Pith. sign in

REVIEW 5 major objections 5 minor 84 references

The paper claims that separating LLM agents into a supervised-fine-tuned sub-goal planner and an offline-to-online RL executor reduces objective drift, raises long-horizon task success, and lowers inference token use across three interactiv

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A hierarchical LLM agent with an SFT-trained subgoal planner and an offline-to-online RL-trained action executor improves long-horizon task performance and token efficiency across three interactive text environments.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A genuinely useful architectural recipe with strong ablations, but its own Table 1 contradicts the 'consistently outperforms' claim and the completion-flag design is under-analyzed. the 5 major comments →

arxiv 2606.03698 v2 pith:SLHGFWK4 submitted 2026-06-02 cs.LG

Multi$^2$: Hierarchical Multi-Agent Decision-Making with LLM-Based Agents in Interactive Environments

Sangeun Park , Minhae Kwon This is my paper
classification cs.LG
keywords hierarchical decision-makingLLM agentsoffline-to-online reinforcement learningobjective drifttoken efficiencyrole specializationmulti-agent systemsbenchmark datasets
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Multi2 argues that long-horizon LLM agents drift because planning and execution are entangled in one context-driven policy, so execution errors compound without correction. It separates them: a fine-tuned planner (System 1) proposes short sub-goals, and an RL-trained executor (System 2) performs atomic actions, with control returning to the planner only when a sub-goal completes. The paper reports that this split outperforms prompt-based and fine-tuned hierarchical baselines on ScienceWorld, ALFWorld, and TextCraft, with the largest margins on longer, harder, and out-of-distribution tasks, and that it reduces inference tokens by invoking the planner on demand. The paper also releases role-specific hierarchical datasets to support training and evaluation of such agents.

Core claim

Multi2 claims that learning-based role specialization—not just hierarchical prompting—stabilizes long-horizon interaction. System 1, a LoRA planner, is SFT-trained to map task-plus-observation to a sub-goal; System 2, a LoRA executor, is trained by policy-anchored offline RL then KL-regularized online RL to map sub-goal-plus-observation to an atomic action and a completion flag. Control returns to System 1 only when the flag fires, saving tokens. Across three environments and three backbones, this design beats prompt-based and fine-tuned hierarchical baselines, with the biggest gains on hard and out-of-distribution tasks; ablations credit the role split, separate adapters, and the two loss t

What carries the argument

The load-bearing mechanism is the two-level control loop with selective invocation: the planner π_φ (SFT loss) proposes sub-goals; the executor π_θ, trained with an offline loss that mixes log-likelihood imitation with a policy-anchored advantage term and an online loss with KL regularization to the offline policy, outputs atomic actions plus a learned completion flag (Algorithm 1, line 15). When the flag is true, System 1 is re-invoked for the next sub-goal; on-demand invocation gives the token-efficiency gain. The two roles share a backbone but keep separate LoRA adapters.

Load-bearing premise

The hierarchy hinges on System 2's learned boolean completion flag: if it fires too early or too late, System 1 is invoked at the wrong moment and the claimed gains vanish, yet the paper reports no accuracy, calibration, or error-propagation analysis for this flag (Algorithm 1, line 15; System 2 prompt, Appendix A.2).

What would settle it

Measure the accuracy of System 2's completion flag against ground-truth sub-goal completion on a sample of rollouts from ScienceWorld, ALFWorld, and TextCraft; then re-run Multi2 with the flag replaced by ground-truth completion. If performance is unchanged, the learned flag is not the load-bearing component; if flag errors are common and predict task failure, the paper's central mechanism is unvalidated.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Multi2 outperforms all compared baselines on most splits of ScienceWorld, ALFWorld, and TextCraft; for example, with Llama-3.1 8B, ScienceWorld ID success reaches 67.61% versus 60.48% for the strongest hierarchical baseline, and TextCraft success rises from 9.50% to 35.60%.
  • The performance gap widens with task difficulty and on out-of-distribution splits, which the paper interprets as evidence that the design specifically mitigates objective drift rather than merely fitting the training distribution.
  • Multi2 is markedly more token-efficient: on ScienceWorld with Llama-3.1 8B, normalized performance per token is roughly 4.5x ReAct on the ID split and 13.8x on the OOD split.
  • Planner-generated sub-goals are a minor failure source in the paper's own analysis (3.7–5.9% of ALFWorld failed episodes), which the authors take as evidence that focusing online adaptation on the executor is well placed.
  • The released hierarchical datasets—sub-goal pairs for the planner and sub-goal-conditioned transitions for the executor—provide reproducible, rule-based training data for future hierarchical LLM agents.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The learned completion flag is the silent hinge: no accuracy or calibration analysis of this flag is reported, so a natural test is to measure flag error against ground-truth sub-goal completion and see whether errors correlate with task failures.
  • Because System 1 is kept fixed during online RL, the framework implicitly bets that planning is largely solved, a bet supported by the paper's own failure analysis; this points future work toward executor-side recovery and adaptation rather than better planners.
  • The training data are distilled from a larger teacher model using deterministic rules; if those teacher trajectories contain systematic suboptimalities, they set a ceiling on both systems, which the current evaluation does not separate out.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes Multi2, a hierarchical two-agent framework for long-horizon interactive LLM agents. System 1 is an SFT-trained high-level planner that emits sub-goals; System 2 is a low-level executor trained with an offline-to-online RL objective combining policy-anchored offline imitation, expectile value regression, and KL-regularized online refinement. The authors introduce and release three hierarchical datasets derived from ScienceWorld, ALFWorld, and TextCraft, and report experiments across three backbone LLMs (Qwen-2.5 3B, Mistral 7B, Llama-3.1 8B) against prompt-based and fine-tuning-based baselines. The central claims are that Multi2 consistently outperforms strong baselines, improves token efficiency, mitigates objective drift, and that the released datasets fill a gap in hierarchical LLM-agent benchmarking.

Significance. If the empirical claims hold, the contribution is meaningful: the role-specialized split between SFT-based planning and offline-to-online RL-based execution is a natural and potentially general recipe for long-horizon agentic tasks, and the released datasets and code would support reproducible follow-up work. The ablations (Figures 6–7, Tables 8–10) generally support the role-specialized design and the value of separate adapters. However, the headline 'consistently outperforms' claim is contradicted by the paper's own Table 1 on the Llama-3.1 8B ScienceWorld OOD split, and the absence of repeated-seed statistics or significance tests makes it impossible to tell whether the observed reversal is noise. Because the paper's main contribution is an empirical performance claim, this inconsistency is load-bearing and must be resolved before the result can be accepted as stated.

major comments (5)
  1. [Abstract, §5.3, Table 1] The claim that Multi2 'consistently outperforms strong agentic baselines' is directly contradicted by Table 1: on Llama-3.1 8B, ScienceWorld OOD, Multi2 achieves 30.68% while Glider achieves 34.36%. This is the only environment/backbone cell where a baseline wins, but no confidence intervals, number of seeds, or significance tests are reported anywhere in the paper or the appendix. Since the headline is an empirical claim, this reversal cannot be dismissed as sampling noise without statistics. Please add repeated-seed results with variance/significance tests, or qualify the abstract and conclusion to 'most splits' / 'on average.'
  2. [§5.1, Table 1, Appendix E.1] Reflexion is evaluated at pass@6 while all other methods, including Multi2, are evaluated at pass@1. The paper explicitly states 'we adopt a strict pass@1 metric' in §5.1, but Reflexion is allowed six attempts, which inflates its results (e.g., ALFWorld Qwen-2.5 ID 37.14 vs. ReAct 6.72). This makes the comparison in Table 1 unfair and undermines the 'consistently outperforms' claim. Report Reflexion at pass@1 as well, or separate multi-trial methods from the single-trial comparison.
  3. [Algorithm 1 line 15; Appendix A.2] The boolean completion flag emitted by System 2 decides when to re-invoke System 1. The paper provides no accuracy analysis, calibration, or sensitivity study for this learned flag. If it fires too early or too late, the hierarchical benefit disappears. Appendix G.4 reports that 'planner-related failures' are rare on ALFWorld, but that is an end-to-end failure attribution, not a measurement of flag precision/recall against the ground-truth sub-goal boundaries. Please report flag accuracy on each environment, or otherwise show that the system is robust to flag errors.
  4. [Appendix D.3, Figures 5 and 9] Task difficulty is defined by the average interaction length of the proposed agent itself. Grouping tasks into Easy/Medium/Hard by Multi2's own horizon makes the difficulty stratification agent-dependent: a baseline that stalls early or loops may land in a different difficulty bin, biasing the 'objective drift robustness' comparison in Figure 5 and the online-adaptation analysis in Figure 9. Define difficulty from task structure (e.g., number of sub-goals, expert-trace length, or environment affordances) that is independent of any evaluated policy, or verify that the grouping is stable across methods.
  5. [Eq. (6), §4.2] The policy-anchored advantage term λA(o_t, π_off(a_t|o_t; g_h); g_h) is not fully specified. It is unclear whether the action inside the advantage is sampled from the current policy or taken from the dataset; if the former, the expectation in (6) is an on-policy term inside an offline objective and requires justification. The text says this 'mitigates over-imitation' and 'improves transfer', but no derivation or analysis is given. Since this term is one of the paper's claimed contributions, please clarify the computation and provide either a derivation or an ablation that isolates its mechanism.
minor comments (5)
  1. [Abstract / Appendix A.4] The project page URL is inconsistent: the abstract gives park-sangeun.github.io/Multi-Square, while Appendix A.4 gives anonymous-projectpage.github.io/Multi-Square. For a double-blind submission, the anonymity-preserving link should be used consistently.
  2. [§3.2] The transition function is written as T: S × A → S; in a POMDP with stochastic transitions it should be S × A → Δ(S) or S × A × S → [0,1]. Also 'temporal discounted factor' should be 'discounted factor.'
  3. [Table 4] The hyperparameter table lists β, λ/α, and η, but the text of Eq. (6)–(7) does not define how β and λ/α are scheduled or whether they are tuned per environment. A sentence on sensitivity to these coefficients would help reproducibility.
  4. [§5.5, Figure 4] The token-efficiency definition is 'performance/tokens', but Figure 4 does not show absolute token counts; the bubble-size normalization is only relative to ReAct. Reporting mean token counts for each method (as done in Table 7) in the main text would make the efficiency claim more transparent.
  5. [Table 1 caption] The caption says 'Cyan highlights the best result,' but the table as reproduced has no color. Please use boldface or a visible marker so the best entries are identifiable in print.

Circularity Check

1 steps flagged

Central RL derivation is self-contained; one self-referential task-difficulty analysis inflates the robustness claim.

specific steps
  1. self definitional [Appendix D.3 (Task Difficulty), used in Section 5.6 / Figure 5]
    "We define task difficulty in ScienceWorld based on the average interaction length required by our agent to solve each task. Intuitively, tasks that require more environment steps tend to involve long-horizon decision-making and provide more probabilities for compounding errors. Concretely, for each task, we run our agent across multiple evaluation episodes and compute the average number of environment steps taken until termination (success or failure)."

    The 'hard task' groups are constructed from Multi2's own average rollout length, so the independent variable in the objective-drift analysis is an output of the proposed method. Section 5.6 then claims 'Multi2 remains consistently strong, thereby widening the gap on harder tasks,' but a task is labeled hard precisely when Multi2 tends to interact with it for many steps. Comparing baselines on groups defined by the proposed agent's behavior cannot independently establish robustness to task difficulty or objective drift; the grouping bakes in part of the claimed effect. This does not affect the main performance table or the RL loss derivation, but it makes the robustness analysis self-referential.

full rationale

The core derivation is not circular. System 1's SFT objective (Eq. 3) is standard behavior cloning on the high-level dataset; System 2's offline losses (Eqs. 4-6) are recognizable TD/expectile/advantage-weighted actor-critic updates using environment rewards and dataset transitions; the online update (Eq. 7) is a standard AWAC-style weighted maximum-likelihood term with KL regularization toward the offline policy. None of these equations reduces to a fitted parameter renamed as a prediction, and no load-bearing claim depends on a self-citation chain: the many citations to the authors' prior work are contextual related-work references, not justifications of the framework's validity. The one genuine self-referential element is the task-difficulty stratification in Appendix D.3, which defines difficulty by the proposed agent's own interaction length and then uses that stratification in Section 5.6 to argue for objective-drift robustness. This affects an analysis figure rather than the central Table 1 comparison, and the main claim is still independently testable, so the overall circularity score is low.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central pipeline rests on the quality of parsed GPT-4 trajectories, the accuracy of System 2's completion flag, and hand-set RL hyperparameters. No new physical or conceptual entities are postulated. The most fragile input is the learned completion flag, because it controls the planner-executor handoff and is not evaluated independently.

free parameters (5)
  • Offline imitation strength beta = 10
    Table 4; weights the exp(beta*A) term in Eq. 6. No sensitivity analysis is reported.
  • Regularization coefficients lambda/alpha = 7 / 7
    Table 4; used in Eqs. 6 and 7 to balance advantage weighting and policy anchoring.
  • Online KL coefficient eta = 0.02
    Table 4; controls how strongly the online policy is anchored to the offline policy in Eq. 7.
  • Expectile tau = 0.7
    Table 4; used in the expectile regression for the value function in Eq. 5.
  • Training budgets = SFT 10 epochs; offline RL 20 epochs; online RL 3500 steps
    Table 4; chosen without stated search; the claimed online self-improvement depends on these budgets.
axioms (6)
  • domain assumption GPT-4-Turbo trajectories from AgentGym are reliable expert demonstrations for both subgoal and action supervision.
    Appendix A builds Dsys1 and Dsys2 from these trajectories; no independent quality validation is provided.
  • domain assumption Deterministic rule-based sub-goal extraction produces correct supervision labels.
    Appendices A and A.4; all SFT and offline RL labels depend on the parser being correct.
  • domain assumption System 2's learned boolean completion flag accurately detects sub-goal achievement.
    Algorithm 1 line 15 and System 2 prompt in A.2; the planner handoff depends on this signal, and no accuracy analysis is given.
  • domain assumption The ID/OOD split at the task-template level prevents training leakage.
    Appendix D.1; the refined ScienceWorld split checks task specifications against the training corpus, but leakage through shared subgoals or actions is not analyzed.
  • standard math Standard RL assumptions (POMDP, TD learning, expectile regression) transfer to token-level LLM policies with LoRA adapters.
    Section 3.2; no convergence or coverage guarantees are established for the LLM action distribution.
  • domain assumption Internal sub-goal shaping rewards in ALFWorld and TextCraft are consistent with true task completion.
    Appendix A.3; the 1/0 shaped reward trains System 2, and a wrong shaping signal would optimize the executor toward the wrong objective.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi$^2$: Hierarchical Multi-Agent Decision-Making with LLM-Based Agents in Interactive Environments." pith.science (2026). https://pith.science/paper/SLHGFWK4

@misc{pith2026260603698,
  author       = {Pith},
  title        = {Pith review of: Multi$^2$: Hierarchical Multi-Agent Decision-Making with LLM-Based Agents in Interactive Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SLHGFWK4}},
  note         = {Machine review of arXiv:2606.03698}
}
Share X Bluesky LinkedIn Reddit HN
abstract

A central goal of large language model (LLM) research is to build agentic systems that can plan, act, and adapt through sustained interaction with dynamic environments. While recent LLM-based agents exhibit impressive contextual reasoning, their long-horizon decision-making remains fragile, often suffering from objective drift, where goals and plans drift over extended interactions. We introduce Multi$^2$, a hierarchical multi-agent decision-making framework that explicitly decomposes agent behavior into complementary roles. A high-level agent (System 1) focuses on context-aware sub-goal generation using supervised fine-tuning (SFT), while a low-level agent (System 2) executes atomic actions through offline-to-online reinforcement learning (RL) in interactive environments. This separation enables stable long-horizon control, mitigates objective drift, and allows efficient adaptation. Across diverse interactive environments, Multi$^2$ consistently outperforms strong agentic baselines, demonstrating improved robustness and coordination in multi-turn interaction. Beyond performance, we introduce and release three hierarchical benchmark datasets, filling a long-standing gap in training and evaluating hierarchical decision-making for LLM-based agents.

Figures

Figures reproduced from arXiv: 2606.03698 by Minhae Kwon, Sangeun Park.

Figure 1
Figure 1. Figure 1: Challenges in long-horizon interaction on the Science￾World using Llama-3.1 8B backbone. We compare ReAct [14] (prompt-based, non-hierarchical), Glider [15] (hierarchical fine￾tuning baseline), and our method (Multi2 ). (a) Horizon Robust￾ness: performance as a function of horizon length, where longer horizons induce larger degradation for baselines. (b) Token Ef￾ficiency: normalized inference-time token e… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Multi2 . (a) Offline Training: System 1 (high-level agent) is trained via SFT on the high-level dataset Dsys1 to generate sub-goals, while System 2 (low-level agent) is initialized via offline RL on the low-level dataset Dsys2 using an actor-critic objective. (b) Online Training: starting from the offline policy, System 2 continues self-improvement through interaction, updating its policy while… view at source ↗
Figure 3
Figure 3. Figure 3: Representative failure examples of baseline agents on the ScienceWorld Find-a-plant task with Qwen-2.5 3B. Each column corresponds to a different method (a–f). Shaded boxes indicate action types: gray denotes valid actions, blue denotes invalid actions rejected by the environment, and yellow denotes unproductive loops that repeat actions without measurable progress [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Inference-time token efficiency on ScienceWorld ID split with Llama-3.1 8B. The x-axis denotes token usage, and the y-axis denotes performance. Bubble size indicates token efficiency, computed as performance/tokens and normalized such that ReAct is 1.0. Solid lines: fine-tuned-based; dashed lines: prompt-based. tasks. Here, ADaPT outperforms GRPO on most tasks be￾cause its hierarchical prompting provides a… view at source ↗
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation results on the ScienceWorld ID split using Llama-3.1 8B (a) and Qwen-2.5 3B (b). (a) Offline Loss Design: Vanilla-IQL (without policy-anchored advantage term) vs. the pro￾posed offline objective. (b) Online Loss Design: Vanilla-AWAC (without KL regularization) vs. the proposed online objective. ing is unstable for high-level planning. Only SFT achieves competitive median performance but is less ro… view at source ↗
Figure 8
Figure 8. Figure 8: Effect of backbone model scale on Multi2 perfor￾mance in ScienceWorld with Qwen-2.5 backbones at three scales (1.5B, 3B, and 7B). (a) In-Distribution Split and (b) Out-of￾Distribution Split. Error bars denote one standard deviation over runs. results support the importance of separate adapters for ef￾fective role specialization of System 1 and System 2. Loss Function Designs [PITH_FULL_IMAGE:figures/full_… view at source ↗
Figure 10
Figure 10. Figure 10: Training dataset example for System 1 and System 2 on ScienceWorld benchmark. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Training dataset example for System 1 and System 2 on ALFWorld benchmark. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Training dataset example for System 1 and System 2 on TextCraft benchmark. Figures 10, 11, and 12 show training dataset examples for each benchmark (ScienceWorld, ALFWorld, and TextCraft). We formulate all environments as partially observable Markov decision processes and use separate prompt templates for System 1 and System 2 to match their input–output requirements. Each System 1 instance in Dsys1 pairs… view at source ↗
Figure 13
Figure 13. Figure 13: Input-output examples of Multi2 on ScienceWorld. (a) System 1 takes the task description and the current state summary (group_action and observation) and generates a sub-goal. (b) System 2 conditions on the sub-goal and the current observation to output an atomic action and a completion flag, repeating until the sub-goal is marked done. System 1 (planner). In the shown episode (task: boil water) in [PITH… view at source ↗
Figure 14
Figure 14. Figure 14: Input-output examples of Multi2 on ALFWorld. (a) System 1 takes the task description and the current state summary (group_action and observation) and generates a sub-goal. (b) System 2 conditions on the sub-goal and the current observation to output an atomic action and a completion flag, repeating until the sub-goal is marked done. System 1 (planner). As shown in [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Input-output examples of Multi2 on TextCraft. (a) System 1 takes the task description and the current state summary (group_action and observation) and generates a sub-goal. (b) System 2 conditions on the sub-goal and the current observation to output an atomic action and a completion flag, repeating until the sub-goal is marked done. System 1 (planner). As shown in [PITH_FULL_IMAGE:figures/full_fig_p025_… view at source ↗
Figure 16
Figure 16. Figure 16: Task difficulty based on average interaction length in the ScienceWorld benchmark. For each task, we compute the average number of environment steps taken by our agent across evaluation episodes, and group tasks into Easy/Medium/Hard by quantiles of this length (short/medium/long horizon) [PITH_FULL_IMAGE:figures/full_fig_p028_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Token efficiency at inference time on ScienceWorld benchmark OOD split with the Llama-3.1 8B backbone. The x-axis denotes token usage, and the y-axis denotes performance. Bubble size is proportional to token efficiency, computed as performance/tokens, and the number inside each bubble denotes normalized token efficiency with ReAct set to 1.0. Solid outlines denote fine-tuning-based methods, while dashed o… view at source ↗
Figure 18
Figure 18. Figure 18: Model scaling trends on ALFWorld benchmark across Qwen-2.5 model scales. (a) In-Distribution Split and (b) Out-of￾Distribution Split [PITH_FULL_IMAGE:figures/full_fig_p033_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Relative improvement rate from online adaptation on ScienceWorld, stratified by task difficulty for (a) In-Distribution Split and (b) Out-of-Distribution Split, measured relative to the offline-only baseline [PITH_FULL_IMAGE:figures/full_fig_p034_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Training curve of the offline-to-online RL stage on ScienceWorld. In [PITH_FULL_IMAGE:figures/full_fig_p035_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Test case result of ReAct, evaluated on the ScienceWorld Find-a-plant task. 36 [PITH_FULL_IMAGE:figures/full_fig_p036_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Test case result of Reflexion, evaluated on the ScienceWorld Find-a-plant task. 37 [PITH_FULL_IMAGE:figures/full_fig_p037_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Test case result of ADaPT, evaluated on the ScienceWorld Find-a-plant task. 38 [PITH_FULL_IMAGE:figures/full_fig_p038_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Test case result of GRPO, evaluated on the ScienceWorld Find-a-plant task. 39 [PITH_FULL_IMAGE:figures/full_fig_p039_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Test case result of Glider, evaluated on the ScienceWorld Find-a-plant task. 40 [PITH_FULL_IMAGE:figures/full_fig_p040_25.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

84 extracted references · 5 linked inside Pith

  1. [1]

    En- hancing decision-making of large language models via actor-critic

    Heng Dong, Kefei Duan, and Chongjie Zhang. En- hancing decision-making of large language models via actor-critic. InInternational Conference on Machine Learning (ICML), 2025

    2025

  2. [2]

    CollabLLM: From passive responders to active collaborators

    Shirley Wu, Michel Galley, Baolin Peng, Hao Cheng, Gavin Li, Yao Dou, Weixin Cai, James Zou, Jure Leskovec, and Jianfeng Gao. CollabLLM: From passive responders to active collaborators. InInter- national Conference on Machine Learning (ICML), 2025

    2025

  3. [3]

    DAMA: Data- and model-aware align- ment of multi-modal LLMs

    Jinda Lu, Junkang Wu, Jinghan Li, Xiaojun Jia, Shuo Wang, YiFan Zhang, Junfeng Fang, Xiang Wang, and Xiangnan He. DAMA: Data- and model-aware align- ment of multi-modal LLMs. InInternational Confer- ence on Machine Learning (ICML), 2025

    2025

  4. [4]

    Inverse rational control with par- tially observable continuous nonlinear dynamics

    Minhae Kwon, Saurabh Daptardar, Paul R Schrater, and Xaq Pitkow. Inverse rational control with par- tially observable continuous nonlinear dynamics. In Advances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  5. [5]

    QuBE: Question-based belief enhancement for agentic LLM reasoning

    Minsoo Kim, Jongyoon Kim, Jihyuk Kim, and Seung Hwang. QuBE: Question-based belief enhancement for agentic LLM reasoning. InEmpirical Methods in Natural Language Processing (EMNLP), 2024

    2024

  6. [6]

    Agentic reasoning: A streamlined frame- work for enhancing LLM reasoning with agentic tools

    Junde Wu, Jiayuan Zhu, Yuyuan Liu, Min Xu, and Yueming Jin. Agentic reasoning: A streamlined frame- work for enhancing LLM reasoning with agentic tools. InAssociation for Computational Linguistics (ACL), 2025

    2025

  7. [7]

    T1: Advancing language model reasoning through reinforcement learning and inference scaling

    Zhenyu Hou, Xin Lv, Rui Lu, Jiajie Zhang, Yujiang Li, Zijun Yao, Juanzi Li, Jie Tang, and Yuxiao Dong. T1: Advancing language model reasoning through reinforcement learning and inference scaling. InIn- ternational Conference on Machine Learning (ICML), 2025

    2025

  8. [8]

    Episodic future think- ing mechanism for multi-agent reinforcement learning

    Dongsu Lee and Minhae Kwon. Episodic future think- ing mechanism for multi-agent reinforcement learning. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  9. [9]

    ReCAP: Recursive context-aware reasoning and planning for large language model agents

    Zhenyu Zhang, Tianyi Chen, Weiran Xu, Alex Pent- land, and Jiaxin Pei. ReCAP: Recursive context-aware reasoning and planning for large language model agents. InAdvances in Neural Information Processing Systems (NeurIPS), 2025

    2025

  10. [10]

    Evaluating LLM-based agents for multi-turn conversations: A survey.arXiv preprint arXiv:2503.22458, 2025

    Shengyue Guan, Jindong Wang, Jiang Bian, Bin Zhu, Jian Lou, and Haoyi Xiong. Evaluating LLM-based agents for multi-turn conversations: A survey.arXiv preprint arXiv:2503.22458, 2025

    arXiv 2025

  11. [11]

    Path drift in large reasoning models: How first-person commitments override safety

    Yuyi Huang, Runzhe Zhan, Lidia Chao, Ailin Tao, and Derek Wong. Path drift in large reasoning models: How first-person commitments override safety. In Empirical Methods in Natural Language Processing (EMNLP), 2025

    2025

  12. [12]

    Drift no more? Context equilibria in multi-turn LLM interac- tions

    Vardhan Dongre, Ryan Rossi, Viet Lai, Seunghyun Yoon, Dilek Hakkani-Tür, and Trung Bui. Drift no more? Context equilibria in multi-turn LLM interac- tions. InAAAI Personalization in the Era of Large Foundation Models Workshop, 2025

    2025

  13. [13]

    Do as I can, not as I say: Grounding language in robotic affordances

    Brian Ichter, Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, Dmitry Kalashnikov, Sergey Levine, Yao Lu, Carolina Parada, Kanishka Rao, Pierre Sermanet, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Mengyuan Yan, Noah Brown, Michael Ahn, Omar Co...

    2023

  14. [14]

    ReAct: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models. InInternational Conference on Learning Representa- tions (ICLR), 2023

    2023

  15. [15]

    Divide and conquer: Grounding LLMs as efficient decision-making agents via offline hierarchical reinforcement learning

    Zican Hu, Wei Liu, Xiaoye Qu, Xiangyu Yue, Chunlin Chen, Zhi Wang, and Yu Cheng. Divide and conquer: Grounding LLMs as efficient decision-making agents via offline hierarchical reinforcement learning. InIn- ternational Conference on Machine Learning (ICML), 2025. 10 Multi2: Hierarchical Multi-Agent Decision-Making with LLM-Based Agents in Interactive Environments

    2025

  16. [16]

    ADaPT: As-needed decomposition and planning with language models

    Archiki Prasad, Alexander Koller, Mareike Hartmann, Peter Clark, Ashish Sabharwal, Mohit Bansal, and Tushar Khot. ADaPT: As-needed decomposition and planning with language models. InFindings of the Association for Computational Linguistics (NAACL), 2024

    2024

  17. [17]

    Plan-and-Act: Improving planning of agents for long-horizon tasks

    Lutfi Erdogan, Hiroki Furuta, Sehoon Kim, Nicholas Lee, Suhong Moon, Gopala Anumanchipalli, Kurt Keutzer, and Amir Gholami. Plan-and-Act: Improving planning of agents for long-horizon tasks. InInter- national Conference on Machine Learning (ICML), 2025

    2025

  18. [18]

    The illusion of diminishing returns: Measuring long horizon execution in LLMs

    Akshit Sinha, Arvindh Arun, Shashwat Goel, Steffen Staab, and Jonas Geiping. The illusion of diminishing returns: Measuring long horizon execution in LLMs. InInternational Conference on Learning Representa- tions (ICLR), 2026

    2026

  19. [19]

    Agent-oriented planning in multi-agent systems

    Ao Li, Yuexiang Xie, Songze Li, Fugee Tsung, Bolin Ding, and Yaliang Li. Agent-oriented planning in multi-agent systems. InInternational Conference on Learning Representations (ICLR), 2025

    2025

  20. [20]

    Multi-agent collaboration via evolv- ing orchestration

    Yufan Dang, Chen Qian, Xueheng Luo, Jingru Fan, Zi- hao Xie, Ruijie Shi, Weize Chen, Cheng Yang, Xiaoyin Che, Ye Tian, Xuantang Xiong, Lei Han, Zhiyuan Liu, and Maosong Sun. Multi-agent collaboration via evolv- ing orchestration. InAdvances in Neural Information Processing Systems (NeurIPS), 2025

    2025

  21. [21]

    Agentic AI: The age of reason- ing—A review.Journal of Automation and Intelli- gence, 2025

    Ume Nisa, Muhammad Shirazi, Mohamed Saip, and Muhammad Pozi. Agentic AI: The age of reason- ing—A review.Journal of Automation and Intelli- gence, 2025

    2025

  22. [22]

    ScienceWorld: Is your agent smarter than a 5th grader? InEmpirical Methods in Natural Language Processing (EMNLP), 2022

    Ruoyao Wang, Peter Jansen, Marc Côté, and Prithviraj Ammanabrolu. ScienceWorld: Is your agent smarter than a 5th grader? InEmpirical Methods in Natural Language Processing (EMNLP), 2022

    2022

  23. [23]

    OASIS: Open-world adaptive self-supervised and imbalanced- aware system

    Miru Kim, Mugon Joe, and Minhae Kwon. OASIS: Open-world adaptive self-supervised and imbalanced- aware system. InACM International Conference on Information and Knowledge Management (CIKM), 2025

    2025

  24. [24]

    Improving network attack classification on imbalanced real-world intrusion incident datasets

    Miru Kim, Mugon Joe, and Minhae Kwon. Improving network attack classification on imbalanced real-world intrusion incident datasets. InInternational Confer- ence on Mobile Systems, Applications and Services (MobiSys), 2025

    2025

  25. [25]

    Con- trastive learning based network attack classifier for imbalanced data.Journal of Communications and Networks, 28(1):86–97, Feb

    Mugon Joe, Miru Kim, and Minhae Kwon. Con- trastive learning based network attack classifier for imbalanced data.Journal of Communications and Networks, 28(1):86–97, Feb. 2026

    2026

  26. [26]

    Per- sonalized split federated learning with early exit: Pre- training and online learning against label shifts.IEEE Internet of Things Journal, 12(22):47069–47082, Nov

    Miru Kim, Heewon Park, and Minhae Kwon. Per- sonalized split federated learning with early exit: Pre- training and online learning against label shifts.IEEE Internet of Things Journal, 12(22):47069–47082, Nov. 2025

    2025

  27. [27]

    Fed-ADE: Adaptive learning rate for federated post-adaptation under distribution shift

    Heewon Park, Mugon Joe, Miru Kim, Kyungjin Im, and Minhae Kwon. Fed-ADE: Adaptive learning rate for federated post-adaptation under distribution shift. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026

    2026

  28. [28]

    Evolving intelligent network attack classifier under label distri- bution shift.IEEE Transactions on Network Science and Engineering, 13(1):7448–7464, Mar

    Miru Kim, Mugon Joe, and Minhae Kwon. Evolving intelligent network attack classifier under label distri- bution shift.IEEE Transactions on Network Science and Engineering, 13(1):7448–7464, Mar. 2026

    2026

  29. [29]

    Personal- ized federated sensing for heterogeneous environment

    Heewon Park, Miru Kim, and Minhae Kwon. Personal- ized federated sensing for heterogeneous environment. IEEE Sensors Letters, 9(4):1–4, 2025

    2025

  30. [30]

    Editable scene simulation for autonomous driving via collaborative LLM-agents

    Yuxi Wei, Zi Wang, Yifan Lu, Chenxin Xu, Changx- ing Liu, Hao Zhao, Siheng Chen, and Yanfeng Wang. Editable scene simulation for autonomous driving via collaborative LLM-agents. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  31. [31]

    DriVLMe: Enhancing LLM- based autonomous driving agents with embodied and social experiences

    Yidong Huang, Jacob Sansom, Ziqiao Ma, Felix Gervits, and Joyce Chai. DriVLMe: Enhancing LLM- based autonomous driving agents with embodied and social experiences. InIEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS), 2024

    2024

  32. [32]

    SToRM: Supervised token reduction for multi-modal LLMs toward efficient end- to-end autonomous driving

    Seo Hyun Kim, Jin Bok Park, Do Yeon Koo, Hogun Park, and Il Yong Chun. SToRM: Supervised token reduction for multi-modal LLMs toward efficient end- to-end autonomous driving. InIEEE International Conference on Robotics and Automation (ICRA), 2026

    2026

  33. [33]

    ASAP: Unsupervised post-training with label distribution shift adaptive learning rate

    Heewon Park, Mugon Joe, Miru Kim, and Minhae Kwon. ASAP: Unsupervised post-training with label distribution shift adaptive learning rate. InACM Inter- national Conference on Information and Knowledge Management (CIKM), 2025

    2025

  34. [34]

    Reflexion: Lan- guage agents with verbal reinforcement learning

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Lan- guage agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  35. [35]

    Make your LLM fully utilize the context

    Shengnan An, Zexiong Ma, Zeqi Lin, Nanning Zheng, Jian Lou, and Weizhu Chen. Make your LLM fully utilize the context. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. 11 Multi2: Hierarchical Multi-Agent Decision-Making with LLM-Based Agents in Interactive Environments

    2024

  36. [36]

    Toward self- improvement of LLMs via imagination, searching, and criticizing

    Ye Tian, Baolin Peng, Linfeng Song, Lifeng Jin, Dian Yu, Haitao Mi, and Dong Yu. Toward self- improvement of LLMs via imagination, searching, and criticizing. InAdvances in Neural Information Pro- cessing Systems (NeurIPS), 2024

    2024

  37. [37]

    The lighthouse of language: Enhancing LLM agents via critique-guided improvement

    Ruihan Yang, Fanghua Ye, Jian Li, Siyu Yuan, Yikai Zhang, Zhaopeng Tu, Xiaolong Li, and Deqing Yang. The lighthouse of language: Enhancing LLM agents via critique-guided improvement. InAdvances in Neu- ral Information Processing Systems (NeurIPS), 2025

    2025

  38. [38]

    The alignment problem from a deep learning perspec- tive

    Richard Ngo, Lawrence Chan, and Sören Mindermann. The alignment problem from a deep learning perspec- tive. InInternational Conference on Learning Repre- sentations (ICLR), 2024

    2024

  39. [39]

    ArCHer: Training language model agents via hierarchical multi-turn RL

    Yifei Zhou, Andrea Zanette, Jiayi Pan, Sergey Levine, and Aviral Kumar. ArCHer: Training language model agents via hierarchical multi-turn RL. InInternational Conference on Machine Learning (ICML), 2024

    2024

  40. [40]

    Robust hierar- chical anomaly detection using feature impact in iot networks.ICT Express, 11(2):358–363, Apr

    Joohong Rheey and Hyunggon Park. Robust hierar- chical anomaly detection using feature impact in iot networks.ICT Express, 11(2):358–363, Apr. 2025

    2025

  41. [41]

    Option discovery us- ing LLM-guided semantic hierarchical reinforcement learning.arXiv preprint arXiv:2503.19007, 2025

    Chak Shek and Pratap Tokekar. Option discovery us- ing LLM-guided semantic hierarchical reinforcement learning.arXiv preprint arXiv:2503.19007, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    Leveraging imitation learning and LLMs for efficient hierarchical reinforcement learning

    Runhan Yang, Jieao Shi, Mengqi Su, and Don- gruo Zhou. Leveraging imitation learning and LLMs for efficient hierarchical reinforcement learning. https://openreview.net/forum?id=6y00rooi7i, 2025

    2025

  43. [43]

    Getting more juice out of the SFT data: Reward learning from human demonstration improves SFT for LLM alignment

    Jiaxiang Li, Siliang Zeng, Hoi Wai, Chenliang Li, Alfredo Garcia, and Mingyi Hong. Getting more juice out of the SFT data: Reward learning from human demonstration improves SFT for LLM alignment. In Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  44. [44]

    Large lan- guage models as generalizable policies for embodied tasks

    Andrew Szot, Max Schwarzer, Harsh Agrawal, Bog- dan Mazoure, Rin Metcalf, Walter Talbott, Natalie Mackraz, R Hjelm, and Alexander Toshev. Large lan- guage models as generalizable policies for embodied tasks. InInternational Conference on Learning Repre- sentations (ICLR), 2024

    2024

  45. [45]

    Unlocking LLMs’ self-improvement capacity with autonomous learning for domain adaptation

    Ke Ji, Junying Chen, Anningzhe Gao, Wenya Xie, Xiang Wan, and Benyou Wang. Unlocking LLMs’ self-improvement capacity with autonomous learning for domain adaptation. InFindings of the Association for Computational Linguistics (ACL), 2025

    2025

  46. [46]

    Data mix- ing optimization for supervised fine-tuning of large language models

    Yuan Li, Zhengzhong Liu, and Eric Xing. Data mix- ing optimization for supervised fine-tuning of large language models. InInternational Conference on Ma- chine Learning (ICML), 2025

    2025

  47. [47]

    Wei Lu, Rachel Luu, and Markus Buehler. Fine-tuning large language models for domain adaptation: Explo- ration of training strategies, scaling, model merging and synergistic capabilities.npj Computational Mate- rials, 11(1):84, 2025

    2025

  48. [48]

    Coevolving with the other you: Fine-tuning LLM with sequential coopera- tive multi-agent reinforcement learning

    Hao Ma, Tianyi Hu, Zhiqiang Pu, Boyin Liu, Xiaolin Ai, Yanyan Liang, and Min Chen. Coevolving with the other you: Fine-tuning LLM with sequential coopera- tive multi-agent reinforcement learning. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  49. [49]

    Instant inverse mod- eling of stochastic driving behavior with deep rein- forcement learning.IEEE Transactions on Consumer Electronics, 71(1):2152–2162, Feb

    Dongsu Lee and Minhae Kwon. Instant inverse mod- eling of stochastic driving behavior with deep rein- forcement learning.IEEE Transactions on Consumer Electronics, 71(1):2152–2162, Feb. 2025

    2025

  50. [50]

    Control- ling large language model with latent action

    Chengxing Jia, Ziniu Li, Pengyuan Wang, Yi Li, Zhenyu Hou, Yuxiao Dong, and Yang Yu. Control- ling large language model with latent action. InIn- ternational Conference on Machine Learning (ICML), 2025

    2025

  51. [51]

    Stability analysis in mixed-autonomous traffic with deep reinforcement learning.IEEE Transactions on Vehicular Technology, 72(3):2848–2862, Mar

    Dongsu Lee and Minhae Kwon. Stability analysis in mixed-autonomous traffic with deep reinforcement learning.IEEE Transactions on Vehicular Technology, 72(3):2848–2862, Mar. 2023

    2023

  52. [52]

    QLASS: Boost- ing language agent inference via Q-guided stepwise search

    Zongyu Lin, Yao Tang, Xingcheng Yao, Da Yin, Ziniu Hu, Yizhou Sun, and Kai Chang. QLASS: Boost- ing language agent inference via Q-guided stepwise search. InInternational Conference on Machine Learn- ing (ICML), 2025

    2025

  53. [53]

    Temporal distance- aware transition augmentation for offline model-based reinforcement learning

    Dongsu Lee and Minhae Kwon. Temporal distance- aware transition augmentation for offline model-based reinforcement learning. InInternational Conference on Machine Learning (ICML), 2025

    2025

  54. [54]

    Online reinforcement learning in stochastic games

    Chen Wei, Yi Hong, and Chi Lu. Online reinforcement learning in stochastic games. InAdvances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  55. [55]

    Efficient online reinforcement learning with offline data

    Philip Ball, Laura Smith, Ilya Kostrikov, and Sergey Levine. Efficient online reinforcement learning with offline data. InInternational Conference on Machine Learning (ICML), 2023

    2023

  56. [56]

    Leveraging offline data in online reinforcement learning

    Andrew Wagenmaker and Aldo Pacchiano. Leveraging offline data in online reinforcement learning. InIn- ternational Conference on Machine Learning (ICML), 2023

    2023

  57. [57]

    Continuous control with deep reinforcement learning

    Timothy Lillicrap, Jonathan Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. InInternational Conference on Learning Representations (ICLR), 2016. 12 Multi2: Hierarchical Multi-Agent Decision-Making with LLM-Based Agents in Interactive Environments

    2016

  58. [58]

    Foresighted decisions for inter-vehicle interactions: An offline reinforcement learning approach

    Dongsu Lee and Minhae Kwon. Foresighted decisions for inter-vehicle interactions: An offline reinforcement learning approach. InIEEE International Conference on Intelligent Transportation Systems (ITSC), 2023

    2023

  59. [59]

    Ad- dressing function approximation error in actor-critic methods

    Scott Fujimoto, Herke Hoof, and David Meger. Ad- dressing function approximation error in actor-critic methods. InInternational Conference on Machine Learning (ICML), 2018

    2018

  60. [60]

    Selec- tive imitation for efficient online reinforcement learn- ing with pre-collected data.ICT Express, 10(6):1308– 1314, Dec

    Chanin Eom, Dongsu Lee, and Minhae Kwon. Selec- tive imitation for efficient online reinforcement learn- ing with pre-collected data.ICT Express, 10(6):1308– 1314, Dec. 2024

    2024

  61. [61]

    Of- fline reinforcement learning with implicit Q-learning

    Ilya Kostrikov, Ashvin Nair, and Sergey Levine. Of- fline reinforcement learning with implicit Q-learning. InInternational Conference on Learning Representa- tions (ICLR), 2022

    2022

  62. [62]

    Price of the au- tonomous strategy with reinforcement learning in mixed-autonomy traffic networks.IEEE Transactions on Intelligent Transportation Systems, 27(2):2741– 2752, Feb

    Chanin Eom and Minhae Kwon. Price of the au- tonomous strategy with reinforcement learning in mixed-autonomy traffic networks.IEEE Transactions on Intelligent Transportation Systems, 27(2):2741– 2752, Feb. 2026

    2026

  63. [63]

    The impact of dataset on offline reinforcement learning performance in uav-based emergency network recov- ery tasks.IEEE Communications Letters, 28(5):1058– 1061, May

    Jeyeon Eo, Dongsu Lee, and Minhae Kwon. The impact of dataset on offline reinforcement learning performance in uav-based emergency network recov- ery tasks.IEEE Communications Letters, 28(5):1058– 1061, May. 2024

    2024

  64. [64]

    Curriculum reinforcement learning for cohesive team in mobile ad hoc networks.IEEE Communications Letters, 26(8):1809–1813, Aug

    Nayoung Kim, Minhae Kwon, and Hyunggon Park. Curriculum reinforcement learning for cohesive team in mobile ad hoc networks.IEEE Communications Letters, 26(8):1809–1813, Aug. 2022

    2022

  65. [65]

    AD4RL: Autonomous driving benchmarks for offline reinforcement learning with value-based dataset

    Dongsu Lee, Chanin Eom, and Minhae Kwon. AD4RL: Autonomous driving benchmarks for offline reinforcement learning with value-based dataset. In IEEE International Conference on Robotics and Au- tomation (ICRA), 2024

    2024

  66. [66]

    Episodic future thinking with offline reinforcement learning for au- tonomous driving.IEEE Internet of Things Journal, 12(11):17012–17023, Jun

    Dongsu Lee and Minhae Kwon. Episodic future thinking with offline reinforcement learning for au- tonomous driving.IEEE Internet of Things Journal, 12(11):17012–17023, Jun. 2025

    2025

  67. [67]

    A unified principle of pessimism for offline reinforce- ment learning under model mismatch

    Yue Wang, Zhongchang Sun, and Shaofeng Zou. A unified principle of pessimism for offline reinforce- ment learning under model mismatch. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  68. [68]

    Is value learning really the main bottleneck in offline RL? InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    Seohong Park, Kevin Frans, Sergey Levine, and Aviral Kumar. Is value learning really the main bottleneck in offline RL? InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  69. [69]

    Beyond online sampling: Bridging offline- to-online alignment via dynamic data transformation for LLMs

    Zhang Zhang, Guhao Feng, Jian Guan, Di He, and Wei Wu. Beyond online sampling: Bridging offline- to-online alignment via dynamic data transformation for LLMs. InEmpirical Methods in Natural Language Processing (EMNLP), 2025

    2025

  70. [70]

    DigiRL: Training in-the-wild device-control agents with au- tonomous reinforcement learning

    Hao Bai, Yifei Zhou, Mert Cemri, Jiayi Pan, Alane Suhr, Sergey Levine, and Aviral Kumar. DigiRL: Training in-the-wild device-control agents with au- tonomous reinforcement learning. InAdvances in Neu- ral Information Processing Systems (NeurIPS), 2024

    2024

  71. [71]

    Unpacking DPO and PPO: Disentangling best practices for learning from preference feedback

    Hamish Ivison, Yizhong Wang, Jiacheng Liu, Zeqiu Wu, Valentina Pyatkin, Nathan Lambert, Noah Smith, Yejin Choi, and Hannaneh Hajishirzi. Unpacking DPO and PPO: Disentangling best practices for learning from preference feedback. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  72. [72]

    Bridging offline and online re- inforcement learning for LLMs.arXiv preprint arXiv:2506.21495, 2025

    Jack Lanchantin, Angelica Chen, Janice Lan, Xian Li, Swarnadeep Saha, Tianlu Wang, Jing Xu, Ping Yu, Weizhe Yuan, Jason Weston, Sainbayar Sukhbaatar, and Ilia Kulikov. Bridging offline and online re- inforcement learning for LLMs.arXiv preprint arXiv:2506.21495, 2025

    Pith/arXiv arXiv 2025

  73. [73]

    Scenario-free au- tonomous driving with multi-task offline-to-online re- inforcement learning.IEEE Transactions on Intelli- gent Transportation Systems, 26(9):13317–13330, Sep

    Dongsu Lee and Minhae Kwon. Scenario-free au- tonomous driving with multi-task offline-to-online re- inforcement learning.IEEE Transactions on Intelli- gent Transportation Systems, 26(9):13317–13330, Sep. 2025

    2025

  74. [74]

    Test-time fine-tuning of image compression models for multi- task adaptability

    Unki Park, Seongmoon Jeong, Youngchan Jang, Gyeong-Moon Park, and Jong Hwan Ko. Test-time fine-tuning of image compression models for multi- task adaptability. InIEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2025

    2025

  75. [75]

    KL-regularized reinforce- ment learning is designed to mode collapse

    Anthony Chen, Jatin Prakash, Jeff Guo, Rob Fergus, and Rajesh Ranganath. KL-regularized reinforce- ment learning is designed to mode collapse. InIn- ternational Conference on Learning Representations (ICLR), 2026

    2026

  76. [76]

    KL-regularised Q-learning: A token- level action-value perspective on online RLHF

    Jason Brown, Lennie Wells, Edward Young, and Ser- gio Bacallado. KL-regularised Q-learning: A token- level action-value perspective on online RLHF. In ICML Workshop on Models of Human Feedback for AI Alignment, 2025

    2025

  77. [77]

    The choice of diver- gence: A neglected key to mitigating diversity collapse in reinforcement learning with verifiable reward

    Long Li, Jiaran Hao, Jason Liu, Zhijian Zhou, Yanting Miao, Wei Pang, Xiaoyu Tan, Wei Chu, Zhe Wang, Shirui Pan, Chao Qu, and Yuan Qi. The choice of diver- gence: A neglected key to mitigating diversity collapse in reinforcement learning with verifiable reward. In International Conference on Learning Representations (ICLR), 2026. 13 Multi2: Hierarchical M...

    2026

  78. [78]

    ALF- World: Aligning text and embodied environments for interactive learning

    Mohit Shridhar, Xingdi Yuan, Marc Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. ALF- World: Aligning text and embodied environments for interactive learning. InInternational Conference on Learning Representations (ICLR), 2021

    2021

  79. [79]

    Qwen2.5 technical report.arXiv preprint arXiv:2412.15115, 2025

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Ke- qin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tianyi...

    Pith/arXiv arXiv 2025

  80. [80]

    Mistral 7b

    Albert Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Chaplot, Diego Casas, Flo- rian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Lavaud, Marie Lachaux, Pierre Stock, Teven Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William Sayed. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023

    Pith/arXiv arXiv 2023

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.