Pith. sign in

REVIEW 4 major objections 5 minor 297 references

LEMUR claims an agent can learn multiple conflicting reward objectives solely from preference feedback, then optimize policies that track the ground-truth oracle on both objectives.

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 · deepseek-v4-flash

2026-08-03 04:31 UTC pith:NFVABGHW

load-bearing objection LEMUR is a plausible but oversold extension of PbRL to MORL: the per-teacher reward decomposition looks sound, but the 'no pre-defined reward function' claim leans on ground-truth teacher anchors that the method quietly requires. the 4 major comments →

arxiv 2607.29559 v1 pith:NFVABGHW submitted 2026-07-31 cs.AI cs.RO

LEMUR: Learning to Align with Multi-Objective Reinforcement Learning from Preference Feedback

Manith Adikari , Bei Peng , Samuele Vinanzi , Angelo Cangelosi This is my paper
classification cs.AI cs.RO
keywords multi-objective reinforcement learningpreference-based RLreward learning from human feedbackPareto frontBradley-Terry modelvector reward relabelingmultiple teacherscontinuous control
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.

LEMUR aims to remove reward engineering from multi-objective reinforcement learning. It treats each conflicting teacher's preferences as evidence for a separate latent objective, learns one reward model per teacher, and optimizes a set of policies against the resulting vector reward using linear scalarization. On four benchmark control and manipulation tasks, LEMUR is the method that most closely tracks the ground-truth oracle on both objectives simultaneously, and it remains competitive under label noise, small feedback budgets, and up to four objectives. The central claim is that explicitly preserving the multi-objective structure of feedback outperforms collapsing it into a single scalar reward, which is why the method matters for real-world tasks with competing goals and no accessible reward function.

Core claim

On the paper's own terms, the discovery is that conflicting preference feedback can be decomposed into separate, objective-specific reward models that share a common vector-reward structure, and that a multi-objective policy population optimized against these learned rewards can recover near-oracle compromise policies. Each teacher's reward model predicts the full objective vector and projects it onto that teacher's preference anchor, coupling the models to one objective space; rewards are recomputed at training time so the policy stays synchronized with the improving reward models. The paper observes that baselines which pool conflicting feedback into one scalar, or fix a welfare scalarizat

What carries the argument

The load-bearing component is a weight-conditioned reward model per teacher: a small MLP that predicts the shared objective vector, with the teacher's utility obtained by a dot product with that teacher's preference anchor. This per-teacher decomposition is coupled to a population-based multi-objective policy optimizer over linear scalarizations of the learned vector reward. Two supporting mechanisms keep it stable: a shared replay buffer across policies, and vector-reward relabeling that recomputes rewards with current models so off-policy training does not drift on stale signals. Unsupervised pre-training populates the buffer with diverse experiences for informative queries.

Load-bearing premise

The evidence assumes that scripted teachers whose preferences are generated from linear scalarizations of a hidden ground-truth vector reward behave like real humans with genuinely multi-objective values.

What would settle it

Have human annotators supply genuine preferences on these tasks, or generate preference labels from a non-linear, context-dependent rule, then measure whether LEMUR's learned policies still track the ground-truth trade-off front and whether reward alignment remains high. If the front collapses or alignment drops sharply, the central claim fails.

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

If this is right

  • If LEMUR is right, multi-objective control tasks can be solved without any ground-truth reward function, removing a major bottleneck for physical systems where rewards are hard to specify.
  • Heterogeneous feedback sources—for instance a non-expert judging task success and an expert judging execution quality—can be kept as separate objectives rather than diluted into one scalar.
  • Adding a new objective or teacher should require only a new reward model and an extended vector reward, without restarting training.
  • The learned policies produce a broad, evenly spaced set of trade-offs, not a single fixed compromise, as measured by higher hypervolume and lower sparsity.
  • With fewer than a thousand queries per teacher on some environments, the method stays sample-efficient enough for settings where human feedback is expensive.

Where Pith is reading between the lines

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

  • I infer the strongest payoff would be in human-subject settings: if real annotators give preferences that do not come from any fixed linear scalarization, the method would need adaptation beyond the scripted-teacher evidence presented.
  • The per-teacher decomposition suggests a natural extension to non-linear scalarization or to preference distributions over a whole community, though the paper does not provide evidence for those regimes.
  • A testable prediction follows from the mechanism: as long as each teacher's utility is a monotone function of a shared vector reward, even noisy labels should leave Pareto-front quality largely intact, which the noise ablation partially supports.
  • One consequence the authors leave implicit is that the same machinery could serve as a reward-learning layer for any vector-reward RL algorithm, not only the specific policy optimizer used here.

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 / 5 minor

Summary. The paper proposes LEMUR, a preference-based multi-objective RL method that learns multiple objective-specific reward models from the conflicting preference feedback of multiple teachers and then trains a population of policies (MORL/D) against the learned vector reward, using a shared replay buffer, reward relabeling, and unsupervised pre-training. Experiments on MO-LunarLander, MO-Hopper, MO-Cheetah, and MO-MetaWorld claim that LEMUR most closely tracks an oracle on both objectives, outperforms aggregation baselines and prior preference-based MORL methods, scales to three and four objectives, and is robust to label noise and reduced feedback budgets. The paper also reports reward-model alignment metrics. The main concerns are inconsistencies in the description of the reward-model architecture, the role of ground-truth teacher anchors, and a reward-alignment comparison that is not uniformly favorable to LEMUR.

Significance. If the central claims hold, LEMUR addresses a genuine gap: extending preference-based RL to multi-teacher, multi-objective settings without hand-specified reward functions. The paper is strong in scope and effort: it benchmarks four environments, discloses baseline adaptations carefully, includes extensive ablations, and is honest about the scripted-teacher limitation. However, the significance as stated depends on resolving whether the method truly operates without pre-defined reward information. The use of ground-truth teacher anchors in the reward model and query distribution is currently a load-bearing ambiguity, and the reward-alignment claim is partially contradicted by the appendix. The paper is therefore promising but not yet ready in its present form.

major comments (4)
  1. [§4.2, Appendix D, Algorithm 1] The reward-model architecture is described inconsistently. §4.2 defines each teacher's model as an MLP predicting the full objective vector, with scalar utility a_j^T \hat r_{\psi_j}(s,a), i.e., input (s,a). Appendix D instead says the model is a weight-conditioned MLP taking (s,a) concatenated with a preference weight vector w. Algorithm 1's Require list omits anchors and line 15 uses r_w = w^T \hat r_\psi(s,a), with no indication of per-teacher models. This makes the method underspecified and not reproducible from the paper. Please state the exact input/output of the reward model, whether anchors are required, and unify notation between the algorithm and the text.
  2. [§4.2, Appendix D, Table 12] The paper claims to learn without pre-defined reward functions, but Table 12 lists the teacher anchors as ground-truth weight vectors w_j, and Appendix D says query weights are sampled from a Dirichlet centered on that anchor. If a_j is given to the reward model as a projection direction and used to shape the query distribution, then the agent is supplied with each teacher's scalarization vector — precisely the kind of reward-side information the contribution claims to infer. If a_j is not given, the architecture and query procedure are underspecified. Either way, the claim 'without pre-defined reward functions' is not supported as stated. Please clarify whether anchors are known a priori, learned, or assumed part of the teacher specification, and adjust the contribution framing accordingly.
  3. [§5.1, Appendix F.2, Table 11] The main text states that LEMUR's reward models 'outperform both PbMORL and FPbRL (Appendix F.2).' This is contradicted on Hopper: in Table 11, FPbRL achieves TAC 0.933±0.000 and Kendall tau 0.933±0.000, while LEMUR achieves 0.856±0.014 and 0.914±0.002. The summary in Table 2 reports only LEMUR's numbers, masking this. Please qualify the claim, or explain why the trajectory-level metrics favor FPbRL despite LEMUR's higher per-state correlation.
  4. [§4.2, Appendix D, Appendix G] There is a tension between the fixed teacher utility defined in Appendix G (r_j(s,a)=w_j^T r(s,a)) and the Dirichlet-sampled query weights in Appendix D. If teacher labels are always generated with the fixed anchor w_j, then conditioning the reward model on varying w has no consistent supervision for the w dimension. If labels are generated with the sampled w, then teachers are not fixed and Table 12 does not describe the actual utilities. This ambiguity affects the interpretation of all reward-learning results and needs to be resolved explicitly.
minor comments (5)
  1. [§5, Appendix C.3–C.4] The main text says all experiments average five seeds, but the query-length ablation (length 35 and length 1) and the entropy-sampling ablation are reported as single seeds. This should be stated in the main text or the arms should be run with multiple seeds.
  2. [Appendix E] MORAL is trained using expert demonstrations collected from ground-truth scalarized rewards. Although this is disclosed, it means MORAL is given reward information that LEMUR and other baselines do not receive. The main-text comparison would be clearer if this asymmetry were restated where the baselines are introduced.
  3. [Appendix F.1, Table 8] PbMORL reports zero variance on Hopper and HalfCheetah hypervolume. Since the metric is computed over evaluation rollouts, zero variance across seeds is surprising; please explain or verify.
  4. [General] No code or supplementary material is released. For a method that is currently underspecified in places, public code would substantially aid reproducibility.
  5. [General] The naming of environments alternates between 'MO-Cheetah' and 'MO-HalfCheetah' across the text and tables; please standardize.

Circularity Check

0 steps flagged

No significant circularity: reward models are trained on preference labels by standard supervised fitting; the anchor-vector input is a modeling/validity concern, not a circular derivation.

full rationale

The paper's derivation chain is not circular. LEMUR trains each teacher-specific reward model with the Bradley-Terry cross-entropy loss (Eq. 1) on labels generated from ground-truth scalarized utilities; the learned vector reward is then used in Eq. 2 for multi-objective policy optimization. The reported returns are evaluated against the environment's true objective returns, which the agent never observes, so the empirical policy results are not forced by the reward-model fitting. The clearest potential concern is the teacher 'preference anchor' a_j used in the reward model (Section 4.2) and as the center of the query-weight Dirichlet (Appendix D); Table 12 identifies these anchors with the scripted ground-truth weight vectors w_j. This means the method as described requires known scalarization weights as an input and weakens the 'without pre-defined reward functions' claim; it also creates an inconsistency with Algorithm 1, whose Require list omits anchors. However, this is a validity/correctness issue rather than a circularity: the reward model r̂_ψj is still learned from preference data, and the alignment scores in Table 2 are standard train-fit evaluations, not independent predictions that reduce to their inputs by construction. No self-citations are load-bearing, no uniqueness theorem is imported, and no known result is merely renamed. The paper itself flags the scripted-teacher limitation in the Conclusion. On the circularity scale, this is a clean non-finding.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

No new physical entities, forces, or dimensions are introduced. The 'latent reward vector' is a modeling construct naming per-teacher reward functions; it is not an entity with independent falsifiable handles outside this paper. The main hidden costs are the chosen teacher anchors and several feedback-loop hyperparameters that are set by hand and demonstrably affect outcomes.

free parameters (5)
  • Teacher anchor weights = e.g., [0.6,0.4] vs [0.4,0.6] per environment (Table 12)
    These define the objectives and conflict level, are used to generate scripted labels, and are embedded in the reward-model conditioning. They are chosen by the experimenters, not learned.
  • Dirichlet concentration kappa = 30.0
    Controls how broadly each teacher's reward model sees preference weights around its anchor; no tuning sweep is reported (Appendix D).
  • Query segment length H = 50
    The ablation in Appendix C.3 shows segment length matters; longer segments improve final performance, so the default choice affects results.
  • Query budget M per teacher = 200 (LunarLander), 500 (Hopper), 300 (Cheetah/MetaWorld)
    Feedback budget strongly affects convergence; the ablation in Figure 5c shows performance improves with budget.
  • MORL/D population size and PSA neighborhood = 6 and 2
    Fixed across all environments; these hyperparameters determine the size and diversity of the recovered Pareto front and are not swept.
axioms (6)
  • domain assumption The MOMDP tuple with vector reward r(s,a) in R^m is the correct formalization of multi-objective decision-making.
    Section 2; standard model (White 1982), but it restricts the kinds of objectives and preferences that can be represented.
  • domain assumption Human preferences over trajectory segments follow the Bradley-Terry model.
    Equation 1; all reward learning is built on this. If real humans produce non-BT, intransitive, or context-dependent preferences, the learned rewards are misspecified.
  • domain assumption Scripted teachers that generate labels from ground-truth linear scalarizations w_j^T r faithfully represent human multi-objective preferences.
    Section 5 and Appendix G; this is the load-bearing evaluation premise and is acknowledged as a limitation in the Conclusion.
  • domain assumption Linear scalarization f_w(r)=w^T r is sufficient to recover desirable trade-offs (the Convex Coverage Set).
    Section 3; the authors note in the Conclusion that non-linear scalarization would be needed for non-convex regions of the frontier.
  • domain assumption MORL/D (MO-SAC) with Pareto simulated annealing is a reliable multi-objective optimizer for the learned rewards.
    Section 4.3 and Appendix D; the entire policy optimization stage depends on this external algorithm.
  • ad hoc to paper The weight-conditioned reward model, where each teacher's scalar utility is a projection of a predicted objective vector onto that teacher's anchor, can represent each teacher's utility.
    Sections 4.2 and D; the architecture is justified by prior work and an ablation, not derived from first principles.

pith-pipeline@v1.3.0-daily-deepseek · 26476 in / 13435 out tokens · 135941 ms · 2026-08-03T04:31:31.464099+00:00 · methodology

0 comments
read the original abstract

Reinforcement Learning (RL) systems are typically trained using a single, well-specified scalar reward function. However, real-world decision-making tasks often involve multiple, competing objectives, such as performance versus efficiency, where ground-truth reward functions are difficult to specify or inaccessible. While Multi-Objective RL (MORL) addresses such trade-offs by modeling rewards as vectors, existing approaches typically assume access to a well-specified reward function for each objective, inheriting the same challenges faced by single-objective RL. Meanwhile, Preference-based RL (PbRL) has shown great potential in solving complex tasks without access to a pre-defined reward function through reward learning from human feedback, yet has largely been studied in single-objective settings. In this work, we bridge this gap with LEMUR: Learning to Align with Multi-Objective Reinforcement Learning with Preference feedback, a novel framework where an agent interactively learns from the preferences of multiple humans to learn optimal multi-objective policies. Our approach jointly learns policies and multiple objective-specific reward models from human feedback, enabling agents to effectively balance competing objectives during learning. We evaluate LEMUR on a variety of benchmark multi-objective tasks, and empirical results demonstrate its superior performance over baseline methods. Our method presents a promising direction for solving multi-objective decision-making tasks without pre-defined reward functions.

Figures

Figures reproduced from arXiv: 2607.29559 by Angelo Cangelosi, Bei Peng, Manith Adikari, Samuele Vinanzi.

Figure 1
Figure 1. Figure 1: Illustration of our framework LEMUR: (1) Unsupervised Pre-training for the MORL agent to explore and collect [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Learning curves on all benchmark environments: [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Task Success Rate (%) learning curves on Meta [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: LEMUR scalability to higher-dimensional objec [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ablation studies on MO-Cheetah evaluating the impact of (a) the shared buffer, vector reward relabeling, and unsupervised pre-training, (b) noisy teacher labels, and (c) varying the total query budget per teacher on agent returns for both objectives; by default LEMUR uses 3900 queries (green). The results are averaged over multiple runs across five seeds. 2023; Ye et al. 2024). Online, iterative RLHF mitig… view at source ↗
Figure 6
Figure 6. Figure 6: Adding a third teacher mid-training (MO-Cheetah). A third teacher is introduced at 4 × 105 steps (dotted line) into an already-training two-teacher run. The new objective (bottom) is learned from scratch while the two existing objectives (top, middle) are preserved, so no retraining from scratch is required. Mean ± std over five seeds. at 4 × 105 environment steps, with training continuing un￾interrupted f… view at source ↗
Figure 8
Figure 8. Figure 8: Effect of query segment length (MO-Cheetah). Ground-truth return per objective for preference queries of length 50 (default), 35 and 1 transition, at a fixed budget of 300 queries per teacher. Shading is ±1 std; the length-35 and length-1 arms are single seeds and carry a nominal band. Shorter segments learn faster early but plateau by ∼2 × 105 steps, while length 50 overtakes them and continues improving.… view at source ↗
Figure 9
Figure 9. Figure 9: Uniform vs. entropy-based query sampling (MO-Cheetah). Ground-truth return per objective for the default uniform sampler and an entropy-based sampler that scores a 10× candidate pool by reward-model uncertainty. Shading is ±1 std; the entropy arm is a single seed and car￾ries a nominal band. Entropy-based selection is consistently ahead, but by a small margin relative to the effect of query length ( [PITH… view at source ↗
Figure 10
Figure 10. Figure 10: Overlapping preferences (MO-Cheetah). With near-aligned teacher anchors, LEMUR converges close to the ground-truth oracle on both objectives, showing that the method degrades gracefully when teachers largely agree. Mean ± std over five seeds. under the same overlap condition [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Varying levels of teacher conflict (MO-Cheetah). Each column is one conflict level, set by the teacher preference anchors: medium (w = [0.6, 0.4]/[0.4, 0.6]), hard ([0.75, 0.25]/[0.25, 0.75]) and harder ([0.9, 0.1]/[0.1, 0.9]). Rows give each objective’s ground-truth return. Solid lines are LEMUR, dashed the ground-truth-reward Oracle under the identical configuration; shading is ±1 std across seeds, with… view at source ↗
Figure 12
Figure 12. Figure 12: Reward-model ablation (MO-Cheetah). The weight-conditioned reward model substantially outperforms the reward-ensemble variant on both objectives under an identical optimizer, explorer, and query budget. Mean ± std over five seeds. the optimizer rather than reproducing the paper verbatim, so that reported gaps reflect differences in reward learning and preference aggregation, the object of study, rather th… view at source ↗
Figure 13
Figure 13. Figure 13: MORAL’s learned scalarisation weights in (MO-Cheetah). The two components of MORAL’s active￾query scalarisation posterior over training. 400 300 200 100 Episode Returns (Objective One) MORAL (SAC) MORAL (PPO) 0 1 2 3 4 5 6 7 Training Steps (×10 ) 500 400 300 200 100 Episode Returns (Objective Two) [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: confirms this: holding MORAL’s adversarial AIRL reward learning and active-query scalarisation poste￾rior fixed and varying only the policy optimizer, the SAC variant dominates PPO on both objectives throughout train￾ing, and the gap widens as training proceeds. The PPO variant additionally displays a pronounced sawtooth characteristic of on-policy updates (shown here under heavy smoothing). Re￾porting MO… 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

297 extracted references · 41 canonical work pages · 13 internal anchors

  1. [1]

    Structure and Interpretation of Computer Programs

    Harold Abelson and Gerald Jay Sussman and Julie Sussman. Structure and Interpretation of Computer Programs. 1985

    1985

  2. [2]

    Visual Information Extraction with Lixto

    Robert Baumgartner and Georg Gottlob and Sergio Flesca. Visual Information Extraction with Lixto. Proceedings of the 27th International Conference on Very Large Databases. 2001

    2001

  3. [3]

    Brachman and James G

    Ronald J. Brachman and James G. Schmolze. An overview of the KL-ONE knowledge representation system. Cognitive Science. 1985

    1985

  4. [4]

    Complexity results for nonmonotonic logics

    Georg Gottlob. Complexity results for nonmonotonic logics. Journal of Logic and Computation. 1992

    1992

  5. [5]

    International conference on machine learning , pages=

    Prediction-guided multi-objective reinforcement learning for continuous robot control , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  6. [6]

    Hypertree Decompositions and Tractable Queries

    Georg Gottlob and Nicola Leone and Francesco Scarcello. Hypertree Decompositions and Tractable Queries. Journal of Computer and System Sciences. 2002

    2002

  7. [7]

    Levesque

    Hector J. Levesque. Foundations of a functional approach to knowledge representation. Artificial Intelligence. 1984

    1984

  8. [8]

    Levesque

    Hector J. Levesque. A logic of implicit and explicit belief. Proceedings of the Fourth National Conference on Artificial Intelligence. 1984

    1984

  9. [9]

    On the compilability and expressive power of propositional planning formalisms

    Bernhard Nebel. On the compilability and expressive power of propositional planning formalisms. Journal of Artificial Intelligence Research. 2000

    2000

  10. [10]

    Leslie Lamport , title =

  11. [11]

    The Knowledge Engineering Review , volume =

    Intelligent Agents: Theory and Practice , author =. The Knowledge Engineering Review , volume =

  12. [12]

    Artificial Intelligence , volume =

    Collaborative Plans for Complex Group Action , author =. Artificial Intelligence , volume =

  13. [13]

    Logics of programs: axiomatics and descriptive power

    David Harel. Logics of programs: axiomatics and descriptive power. 1978

    1978

  14. [14]

    Clarkson

    Kenneth L. Clarkson. Algorithms for Closest-Point Problems (Computational Geometry). 1985

    1985

  15. [15]

    A More Perfect Union

    Barack Obama. A More Perfect Union. 2008

    2008

  16. [16]

    The fountain of youth

    Joseph Scientist. The fountain of youth. 2009

    2009

  17. [17]

    Sam Anzaroot and Andrew McCallum , title =

  18. [18]

    Proceedings of the 20th International Colloquium on Automata, Languages and Programming , series =

    Maintaining Discrete Probability Distributions Optimally , author =. Proceedings of the 20th International Colloquium on Automata, Languages and Programming , series =. 1993 , publisher =

    1993

  19. [19]

    Donald E. Knuth. The Art of Computer Programming, Vol. 1: Fundamental Algorithms. 1997

    1997

  20. [20]

    Anisi , title =

    David A. Anisi , title =

  21. [21]

    Journal of Artificial Intelligence Research , author =

    Multi-. Journal of Artificial Intelligence Research , author =. 2024 , pages =. doi:10.1613/jair.1.15702 , abstract =

    work page doi:10.1613/jair.1.15702 2024

  22. [22]

    Kim, Changyeon and Park, Jongjin and Shin, Jinwoo and Lee, Honglak and Abbeel, Pieter and Lee, Kimin , year =

  23. [23]

    Reinforcement Learning , author =

    Learning. Reinforcement Learning , author =

  24. [24]

    Varys, Kryspin , year =

  25. [25]

    Hussonnois, Maxence and Karimpanal, Thommen George and Rana, Santu , month = jan, year =. Human-. doi:10.48550/arXiv.2501.17431 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2501.17431

  26. [26]

    Divide and Conquer: Provably Unveiling the Pareto Front with Multi-Objective Reinforcement Learning

    Röpke, Willem and Reymond, Mathieu and Mannion, Patrick and Roijers, Diederik M. and Nowé, Ann and Rădulescu, Roxana , month = feb, year =. Divide and. doi:10.48550/arXiv.2402.07182 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2402.07182

  27. [27]

    Fairness in

    Siddique, Umer and Sinha, Abhinav and Cao, Yongcan , month = sep, year =. Fairness in. doi:10.48550/arXiv.2306.09995 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2306.09995

  28. [28]

    Personalized

    Jang, Joel and Kim, Seungone and Lin, Bill Yuchen and Wang, Yizhong and Hessel, Jack and Zettlemoyer, Luke and Hajishirzi, Hannaneh and Choi, Yejin and Ammanabrolu, Prithviraj , month = oct, year =. Personalized. doi:10.48550/arXiv.2310.11564 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.11564

  29. [29]

    Advances in Neural Information Processing Systems , volume=

    Rewarded soups: towards pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards , author=. Advances in Neural Information Processing Systems , volume=

  30. [30]

    Promptable Behaviors: Personalizing Multi-Objective Rewards from Human Preferences

    Hwang, Minyoung and Weihs, Luca and Park, Chanwoo and Lee, Kimin and Kembhavi, Aniruddha and Ehsani, Kiana , month = dec, year =. Promptable. doi:10.48550/arXiv.2312.09337 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2312.09337

  31. [31]

    Proceedings of the 41st International Conference on Machine Learning , pages =

    Position: A Roadmap to Pluralistic Alignment , author =. Proceedings of the 41st International Conference on Machine Learning , pages =. 2024 , editor =

    2024

  32. [32]

    and Chadwick, Martin J

    Bakker, Michiel A. and Chadwick, Martin J. and Sheahan, Hannah R. and Tessler, Michael Henry and Campbell-Gillingham, Lucy and Balaguer, Jan and McAleese, Nat and Glaese, Amelia and Aslanides, John and Botvinick, Matthew M. and Summerfield, Christopher , month = nov, year =. Fine-tuning language models to find agreement among humans with diverse preferenc...

    Pith/arXiv arXiv doi:10.48550/arxiv.2211.15006

  33. [33]

    Adaptive

    Harland, Hadassah and Dazeley, Richard and Vamplew, Peter and Senaratne, Hashini and Nakisa, Bahareh and Cruz, Francisco , month = oct, year =. Adaptive. doi:10.48550/arXiv.2410.23630 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.23630

  34. [34]

    Gupta, Raghav and Sullivan, Ryan and Li, Yunxuan and Phatale, Samrat and Rastogi, Abhinav , file =. Robust

  35. [35]

    2024 , pages =

    Science , author =. 2024 , pages =. doi:10.1126/science.adq2852 , abstract =

    doi:10.1126/science.adq2852 2024

  36. [36]

    Weisswange, Thomas H and Javed, Hifza and Dietrich, Manuel and Pham, Tuan Vu and Parreira, Maria Teresa and Sack, Michael and Jamali, Nawid , year =. What

  37. [37]

    and Yang, Diyi and Vosoughi, Soroush , month = oct, year =

    Liu, Ruibo and Yang, Ruixin and Jia, Chenyan and Zhang, Ge and Zhou, Denny and Dai, Andrew M. and Yang, Diyi and Vosoughi, Soroush , month = oct, year =. Training

  38. [38]

    Chen, Huaben and Ji, Wenkang and Xu, Lufeng and Zhao, Shiyu , month = oct, year =. Multi-

  39. [39]

    and Hassenzahl, Marc , month = jul, year =

    Pham, Tuan Vu and Weisswange, Thomas H. and Hassenzahl, Marc , month = jul, year =. Embodied. Designing. doi:10.1145/3643834.3660696 , abstract =

    arXiv

  40. [40]

    IEEE Access , author =

    Multi-. IEEE Access , author =. 2018 , pages =. doi:10.1109/ACCESS.2018.2831228 , abstract =

    arXiv 2018

  41. [41]

    Advice to

    Belpaeme, Tony , editor =. Advice to. Human-. 2020 , doi =

    2020

  42. [42]

    Applied Intelligence , author =

    A review of cooperative multi-agent deep reinforcement learning , volume =. Applied Intelligence , author =. 2023 , pages =. doi:10.1007/s10489-022-04105-y , abstract =

    doi:10.1007/s10489-022-04105-y 2023

  43. [43]

    Frontiers in Computer Science , author =

    A multimodal approach for modeling engagement in conversation , volume =. Frontiers in Computer Science , author =. 2023 , pages =. doi:10.3389/fcomp.2023.1062342 , abstract =

    arXiv 2023

  44. [44]

    International Journal of Social Robotics , author =

    The. International Journal of Social Robotics , author =. 2011 , pages =. doi:10.1007/s12369-010-0082-7 , abstract =

    doi:10.1007/s12369-010-0082-7 2011

  45. [46]

    Analysis of

    Shiota, Tsukasa and Yamamura, Takashi and Shimada, Kazutaka , editor =. Analysis of. Collaboration. 2018 , doi =

    2018

  46. [47]

    Multimodal Technologies and Interaction , author =

    A. Multimodal Technologies and Interaction , author =. 2017 , pages =. doi:10.3390/mti1030014 , abstract =

    doi:10.3390/mti1030014 2017

  47. [48]

    Trends in Cognitive Sciences , author =

    Mind. Trends in Cognitive Sciences , author =. 2021 , pages =. doi:10.1016/j.tics.2020.11.009 , language =

    work page doi:10.1016/j.tics.2020.11.009 2021

  48. [49]

    iScience , author =

    The collaborative mind: intention reading and trust in human-robot interaction , volume =. iScience , author =. 2021 , pages =. doi:10.1016/j.isci.2021.102130 , abstract =

    arXiv 2021

  49. [50]

    Journal of Artificial Intelligence Research , author =

    Compositionality. Journal of Artificial Intelligence Research , author =. 2020 , pages =. doi:10.1613/jair.1.11674 , abstract =

    doi:10.1613/jair.1.11674 2020

  50. [51]

    Transformers are

    Melo, Luckeciano C , file =. Transformers are

  51. [52]

    Evaluating

    Lanctot, Marc and Larson, Kate and Bachrach, Yoram and Marris, Luke and Li, Zun and Bhoopchand, Avishkar and Anthony, Thomas and Tanner, Brian and Koop, Anna , month = dec, year =. Evaluating. doi:10.48550/arXiv.2312.03121 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2312.03121

  52. [53]

    Developing, Evaluating and Scaling Learning Agents in Multi-Agent Environments

    Gemp, Ian and Anthony, Thomas and Bachrach, Yoram and Bhoopchand, Avishkar and Bullard, Kalesha and Connor, Jerome and Dasagi, Vibhavari and Vylder, Bart De and Duenez-Guzman, Edgar and Elie, Romuald and Everett, Richard and Hennes, Daniel and Hughes, Edward and Khan, Mina and Lanctot, Marc and Larson, Kate and Lever, Guy and Liu, Siqi and Marris, Luke an...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2209.10958

  53. [54]

    Autonomous Agents and Multi-Agent Systems , author =

    Formal contracts mitigate social dilemmas in multi-agent reinforcement learning , volume =. Autonomous Agents and Multi-Agent Systems , author =. 2024 , pages =. doi:10.1007/s10458-024-09682-5 , abstract =

    work page doi:10.1007/s10458-024-09682-5 2024

  54. [55]

    Journal of Artificial Intelligence Research , author =

    Learning to. Journal of Artificial Intelligence Research , author =. 2024 , pages =. doi:10.1613/jair.1.15167 , abstract =

    work page doi:10.1613/jair.1.15167 2024

  55. [56]

    Journal of Artificial Intelligence Research , author =

    A. Journal of Artificial Intelligence Research , author =. 2023 , pages =. doi:10.1613/jair.1.14174 , abstract =

    doi:10.1613/jair.1.14174 2023

  56. [57]

    White, Colin and Safari, Mahmoud and Sukthanker, Rhea and Ru, Binxin and Elsken, Thomas and Zela, Arber and Dey, Debadeepta and Hutter, Frank , month = jan, year =. Neural. doi:10.48550/arXiv.2301.08727 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2301.08727

  57. [58]

    and Barto, Andrew , year =

    Sutton, Richard S. and Barto, Andrew , year =. Reinforcement learning: an introduction , isbn =

  58. [59]

    and Barto, Andrew G

    Sutton, Richard S. and Barto, Andrew G. , year =. Reinforcement learning: an introduction , isbn =

  59. [61]

    Royal Society Open Science , author =

    General intelligence requires rethinking exploration , volume =. Royal Society Open Science , author =. 2023 , pages =. doi:10.1098/rsos.230539 , abstract =

    work page doi:10.1098/rsos.230539 2023

  60. [62]

    and Islam, Usman and Willis, Richard and Sunehag, Peter , month = dec, year =

    Du, Yali and Leibo, Joel Z. and Islam, Usman and Willis, Richard and Sunehag, Peter , month = dec, year =. A. doi:10.48550/arXiv.2312.05162 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2312.05162

  61. [63]

    , month = oct, year =

    Ivanov, Dima and Dütting, Paul and Talgam-Cohen, Inbal and Wang, Tonghan and Parkes, David C. , month = oct, year =. Principal-

  62. [64]

    Mediated

    Ivanov, Dmitry and Zisman, Ilya and Chernyshev, Kirill , month = jun, year =. Mediated. doi:10.5555/3545946.3598618 , abstract =

    arXiv

  63. [65]

    Autonomous Agents and Multi-Agent Systems , author =

    Warmth and competence in human-agent cooperation , volume =. Autonomous Agents and Multi-Agent Systems , author =. 2024 , pages =. doi:10.1007/s10458-024-09649-6 , abstract =

    work page doi:10.1007/s10458-024-09649-6 2024

  64. [66]

    Cognition , author =

    Beyond the matrix:. Cognition , author =. 2025 , pages =. doi:10.1016/j.cognition.2024.105993 , abstract =

    arXiv 2025

  65. [67]

    and Everett, Richard and Weidinger, Laura and Isaac, William S

    Köster, Raphael and McKee, Kevin R. and Everett, Richard and Weidinger, Laura and Isaac, William S. and Hughes, Edward and Duéñez-Guzmán, Edgar A. and Graepel, Thore and Botvinick, Matthew and Leibo, Joel Z. , month = dec, year =. Model-free conventions in multi-agent reinforcement learning with heterogeneous preferences , url =

  66. [68]

    Cooperative

    Hadfield-Menell, Dylan and Dragan, Anca and Abbeel, Pieter and Russell, Stuart , month = feb, year =. Cooperative

  67. [69]

    IEEE Robotics and Automation Letters , author =

    Long-. IEEE Robotics and Automation Letters , author =. 2022 , pages =. doi:10.1109/LRA.2021.3124524 , abstract =

    arXiv 2022

  68. [70]

    Current Robotics Reports , author =

    A. Current Robotics Reports , author =. 2020 , pages =. doi:10.1007/s43154-020-00019-0 , abstract =

    doi:10.1007/s43154-020-00019-0 2020

  69. [71]

    Social Cognitive and Affective Neuroscience , author =

    Computational modelling of social cognition and behaviour—a reinforcement learning primer , copyright =. Social Cognitive and Affective Neuroscience , author =. 2020 , pages =. doi:10.1093/scan/nsaa040 , abstract =

    work page doi:10.1093/scan/nsaa040 2020

  70. [72]

    Trends in Cognitive Sciences , author =

    The computational challenge of social learning , volume =. Trends in Cognitive Sciences , author =. 2021 , pages =. doi:10.1016/j.tics.2021.09.002 , language =

    work page doi:10.1016/j.tics.2021.09.002 2021

  71. [73]

    Computational Intelligence and Neuroscience , author =

    A. Computational Intelligence and Neuroscience , author =. 2022 , pages =. doi:10.1155/2022/2341898 , abstract =

    work page doi:10.1155/2022/2341898 2022

  72. [74]

    Adapting a

    Ritschel, Hannes and Baur, Tobias and Andre, Elisabeth , month = aug, year =. Adapting a. 2017 26th. doi:10.1109/ROMAN.2017.8172330 , abstract =

    arXiv 2017

  73. [75]

    IEEE Transactions on Affective Computing , author =

    Training. IEEE Transactions on Affective Computing , author =. 2022 , pages =. doi:10.1109/TAFFC.2022.3190233 , abstract =

    arXiv 2022

  74. [76]

    Qureshi, Ahmed Hussain and Nakamura, Yutaka and Yoshikawa, Yuichiro and Ishiguro, Hiroshi , month = may, year =. Show,. 2017. doi:10.1109/ICRA.2017.7989193 , abstract =

    arXiv 2017

  75. [77]

    Metin and Yemez, Yucel , month = aug, year =

    Hussain, Nusrah and Erzin, Engin and Sezgin, T. Metin and Yemez, Yucel , month = aug, year =. Speech

  76. [78]

    Metin and Yemez, Yucel , month = aug, year =

    Hussain, Nusrah and Erzin, Engin and Sezgin, T. Metin and Yemez, Yucel , month = aug, year =. Batch

  77. [79]

    Russell, Stuart J. and Norvig, Peter and Chang, Ming-wei and Devlin, Jacob and Dragan, Anca and Forsyth, David and Goodfellow, Ian and Malik, Jitendra and Mansinghka, Vikas and Pearl, Judea and Wooldridge, Michael J. , year =. Artificial intelligence: a modern approach , isbn =

  78. [80]

    Human-to-

    Bahl, Shikhar and Gupta, Abhinav and Pathak, Deepak , month = jul, year =. Human-to-

  79. [81]

    IEEE Robotics and Automation Letters , author =

    Autonomy in. IEEE Robotics and Automation Letters , author =. 2021 , pages =. doi:10.1109/LRA.2021.3100603 , abstract =

    arXiv 2021

  80. [82]

    Efficient

    Nikolaidis, Stefanos and Ramakrishnan, Ramya and Gu, Keren and Shah, Julie , month = mar, year =. Efficient. Proceedings of the. doi:10.1145/2696454.2696455 , abstract =

    arXiv

Showing first 80 references.