauthorized UA V countermeasure for low-altitude econ- omy: Joint communications and jamming based on MIMO cellular systems,

Z. Liet al., “Unauthorized UA V countermeasure for low-altitude econ- omy: Joint communications and jamming based on MIMO cellular systems,”IEEE Internet Things J., vol. 12, no. 6, pp. 6659–6672, 2025

work page 2025

[3] [2]

Multi-UA V-enabled energy-efficient data delivery for low-altitude economy: Joint coded caching, user grouping, and UA V deployment,

Q. Weiet al., “Multi-UA V-enabled energy-efficient data delivery for low-altitude economy: Joint coded caching, user grouping, and UA V deployment,”IEEE Internet Things J., pp. 1–1, 2025

work page 2025

[4] [4]

Survey on large language model-enhanced reinforcement learning: Concept, taxonomy, and methods,

Y . Caoet al., “Survey on large language model-enhanced reinforcement learning: Concept, taxonomy, and methods,”IEEE Trans. Neural Netw. Learn. Syst., pp. 1–21, 2024

work page 2024

[5] [5]

Generative AI agents with large language model for satellite networks via a mixture of experts transmission,

R. Zhanget al., “Generative AI agents with large language model for satellite networks via a mixture of experts transmission,”IEEE J. Sel. Areas Commun., vol. 42, no. 12, pp. 3581–3596, 2024

work page 2024

[6] [6]

Chain-of-thought prompting elicits reasoning in large language models,

J. Weiet al., “Chain-of-thought prompting elicits reasoning in large language models,” inProc. NeurIPS, vol. 35, 2022, pp. 24 824–24 837

work page 2022

[7] [7]

Reward design with language models,

M. Kwonet al., “Reward design with language models,” inProc. ICLR, 2023

work page 2023

[8] [8]

History compression via language models in rein- forcement learning,

F. Paischeret al., “History compression via language models in rein- forcement learning,” inProc. ICML, 2022, pp. 17 156–17 185

work page 2022