Pith. sign in

REVIEW 4 major objections 7 minor 173 references

Credit assignment in multi-agent learning works better when agents are ranked by relative contribution instead of scored absolutely.

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 ↗

T0 review · grok-4.5

2026-07-31 21:52 UTC pith:PREIRM4I

load-bearing objection Solid composition paper: rank aggregation over LMM pairwise judgments for open-team embodied credit, with a real benchmark and honest limits—worth referee time if artifacts land. the 4 major comments →

arxiv 2607.27967 v1 pith:PREIRM4I submitted 2026-07-30 cs.AI

MARS-RA: Rank Aggregation for Credit Assignment via Multimodal Comparisons in Embodied Multi-Agent Cooperation

Dawei Wang , Di Zhao , Xinyuan Liu , Marci Chi Ma , Xiaoyang Liu , Chengming Zhou , Gary Ushaw , Richard Davison This is my paper
classification cs.AI
keywords multi-agent reinforcement learningcredit assignmentrank aggregationembodied AIlarge multimodal modelspotential-based reward shapingopen multi-agent systems
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.

In cooperative multi-agent reinforcement learning for embodied systems, agents often cannot tell how much each one contributed to a sparse, delayed team reward, especially when the set of active agents changes mid-task. This paper claims that the problem becomes tractable if credit is treated as a rank-aggregation task: a large multimodal model compares pairs of agents from their egocentric images, the comparisons are aggregated into contribution scores, and those scores shape rewards through a potential function. Relative comparisons tolerate noise and changing team size better than absolute scores. On a new open-team benchmark and on standard cooperative games, the method steers agents toward effective cooperation and beats strong baselines. Theory links the aggregated scores to latent preferences and, under ideal conditions, to Shapley values.

Core claim

Reformulating MARL credit assignment as Bradley–Terry rank aggregation over LMM-generated pairwise contribution comparisons yields per-agent scores that, when used as potential-based reward shaping, produce denser learning signals that remain valid under noise and dynamic agent participation, guiding effective cooperation on open embodied tasks.

What carries the argument

MARS-RA: LMM pairwise comparisons of active agents’ egocentric views are fit by a Bradley–Terry model to latent contribution scores, which are softmax-normalized into a potential and added as F = γψ(s′,t′) − ψ(s,t) shaping rewards without changing the optimal policy.

Load-bearing premise

A large multimodal model looking at agents’ camera views and a text prompt holds a stable, useful notion of who is helping the team more, so that noisy pairwise answers still concentrate on true relative contribution.

What would settle it

On MARS-Bench or similar open-team tasks, replace or corrupt the LMM comparisons so accuracy falls near chance while keeping the same aggregation and shaping pipeline; if success rates then collapse to the plain MAPPO baseline, the claim that relative LMM ranking supplies the useful credit signal is falsified.

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

If this is right

  • Open multi-robot teams with agents entering and leaving can receive dense credit signals without fixing the team size in advance.
  • Improving LMM pairwise judgment accuracy or query count should further raise cooperative success under the same training loop.
  • Potential-based shaping from aggregated ranks can be dropped onto existing policy-gradient MARL backbones without redesigning value factorization.
  • Shapley-style contribution semantics become an interpretive lens for LMM-derived scores even when exact Shapley computation is intractable.

Where Pith is reading between the lines

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

  • The same relative-comparison pipeline could supply online credit in human–robot teams where a vision-language model watches shared video rather than egocentric robot cameras.
  • If pairwise accuracy saturates, the remaining gains may come from better temporal aggregation across long horizons rather than more queries per step.
  • Benchmarks that force dynamic agent counts will pressure other credit methods that still assume a fixed factorization graph.

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

4 major / 7 minor

Summary. The paper proposes MARS-RA, which reformulates cooperative MARL credit assignment as Bradley–Terry rank aggregation over LMM-generated pairwise contribution comparisons on egocentric images, then injects the resulting scores via potential-based reward shaping (PBRS). It targets embodied settings with sparse/delayed feedback and a dynamically changing active set, introduces MARS-Bench (Pass Gate, Herd Sheep, Collect Ball in ManiSkill3/XLeRobot with battery-driven entry/exit), and supplies convergence/robustness analysis (Prop. 1) plus an interpretive Shapley link (Prop. 2). Empirically, MARS-RA (MAPPO backbone, Gemini-2.5-Pro comparisons) outperforms MAPPO, LMM Score, QMIX, COMA, SQDDPG, SAMA, and V-GEPF on MARS-Bench (often >47–70% success), is competitive on Overcooked and stronger on 10-agent Pistonball, with ablations on LMM accuracy/query count, Rank Centrality vs BT, ρ, and a 25-run real-robot Pass Gate transfer (~64%).

Significance. If the results hold, the work offers a practical and relatively clean interface between foundation multimodal models and MARL credit assignment: relative comparisons plus classical rank aggregation and policy-invariant PBRS, rather than absolute LMM scores or hand-designed factorizations. MARS-Bench usefully stresses open team size, partial egocentric vision, and sparse rewards—gaps relative to fixed-team embodied benchmarks. Strengths include multi-seed learning curves with CIs, several ablations (backbone, absolute LMM scores, aggregation method, ρ, LMM family/query count), generalization checks beyond the new bench, and a small real-robot deployment that does not require LMMs at test time. Prop. 1 is a standard MLE concentration restatement under a connected comparison graph; Prop. 2 is appropriately labeled interpretive. The main scientific value is empirical and systems-level rather than a new credit-assignment theorem.

major comments (4)
  1. [§8.4, Fig. 6, Table 4, Appendix E] §8.4, Fig. 6 (left), Table 4, and Appendix E define “pairwise comparison accuracy” solely as agreement with MARS-Bench’s dense per-agent reward and then treat that quantity as the quality of contribution judgment. Dense shaping rewards are simulator-internal progress signals, not counterfactual/Shapley marginal contributions, and are unavailable in the real embodied regime the introduction emphasizes. The reported accuracy–success correlation and the interpretation that better “contribution” comparisons drive gains are therefore only partially identified: an alternative reading is that the LMM acts as a noisy dense-reward oracle. The paper should either (i) replace or supplement this proxy with an independent contribution measure (e.g., leave-one-out return, approximate Shapley on logged rollouts, or human preference labels on held-out clips), or (ii) explicitly reframe claims as “LMM pr
  2. [§6.1–6.2, Abstract, §1] §6.1 Prop. 1 only bounds ‖ĉ_t − c*_t‖ under the assumption that the LMM implements a latent BT preference c*_t and that the comparison graph is connected; it does not show that c*_t equals true marginal contribution. §6.2 Prop. 2 is conditional (“if c*_t = v*_t”) and interpretive. The abstract and introduction still present “theoretical justification for … credit assignment” and robustness to embodied noise in a way that can be read as recovering true credits. Please separate cleanly: (A) statistical recovery of the LMM’s latent ranking, (B) optional Shapley semantics under ideal rationality, (C) empirical policy improvement. As written, readers can over-read Prop. 1 as solving the credit-assignment identification problem the paper poses in §1.
  3. [§1, §5.1, §7, Fig. 4, contribution (3)] The headline setting is openness in the active set (Open Dec-POMDP; battery removal/respawn in §7). Main results (Fig. 4) compare full MARS-RA to baselines on the open tasks, but there is no controlled ablation that freezes |I_t|=n versus dynamic |I_t| while holding the rest of the pipeline fixed, nor a report of how often |I_t|<2 causes skipped comparisons and how credits are defined for inactive agents beyond zero matrix rows. Without that, it is hard to credit the rank-aggregation formulation—versus LMM-derived dense shaping alone—for robustness to changing team size, which is listed as a primary motivation and contribution (3).
  4. [§5.1, §8, Appendix C, Fig. 5(f)] §8 and Appendix C: MARS-RA issues many LMM pairwise queries (ordered pairs among active agents, optionally repeated; action repeat 32 with stacked frames). Training time is reported (~30h vs ~16h for MAPPO) but there is no matched compute or query-budget comparison against SAMA/V-GEPF/LMM Score, and no scaling curve in |I_t| beyond the 2–4 agent bench and a single 10-agent Pistonball setting. Because pairwise cost is quadratic in active agents, the central practicality claim for open embodied teams needs either a fixed comparison budget protocol, active-pair sampling, or explicit cost–performance Pareto against single-pass LMM scoring.
minor comments (7)
  1. [Fig. 4 caption] Figure 4 caption says “three tasks in MARS-RA”; it should say MARS-Bench.
  2. [§5.3, Eq. (4)–(5)] Eq. (4): ψ = softmax(ĉ_t) is written as a scalar potential on (s_t,t) but softmax yields a vector over agents. Clarify whether each agent i uses ψ_i = softmax(ĉ_t)_i in its own shaped reward, and how joint vs per-agent PBRS is applied under MAPPO.
  3. [§4, §8.4] Assumptions §4 are thin (corpus diversity; multi-image prompts; image-evaluable tasks). A short failure-mode paragraph tying §8.4’s “all agents face a wall” errors to Assumption (3) would help.
  4. [§2] Related work on LLM credit assignment (LLM-MCA, LCA, SAMA, LERO) is cited; a sharper contrast table (observation modality, open teams, absolute vs pairwise, PBRS vs subgoals) would clarify novelty versus SAMA/V-GEPF.
  5. [§8.1, Table 3] Report the exact ρ used in main Fig. 4 curves (Table 3 sweeps ρ but main text is silent).
  6. [§8.4, Appendix F] Real-world protocol (n=25, 64%) is encouraging; please state sim-to-real gaps (latency, perception, battery model) and whether any fine-tuning occurred on hardware.
  7. [§1, throughout] Minor typos/style: “we proposeMARS-RA”, “constructMARS-Bench”, inconsistent spacing before citations; arXiv id/date in the banner look placeholder-like relative to the reference list.

Circularity Check

0 steps flagged

No significant circularity: credit scores are defined from LMM comparisons via standard BT-MLE and PBRS; success-rate claims are not forced by that definition.

full rationale

The load-bearing pipeline is (i) LMM pairwise judgments on egocentric images, (ii) Bradley–Terry MLE aggregation to ĉ_t (Eqs. 2–3), (iii) potential-based shaping ψ=softmax(ĉ_t) with the classical Ng/Devlin invariance guarantee (Eqs. 4–5), and (iv) empirical success rates versus MAPPO and other baselines. Prop. 1 only restates standard connected-graph MLE concentration of ˆc_t to the LMM’s own latent preference c*_t (citing Negahban/Shah); it does not define success in terms of ĉ_t. Prop. 2 is explicitly conditional and interpretive (“If … c*_t = v*_t …”), not an empirical recovery claim. Dense-reward agreement is used only as an offline proxy for LMM comparison accuracy (Fig. 6, Table 4), not as the training target or the reported primary metric. Core citations (Bradley–Terry, Negahban, Shah, Ng PBRS, Shapley) are external classics; author self-citations are peripheral. Nothing in the derivation equates the method’s inputs with its success-rate outputs by construction. Identification risk (whether c*_t tracks true marginal contribution) is a validity concern, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 2 invented entities

The central empirical claim rests on standard MARL/PBRS/BT machinery plus domain assumptions that LMMs can judge contribution from egocentric pixels and that dense-reward agreement proxies comparison quality. Free knobs include shaping weight ρ, query count, and choice of LMM. No new physical entities; MARS-Bench and the MARS-RA pipeline are engineered constructs, not postulated natural kinds.

free parameters (4)
  • ρ (potential shaping weight) = best around 0.1–1.0 (table of success rates)
    Scales F relative to environment reward; swept post hoc on MARS-Bench with best band {0.1, 0.5, 1.0}.
  • LMM query count K per decision = main experiments often 1; analysis uses 1–16
    Number of pairwise comparisons; performance rises with K especially for weaker models.
  • action repeat / frame stack = repeat 32; stack 4
    Decisions every 32 env steps; 4 grayscale frames → 128×128×4 observations—engineering choices affecting credit timing.
  • MAPPO and baseline hyperparameters = e.g. lr=0.001, γ=0.99, clip=0.2
    lr, batch, clip, γ, λ etc. fixed to shared table / prior papers; not derived.
axioms (6)
  • standard math Potential-based reward shaping preserves policy invariance / Nash structure under the stated dynamic potential (Ng et al.; Devlin & Kudenko).
    Invoked in §5.3 to justify inserting contribution scores as F = γψ' − ψ.
  • standard math Bradley–Terry MLE on a connected comparison graph concentrates to latent scores at rate ~sqrt((n log n)/K) under standard RSC/Hessian-Laplacian arguments.
    Prop. 1 / Appendix B cite Negahban, Shah et al.; used for convergence/robustness claim.
  • domain assumption LMMs trained on diverse corpora can process multi-image inputs and follow prompts to judge relative team contribution in the target tasks.
    Assumptions §4 (1)–(3); load-bearing for automated comparisons without human labels.
  • ad hoc to paper There exists a latent ground-truth preference vector c*_t that the LMM implements as a BT comparator (possibly aligned with Shapley values under ideal rationality).
    Stated for Prop. 1–2; not independently measured—scores recover c*, not proven true contribution.
  • domain assumption Open Dec-POMDP model with battery-driven enter/exit captures the embodied openness of interest.
    Preliminaries §3 and MARS-Bench design; success metrics defined under this process.
  • ad hoc to paper Agreement between LMM pairwise winners and dense per-agent environment rewards is a valid accuracy measure for contribution judgment.
    §8.4 / Appendix E; dense rewards are privileged internal signals, not external contribution ground truth.
invented entities (2)
  • MARS-RA pipeline (LMM pairwise matrix → BT credits → softmax potential shaping) no independent evidence
    purpose: Operational credit-assignment mechanism for open embodied MARL training.
    Engineered method, not a claimed natural law; evaluated empirically.
  • MARS-Bench (Pass Gate, Herd Sheep, Collect Ball with dynamic active agents) no independent evidence
    purpose: Provide open-team embodied cooperative tasks missing from prior fixed-team benchmarks.
    New evaluation suite in ManiSkill3/XLeRobot; community value if adopted.

pith-pipeline@v1.2.0-daily-grok45 · 25375 in / 3661 out tokens · 71889 ms · 2026-07-31T21:52:56.696564+00:00 · methodology

0 comments
read the original abstract

Credit assignment is a fundamental challenge in cooperative multi-agent reinforcement learning, particularly in embodied AI settings characterized by limited and delayed feedback as well as dynamically changing numbers of active agents. We propose MARS-RA, a framework that reformulates credit assignment as a rank aggregation problem using contribution-based pairwise comparisons among agents generated by large multimodal models. This shift from absolute to relative estimation ensures robustness against noise and dynamic agent participation, converting comparison results into contribution scores for potential-based reward shaping. We provide theoretical justification for the convergence and robustness of the proposed framework, and show that Shapley values can be used as an interpretive reference. Experimental results on challenging tasks of different types indicate that MARS-RA can guide agents toward effective cooperation.

Figures

Figures reproduced from arXiv: 2607.27967 by Chengming Zhou, Dawei Wang, Di Zhao, Gary Ushaw, Marci Chi Ma, Richard Davison, Xiaoyang Liu, Xinyuan Liu.

Figure 1
Figure 1. Figure 1: The MARS-RA framework for MARL credit assignment in cooperative embodied AI systems. Agents’ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Example of the prompt used in the Pass Gate [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a–c) Visualizations of the Pass Gate, Herd [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Learning curves of all compared methods on the three tasks in MARS-RA, trained for 50 million [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance comparison on five Overcooked tasks and Pistonball after training for 1 million environment [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Left: Pairwise comparison accuracy of the [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Left: A real-world 3D-scanned indoor envi [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Model architectures used by all MARL algo [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Screenshots of five tasks in the Overcooked environment: (a) Cramped Room, (b) Asymmetric Advantages, [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Screenshot of the Pistonball environment. [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Training Curves of MARS-RA with Different LMMs and Numbers of Queries for Pairwise Comparisons. [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Pairwise comparison errors by LMMs mainly occur when both agents’ egocentric observations are [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The learning curve obtained during training [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: A representative successful real-world trajectory of XLeRobots in the Pass Gate task, where a policy [PITH_FULL_IMAGE:figures/full_fig_p020_14.png] view at source ↗

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

173 extracted references · 46 linked inside Pith

  1. [1]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  2. [2]

    Publications Manual , year = "1983", publisher =

    1983

  3. [3]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    arXiv 1981

  4. [4]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  5. [5]

    Dan Gusfield , title =. 1997

    1997

  6. [6]

    2006 , publisher=

    Pattern recognition and machine learning , author=. 2006 , publisher=

    2006

  7. [7]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  8. [8]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  9. [9]

    FirstName LastName , title =

  10. [10]

    FirstName Alpher , title =

  11. [11]

    Journal of Foo , volume = 13, number = 1, pages =

    FirstName Alpher and FirstName Fotheringham-Smythe , title =. Journal of Foo , volume = 13, number = 1, pages =

  12. [12]

    Journal of Foo , volume = 14, number = 1, pages =

    FirstName Alpher and FirstName Fotheringham-Smythe and FirstName Gamow , title =. Journal of Foo , volume = 14, number = 1, pages =

  13. [13]

    FirstName Alpher and FirstName Gamow , title =

  14. [14]

    2023 International Conference on Inventive Computation Technologies (ICICT) , pages=

    Robot as staff: Robot for alibaba e-commerce warehouse process , author=. 2023 International Conference on Inventive Computation Technologies (ICICT) , pages=. 2023 , organization=

    2023

  15. [15]

    com: Operations research algorithms drive intelligent warehouse robots to work , author=

    JD. com: Operations research algorithms drive intelligent warehouse robots to work , author=. INFORMS Journal on Applied Analytics , volume=. 2022 , publisher=

    2022

  16. [16]

    2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=

    Detecting Wildfires on UAVs with Real-time Segmentation Trained by Larger Teacher Models , author=. 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=. 2025 , organization=

    2025

  17. [17]

    Science Robotics , volume=

    Swarm of micro flying robots in the wild , author=. Science Robotics , volume=. 2022 , publisher=

    2022

  18. [18]

    Mind , volume=

    Computing machinery and intelligence (1950) , author=. Mind , volume=

    1950

  19. [19]

    1998 , publisher=

    Being there: Putting brain, body, and world together again , author=. 1998 , publisher=

    1998

  20. [20]

    arXiv preprint arXiv:2509.20021 , year=

    Embodied ai: From llms to world models , author=. arXiv preprint arXiv:2509.20021 , year=

    arXiv

  21. [21]

    Proceedings of the IRE , volume=

    Steps toward artificial intelligence , author=. Proceedings of the IRE , volume=. 2007 , publisher=

    2007

  22. [22]

    arXiv preprint arXiv:1705.08926 , year=

    Counterfactual multi-agent policy gradients (2017) , author=. arXiv preprint arXiv:1705.08926 , year=

    Pith/arXiv arXiv 2017

  23. [23]

    arXiv preprint arXiv:1706.05296 , year=

    Value-decomposition networks for cooperative multi-agent learning , author=. arXiv preprint arXiv:1706.05296 , year=

    Pith/arXiv arXiv

  24. [24]

    Journal of Machine Learning Research , volume=

    Monotonic value function factorisation for deep multi-agent reinforcement learning , author=. Journal of Machine Learning Research , volume=

  25. [25]

    International Conference on Machine Learning , pages=

    Megaverse: Simulating embodied agents at one million experiences per second , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  26. [26]

    arXiv preprint arXiv:2006.07869 , year=

    Benchmarking multi-agent deep reinforcement learning algorithms in cooperative tasks , author=. arXiv preprint arXiv:2006.07869 , year=

    Pith/arXiv arXiv 2006

  27. [27]

    2018 , publisher=

    The book of why: the new science of cause and effect , author=. 2018 , publisher=

    2018

  28. [28]

    , author=

    Preschoolers are able to take merit into account when distributing goods. , author=. Developmental psychology , volume=. 2012 , publisher=

    2012

  29. [29]

    2008 , publisher=

    The Rational Imagination: How People Create Alternatives to Reality , author=. 2008 , publisher=

    2008

  30. [30]

    Nature Human Behaviour , pages=

    Computational basis of hierarchical and counterfactual information processing , author=. Nature Human Behaviour , pages=. 2025 , publisher=

    2025

  31. [31]

    Social cognitive and affective neuroscience , volume=

    Counterfactual thinking: an fMRI study on changing the past for a better future , author=. Social cognitive and affective neuroscience , volume=. 2013 , publisher=

    2013

  32. [32]

    arXiv preprint arXiv:2502.16863 , year=

    Leveraging Large Language Models for Effective and Explainable Multi-Agent Credit Assignment , author=. arXiv preprint arXiv:2502.16863 , year=

    Pith/arXiv arXiv

  33. [33]

    The American journal of psychology , volume=

    An experimental study of apparent behavior , author=. The American journal of psychology , volume=. 1944 , publisher=

    1944

  34. [34]

    Artificial Intelligence Review , volume=

    Deep multiagent reinforcement learning: Challenges and directions , author=. Artificial Intelligence Review , volume=. 2023 , publisher=

    2023

  35. [35]

    arXiv preprint arXiv:2505.05108 , year=

    Multi-agent embodied ai: Advances and future directions , author=. arXiv preprint arXiv:2505.05108 , year=

    Pith/arXiv arXiv

  36. [36]

    arXiv preprint arXiv:2312.11805 , year=

    Gemini: a family of highly capable multimodal models , author=. arXiv preprint arXiv:2312.11805 , year=

    Pith/arXiv arXiv

  37. [37]

    International conference on machine learning , pages=

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  38. [38]

    Advances in neural information processing systems , volume=

    Visual instruction tuning , author=. Advances in neural information processing systems , volume=

  39. [39]

    arXiv preprint arXiv:2408.08105 , year=

    Multimodal causal reasoning benchmark: Challenging vision large language models to discern causal links across modalities , author=. arXiv preprint arXiv:2408.08105 , year=

    Pith/arXiv arXiv

  40. [40]

    arXiv preprint arXiv:2509.17784 , year=

    Revealing Multimodal Causality with Large Language Models , author=. arXiv preprint arXiv:2509.17784 , year=

    arXiv

  41. [41]

    arXiv preprint arXiv:2505.03792 , year=

    Towards Efficient Online Tuning of VLM Agents via Counterfactual Soft Reinforcement Learning , author=. arXiv preprint arXiv:2505.03792 , year=

    Pith/arXiv arXiv

  42. [42]

    Autonomous Robots , volume=

    Learning scalable and efficient communication policies for multi-robot collision avoidance , author=. Autonomous Robots , volume=. 2023 , publisher=

    2023

  43. [43]

    arXiv preprint arXiv:2309.17433 , year=

    Dream: Decentralized reinforcement learning for exploration and efficient energy management in multi-robot systems , author=. arXiv preprint arXiv:2309.17433 , year=

    Pith/arXiv arXiv

  44. [44]

    2022 International Conference on Computer Communications and Networks (ICCCN) , pages=

    Energy-aware multi-agent reinforcement learning for collaborative execution in mission-oriented drone networks , author=. 2022 International Conference on Computer Communications and Networks (ICCCN) , pages=. 2022 , organization=

    2022

  45. [45]

    IEEE Robotics and Automation Letters , year=

    Safe Multi-Agent Reinforcement Learning for Behavior-Based Cooperative Navigation , author=. IEEE Robotics and Automation Letters , year=

  46. [46]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  47. [47]

    arXiv preprint arXiv:2502.14496 , year=

    Enhancing Language Multi-Agent Learning with Multi-Agent Credit Re-Assignment for Interactive Environment Generalization , author=. arXiv preprint arXiv:2502.14496 , year=

    Pith/arXiv arXiv

  48. [48]

    International Conference on Artificial Intelligence and Statistics , pages=

    Multi-agent credit assignment with pretrained language models , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2025 , organization=

    2025

  49. [49]

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , pages=

    2019

  50. [50]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    Pith/arXiv arXiv

  51. [51]

    arXiv preprint arXiv:2407.07895 , year=

    Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models , author=. arXiv preprint arXiv:2407.07895 , year=

    Pith/arXiv arXiv

  52. [52]

    M4Bench: A Benchmark of Multi-domain Multi-granularity Multi-image Understanding for Multi-modal Large Language Models , author=

  53. [53]

    arXiv preprint arXiv:2406.09411 , year=

    Muirbench: A comprehensive benchmark for robust multi-image understanding , author=. arXiv preprint arXiv:2406.09411 , year=

    Pith/arXiv arXiv

  54. [54]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Identifying and mitigating position bias of multi-image vision-language models , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  55. [55]

    arXiv preprint arXiv:2508.00726 , year=

    Mihbench: Benchmarking and mitigating multi-image hallucinations in multimodal large language models , author=. arXiv preprint arXiv:2508.00726 , year=

    Pith/arXiv arXiv

  56. [56]

    arXiv preprint arXiv:2506.22434 , year=

    MiCo: Multi-image Contrast for Reinforcement Visual Reasoning , author=. arXiv preprint arXiv:2506.22434 , year=

    Pith/arXiv arXiv

  57. [57]

    IEEE/ASME Transactions on Mechatronics , year=

    Aligning cyber space with physical world: A comprehensive survey on embodied ai , author=. IEEE/ASME Transactions on Mechatronics , year=

  58. [58]

    Science , volume=

    New approaches to robotics , author=. Science , volume=. 1991 , publisher=

    1991

  59. [59]

    arXiv preprint arXiv:2310.00835 , year=

    Tram: Benchmarking temporal reasoning for large language models , author=. arXiv preprint arXiv:2310.00835 , year=

    Pith/arXiv arXiv

  60. [60]

    arXiv preprint arXiv:2311.17667 , year=

    Timebench: A comprehensive evaluation of temporal reasoning abilities in large language models , author=. arXiv preprint arXiv:2311.17667 , year=

    Pith/arXiv arXiv

  61. [61]

    Proceedings of the 28th ACM International Conference on Multimedia , pages=

    Bridging the gap between vision and language domains for improved image captioning , author=. Proceedings of the 28th ACM International Conference on Multimedia , pages=

  62. [62]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Vilem: Visual-language error modeling for image-text retrieval , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  63. [63]

    Advances in Neural Information Processing Systems , volume=

    Language is not all you need: Aligning perception with language models , author=. Advances in Neural Information Processing Systems , volume=

  64. [64]

    IEEE Transactions on Multimedia , year=

    Lmeye: An interactive perception network for large language models , author=. IEEE Transactions on Multimedia , year=

  65. [65]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Omnia de egotempo: Benchmarking temporal understanding of multi-modal llms in egocentric videos , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  66. [66]

    arXiv preprint arXiv:2412.11391 , year=

    Temporal Contrastive Learning for Video Temporal Reasoning in Large Vision-Language Models , author=. arXiv preprint arXiv:2412.11391 , year=

    Pith/arXiv arXiv

  67. [67]

    arXiv preprint arXiv:2310.12921 , year=

    Vision-language models are zero-shot reward models for reinforcement learning , author=. arXiv preprint arXiv:2310.12921 , year=

    Pith/arXiv arXiv

  68. [68]

    arXiv preprint arXiv:2402.03681 , year=

    Rl-vlm-f: Reinforcement learning from vision language foundation model feedback , author=. arXiv preprint arXiv:2402.03681 , year=

    Pith/arXiv arXiv

  69. [69]

    Industrial Robot: the international journal of robotics research and application , volume=

    The role of robots in logistics , author=. Industrial Robot: the international journal of robotics research and application , volume=. 2024 , publisher=

    2024

  70. [70]

    Icml , volume=

    Policy invariance under reward transformations: Theory and application to reward shaping , author=. Icml , volume=. 1999 , organization=

    1999

  71. [71]

    Electronics , volume=

    Reinforcement learning-based multimodal model for the stock investment portfolio management task , author=. Electronics , volume=. 2024 , publisher=

    2024

  72. [72]

    2024 IEEE International Conference on Robotics and Biomimetics (ROBIO) , pages=

    Swarmpath: Drone swarm navigation through cluttered environments leveraging artificial potential field and impedance control , author=. 2024 IEEE International Conference on Robotics and Biomimetics (ROBIO) , pages=. 2024 , organization=

    2024

  73. [73]

    Science Robotics , volume=

    Cerberus in the darpa subterranean challenge , author=. Science Robotics , volume=. 2022 , publisher=

    2022

  74. [74]

    arXiv preprint arXiv:2310.02071 , year=

    Towards end-to-end embodied decision making via multi-modal large language model: Explorations with gpt4-vision and beyond , author=. arXiv preprint arXiv:2310.02071 , year=

    Pith/arXiv arXiv

  75. [75]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    From multimodal llms to generalist embodied agents: Methods and lessons , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  76. [76]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Proagent: building proactive cooperative agents with large language models , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  77. [77]

    arXiv preprint arXiv:2509.14380 , year=

    CRAFT: Coaching Reinforcement Learning Autonomously using Foundation Models for Multi-Robot Coordination Tasks , author=. arXiv preprint arXiv:2509.14380 , year=

    Pith/arXiv arXiv

  78. [78]

    arXiv preprint arXiv:2503.02077 , year=

    M3hf: Multi-agent reinforcement learning from multi-phase human feedback of mixed quality , author=. arXiv preprint arXiv:2503.02077 , year=

    Pith/arXiv arXiv

  79. [79]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Contrastive identity-aware learning for multi-agent value decomposition , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  80. [80]

    Neurocomputing , pages=

    XMIX: Graph-based temporal credit assignment and attention-augmented value decomposition for multi-agent cooperative reinforcement learning , author=. Neurocomputing , pages=. 2025 , publisher=

    2025

Showing first 80 references.