Pith. sign in

REVIEW 4 major objections 5 minor 37 references

A guided graph diffusion model can recover near-optimal Nash bargaining allocations from agent features alone, with no Pareto-frontier solver at inference.

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

2026-08-02 17:03 UTC pith:QR5P5BTX

load-bearing objection A promising architecture with informative ablations, but the headline numbers are not credible until the test-set tuning is fixed and the theory is aligned with the evaluated algorithm. the 4 major comments →

arxiv 2603.29297 v1 pith:QR5P5BTX submitted 2026-03-31 cs.GT

Differentiable Normative Guidance for Nash Bargaining Solution Recovery

Moirangthem Tiken Singh , Surajit Borkotokey , Rajnish Kumar This is my paper
classification cs.GT MSC 91A1268T07
keywords automated negotiationNash bargaining solutiondiffusion modelsindividual rationalitygraph attentionnormative guidanceutility allocationPareto efficiency
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.

This paper tries to show that a generative model can do what classical bargaining solvers need the whole Pareto frontier to do: produce an allocation close to the Nash bargaining solution from only the agents' features. The authors argue that training on human negotiation data reproduces biased, interior outcomes, so they add a differentiable normative loss to a graph-conditioned diffusion generator that penalizes individual-rationality violations and Nash-product shortfalls during the final denoising steps. If right, this means negotiation agents could output fair, individually rational splits in one shot, with no frontier geometry needed at inference. Across synthetic and two human corpora, they report 100% IR compliance and Nash efficiencies of 99.45%, 54.24%, and 88.67%, far above unconstrained generative baselines. They also prove, for a continuous analogue, that the guided process enters the individually rational region in finite time and, under score-matching assumptions, converges to the Nash bargaining solution.

Core claim

The central claim is that steering a diffusion model with a smooth, differentiable composite loss can enforce normative bargaining axioms during generation. The loss combines a log-barrier on the Nash product, a softplus penalty for falling below disagreement points, and a frontier-radius penalty; applied only in the final fraction of reverse steps, it pulls samples from the empirically biased training distribution onto the Pareto arc near the NBS. The authors also claim that, with enough penalty weight, the continuous-time guided process reaches the individually rational region in finite time and, under score-matching assumptions, converges to the NBS as noise collapses.

What carries the argument

The load-bearing object is the guidance loss: three smooth softplus-based terms—a Nash log-barrier, a strict IR penalty with margin, and a frontier-adherence penalty—whose gradients steer the clean utility estimate at each low-noise denoising step. It is coupled with a graph attention encoder that maps each agent's features and disagreement point into a permutation-invariant embedding, and a diffusion denoiser that generates the joint utility vector; the guidance is active only in the last fraction of steps, which the experiments identify as the dominant factor.

Load-bearing premise

The theoretical guarantees hold for a continuous process with guidance on at every step, but the implemented system turns guidance on only near the end; if that scheduling gap matters, the perfect IR compliance is an experimental result, not a proven one.

What would settle it

Run the inference procedure with guidance active for every reverse step on the synthetic set; the paper's own ablation already shows IR compliance dropping to 99.8%, so a sharper test is to find any test instance where the late-activated discrete schedule violates IR despite the penalty weight exceeding the proof's stated bound—or to compute the Lyapunov function along the discrete trajectory and observe it increasing inside the supposedly attractive region.

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

If this is right

  • Generative negotiation agents could produce NBS-approximate outcomes without solving for the Pareto frontier at inference, removing a major obstacle for real corpora.
  • A model trained mostly to reproduce human data can escape biased, interior allocations when a normative objective is injected during generation.
  • The activation window for guidance matters more than the exact loss weights; hard-constraint guidance from step zero hurts both efficiency and IR compliance.
  • Post-hoc projection onto the frontier does not recover Nash-optimal allocations from a biased sample, so iterative in-generation steering is necessary.
  • On discrete-item negotiations, the method is close to full guidance from the start, suggesting the benefit of late activation depends on domain structure.

Where Pith is reading between the lines

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

  • If the finite-time IR guarantee can be extended from the continuous, always-on ODE to the discrete, partially guided schedule actually deployed, the 100% IR results would become a certified property rather than a tuned outcome; the paper leaves that extension open.
  • The guidance loss is parameterized only through disagreement points and budget/priority features, so the same machinery could be tested in multilateral settings with more than two agents, where the Nash product has more terms and gradient variance; the paper identifies this as future work.
  • Because the guidance step size saturated at the upper search boundary, the reported efficiencies are lower bounds under the tested grid; larger step sizes or adaptive schedules may improve the human-corpus numbers.
  • A testable extension is counterfactual fairness auditing: equalize disagreement points in the graph encoder and measure how much outcome asymmetry is attributable to structural imbalance—an audit the architecture makes natural.

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. This paper proposes a guided graph diffusion framework for generating utility allocations in bilateral negotiations. A GATv2 encoder builds a context embedding from agent features; a conditional MLP denoiser with DDIM sampling produces candidate allocations; a composite guidance loss (Eq. 11) combining a softplus Nash-product term, an IR penalty and a frontier penalty is applied during the last t_start fraction of reverse steps. The paper claims finite-time IR convergence and asymptotic NBS convergence (Lemma 3.1, Theorem 3.2) for a continuous full-time ODE, and reports high empirical performance: 99.45% Nash efficiency on a synthetic NTU dataset, 54.24% on CaSiNo, 88.67% on Deal or No Deal, with 100% IR compliance, substantially above unconstrained generative baselines.

Significance. The core idea is attractive: use differentiable normative constraints to steer a generative model away from biased human data. If the claims hold, the work would be a meaningful step toward inference-time game-theoretic compliance in automated negotiation. Strengths include the synthetic oracle benchmark, ablations isolating late activation and post-hoc projection, and the use of non-parametric significance tests on per-instance Nash products. However, the current empirical support is weakened by test-set hyperparameter selection and by a gap between the theoretical ODE setting and the discrete, late-guidance algorithm that is actually evaluated. The paper also releases no code or artifacts, so the sensitivity analyses are not independently checkable.

major comments (4)
  1. [§4.3, Tables 5–6] Section 4.3 states that the grid search over λ, t_start, α, β, γ, and S was 'evaluated on every test split using the full suite of metrics, and the best-performing parameter set for each dataset was selected via a composite objective.' Table 5 reports the selected configurations and Table 6 reports the paper's headline numbers. This is selection on the test set; with roughly 320,000 configurations, best-of-test selection can materially inflate every metric. Consequently 99.45% / 54.24% / 88.67% / 100% IR are not unbiased estimates of performance on new data. The 'domain-invariant λ' conclusion is also a selection artifact because λ hits the grid boundary. The paper must either use the available validation split for selection and evaluate the final model once on test, or perform nested cross-validation.
  2. [§3.4 vs. Algorithm 1] Lemma 3.1 and Theorem 3.2 are proved for the continuous reverse-time ODE in Eq. (13), with -∇u L_guide active over the entire trajectory and with a perfect score assumption. Algorithm 1 applies L_guide only when t/T < t_start (line 7), uses S discrete DDIM steps, gradient normalization, clipping, and projection. None of these discrete/numerical aspects appear in the proofs. Thus the 100% IR guarantee and NBS convergence are not established for the implemented procedure; late activation could break the Lyapunov argument. The assertion in §3.4 that the final utility approximates NBS with an L2 score-matching error bound is also stated without proof. Please either revise the theorems to cover the discrete late-guidance algorithm or clearly label them as idealized and provide finite-sample support.
  3. [§4.4 and Eq. (11)] The primary metric 'Nash Efficiency' is the ratio of generated Nash product to oracle NBS Nash product. The guidance loss's first term in Eq. (11), -α Σ log(softplus(ui - di) + ε), is essentially maximizing the log Nash product. Therefore high Nash efficiency is partly a direct consequence of optimizing the evaluation metric, not an emergent recovery property. This is not circular in a destructive sense, because NBS is characterized by Nash-product maximization, but the paper needs independent metrics (e.g., distance to true NBS, a Pareto-efficiency rate, or other fairness indices) and should avoid claiming 'recovery' solely from Nash efficiency. The problem is amplified by the paper's own admission in §6 that frontier distance only penalizes outward overshoot, leaving Nash efficiency as the only reliable metric.
  4. [§4.1, §4.4, Table 7] Section 4.1 describes the SLSQP-computed NBS reference only for the synthetic NTU dataset, but §4.4 and the Table 7 footnote state that Nash Efficiency is normalized per-instance by the SLSQP solution u*_NBS for all datasets. How are the Pareto frontier and NBS computed for CaSiNo and Deal-or-No-Deal, where the true feasible set is not available? If the assumed feasible set is the Euclidean ball F = {u ≥ 0, ||u|| ≤ r} of §3.3, then the efficiency numbers are relative to an artificial geometry, not the underlying item-based bargaining set. Please specify exactly which feasible set and reference NBS are used for each dataset, and report those reference values alongside the generated values in Table 6.
minor comments (5)
  1. [Abstract/Introduction] Duplicate word: 'Autonomous artificial intelligence agents agents are increasingly being deployed'.
  2. [References] References [13] and [31] are the same paper (Brody, Alon, Yahav) and should be consolidated.
  3. [Declaration] The Declaration of Competing Interest and the AI-assistance declaration refer to 'the author' in singular, although there are three authors.
  4. [Table 5] The composite objective in the Table 5 caption uses 'Nash' without specifying whether it is Nash Product or Nash Efficiency; this should be clarified because the two are related but not identical.
  5. [§5.1] The sensitivity score 'range/mean' is not precisely defined; state whether the range is over grid values or over the resulting metric values, and how averaging over four metrics is performed.

Circularity Check

3 steps flagged

Headline Nash-efficiency numbers are test-set-tuned, and the guidance loss optimizes the same Nash-product metric used to report them; the central prediction is not independent.

specific steps
  1. fitted input called prediction [Section 4.3, Hyperparameter Search (Tables 5 and 6)]
    "An extensive grid of configurations was evaluated on every test split using the full suite of metrics, and the best-performing parameter set for each dataset was selected via a composite objective that balances Nash efficiency, individual rationality compliance, and frontier proximity."

    The grid search is evaluated on every test split, and the composite selection objective includes Nash efficiency and IR compliance, which are exactly the metrics reported in Table 6. The headline figures (99.45%, 54.24%, 88.67%) are therefore best-of-test selections, not predictions from a fixed model on held-out data. The claim that lambda is domain-invariant because it hit the grid boundary across all datasets is likewise an artifact of selecting the best test-set value, not evidence of generalization.

  2. self definitional [Section 3.3, Eq. (11), and Section 4.4, Nash Efficiency]
    "Lguide(u,d) = −α Σ_i log(softplus(ui−di)+ϵ) [LN: Nash efficiency] + β Σ_i softplus(di−ui+δ) [LIR: strict IR] + γ·softplus(∥u∥2−r2) [LF: frontier adherence] ... Nash efficiency normalizes the achieved Nash product against the exact SLSQP-computed NBS of each test instance."

    The reported Nash Efficiency is the generated Nash product divided by the NBS Nash product. The first term of Lguide is a smooth monotone approximation of log(Nash product) and is explicitly labeled 'Nash efficiency'. Thus the guidance loss directly maximizes the same quantity used as the evaluation metric. Combined with the feasible-set projection and the known radius r used in Eq. (12) and Table 3, the high Nash-efficiency numbers are substantially achieved by optimizing the metric itself rather than by independently recovering NBS from agent features.

  3. self definitional [Section 3.4, Theorem 3.2 (Eq. 13)]
    "Inside the feasible region F, the IR penalty evaluates to zero, and the Nash term dominates the guidance gradient: ∇uLguide = −α∇u log N(u). ... Therefore, as the noise variance collapses (t→0), the state converges deterministically: lim_{t→0} u_t = arg max_{u∈∂M+∩F} ∏_{i=1}^n (u_i − d_i) = u*NBS."

    Theorem 3.2 assumes perfect score matching and, via Assumption 3.1, that the learned manifold's boundary is the Pareto frontier. The proof then shows that following the gradient of log Nash product converges to the maximizer of the Nash product, which is the definition of NBS. This is a restatement of the guidance construction, not an independent derivation. Moreover, the theorem is stated for a continuous ODE with guidance active for the entire trajectory, while Algorithm 1 activates guidance only when t/T < t_start, so the theorem does not establish the discrete, late-activation procedure actually evaluated.

full rationale

The paper does not rely on questionable self-citations; the references are external and the graph encoder/diffusion training are independent architectural contributions. However, the central empirical claim is not independent. Section 4.3 explicitly selects hyperparameters on every test split using a composite objective that includes the headline metrics, so Table 6 reports the result of tuning on the evaluation data rather than an out-of-sample prediction. This is the strongest circular step. A second, structural circularity is that the guidance loss in Eq. (11) directly approximates log Nash product, and the reported Nash Efficiency is the normalized Nash product; high Nash-efficiency scores are therefore partly obtained by directly optimizing the metric. The theoretical guarantee in Theorem 3.2 is also self-definitional: it assumes the manifold frontier and then proves gradient ascent on log Nash product reaches the Nash-product maximizer. Separately, there is a genuine theory/algorithm mismatch: the proofs analyze a continuous ODE with guidance throughout, while the implemented algorithm applies guidance only in the final fraction of steps; this is a correctness risk rather than a circularity, but it further weakens the inference from theorem to table. Given the test-set selection and the identity between the optimized objective and the measured metric, the headline numbers cannot be taken as independent estimates of NBS recovery. Score 6 reflects partial circularity: the method still learns from data and outperforms unguided baselines, but the reported absolute efficiencies are not established as predictions.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The method does not introduce new physical entities; its load-bearing extras are (i) test-set-tuned guidance hyperparameters, (ii) the radial Euclidean-ball feasible set, and (iii) idealizations about the learned score and manifold. The most fragile additions are the manifold assumptions and the late-activation proof gap.

free parameters (8)
  • lambda (lambda_guide) = 0.35 (all datasets)
    Guidance step size; grid-searched and selected on the test split, saturating the upper boundary of the search grid (§5.1).
  • t_start (guide_start_frac) = 0.25 / 0.40 / 0.30 per dataset
    Activation window for guidance; selected by the test-set composite objective; identified as a dominant sensitivity parameter (§5.1).
  • alpha (alpha_norm) = 10.0 / 150.0 / 10.0
    Nash multiplier in the guidance loss; selected per dataset on the test split (Table 5).
  • beta (beta_ir) = 8.0 / 20.0 / 0.5
    IR penalty coefficient; annealed during Phase 2 and then selected per dataset on the test split (Table 5).
  • gamma (gamma_frontier) = 15.0 / 10.0 / 10.0
    Frontier penalty coefficient; selected per dataset on the test split (Table 5).
  • S (DDIM steps) = 15 / 10 / 10
    Number of inference steps; fewer steps selected for all datasets to reduce early-chain drift (§5.1).
  • r (Pareto frontier radius) = 1.0
    Hand-set radius defining the feasible set F = {u >= 0, ||u|| <= r}; used in the frontier penalty and projection (Eq. 11, Table 3).
  • delta (IR margin) = 0.05
    Hand-set strictness margin in the softplus IR penalty to enforce ui >= di with a buffer (Eq. 11).
axioms (6)
  • domain assumption Assumption 3.1: The learned data manifold M has a strictly concave, continuous Pareto frontier ∂M+.
    Invoked in the NBS convergence proof (Theorem 3.2); not verified for CaSiNo or Deal or No Deal, where true frontiers are discrete or estimated.
  • domain assumption Assumption 3.2: The unconditional score function is L-Lipschitz and bounded: ||∇u log p_t(u)|| <= M.
    Standard idealization used to bound the IR-penalty condition in Lemma 3.1; not empirically checked for the trained MLP denoiser.
  • domain assumption Assumption 3.3: The Nash product is strictly log-concave over M ∩ F.
    Needed to guarantee a unique global maximum in Theorem 3.2; may fail for non-convex learned manifolds. The paper handwaves that Gaussian smoothing induces convexity at high noise scales.
  • ad hoc to paper The feasible set is a Euclidean ball: F = {u >= 0, ||u|| <= r}.
    The frontier penalty and projection ΠF use a radial ball set by r=1.0; the true Pareto frontier of real datasets need not be radial.
  • domain assumption Complete information: disagreement points and priority weights are fully observed at inference.
    Acknowledged as a limitation (§6, §7); the method does not handle partial observability or adversarial concealment.
  • ad hoc to paper Late activation of guidance does not invalidate the convergence guarantees formulated for full-time guidance.
    Lemma 3.1 and Theorem 3.2 consider an ODE with continuous L_guide, while Algorithm 1 only activates guidance when t/T < t_start; no proof bridges this gap.

pith-pipeline@v1.3.0-alltime-deepseek · 19300 in / 12957 out tokens · 124027 ms · 2026-08-02T17:03:09.409621+00:00 · methodology

0 comments
read the original abstract

Autonomous artificial intelligence agents in negotiation systems must generate equitable utility allocations satisfying individual rationality (IR), ensuring each agent receives at least its outside option, and the Nash Bargaining Solution (NBS), which maximizes joint surplus. Existing generative models often learn suboptimal human behaviors, producing solutions far from Pareto efficiency, while classical methods require full Pareto frontier knowledge, which is unavailable in real datasets. We propose a guided graph diffusion framework that generates individually rational utility vectors while approximating the NBS without frontier knowledge at inference time. Negotiations are modeled as directed graphs with graph attention capturing asymmetric agent attributes, and a conditional diffusion model maps these to utility vectors. A differentiable composite guidance loss, applied in the final reverse diffusion steps, penalizes IR violations and Nash product gaps. We prove that, under sufficient penalty weighting, solutions enter the IR region in finite time. Across datasets, the method achieves 100% IR compliance. Nash efficiency reaches 99.45% on synthetic data (within 0.55 percentage points of an oracle), and 54.24% (CaSiNo) and 88.67% (Deal or No Deal), improving 20-60 percentage points over unconstrained generative baselines.

Figures

Figures reproduced from arXiv: 2603.29297 by Moirangthem Tiken Singh, Rajnish Kumar, Surajit Borkotokey.

Figure 1
Figure 1. Figure 1: System architecture for equitable utility generation. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Hyperparameter sensitivity profile for the Synthetic NTU dataset. [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Joint 2-D grid search over λ (lambda_guide, y-axis) and tstart (guide_start_frac, x-axis) across four evaluation metrics (Synthetic NTU dataset). The dashed blue box marks the base configuration (λ = 0.03, tstart = 0.30); the solid box marks the composite-optimal configuration. Large λ combined with early activation maximizes Nash Product and IR Compliance but increases Frontier Distance; small λ with late… view at source ↗
Figure 4
Figure 4. Figure 4: Visual evaluation of generated utility allocations across the three negotiation domains under optimized guidance [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Without the two stabilization mechanisms described in Section 3.2, applying the log-barrier gradient ∇uˆ0Lguide directly to the DDIM state induces severe trajectory divergence. As uˆ0 approaches the IR bound￾ary (ui → di), the Nash term LN produces unbounded gradients. Empirical tracking confirmed that without the soft-clamp (Fix-B: ut ← clip(ut, −cdrift, cdrift)), the latent state accumulates up to 3.0 un… view at source ↗
Figure 5
Figure 5. Figure 5: Aggregate trajectory statistics (mean ±1 std, n = 30 test cases) for guided (green) and unguided (orange) DDIM chains across the three negotiation domains. The green shaded band indicates the active guidance window (t/T < tstart). Top (Synthetic NTU): IR Compliance (right panel) shows the unguided model suffering from late-stage cumulative drift, while the guided chain maintains strict 1.000 compliance thr… view at source ↗
Figure 6
Figure 6. Figure 6: Latent trajectory dynamics — 4 representative cases (Synthetic NTU, S = 50 visualization steps). Each column is an independent negotiation scenario; guided (green) and unguided (orange) trajectories share identical noise initializations so that differences are attributable solely to ∇uˆ0Lguide. Row 1: Utility-space path of uˆ0 over denoising steps; color intensity encodes progress (light → dark). Stars (⋆)… 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

37 extracted references · 6 canonical work pages

  1. [1]

    C. M. Jonker, R. Aydoğan, T. Baarslag, K. Fujita, T. Ito, K. Hindiks, Automated negotiating agents competition (anac), in: Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, AAAI’17, AAAI Press, 2017, p. 5070–5072

    2017

  2. [2]

    Vaccaro, M

    M. Vaccaro, M. Caosun, H. Ju, S. Aral, J. R. Curhan, Advancing ai negotiations: A large-scale au- tonomous negotiation competition, arXiv preprint arXiv:2503.06416 (2025)

    arXiv 2025

  3. [3]

    J. F. Nash, Jr., The bargaining problem, Econometrica 18 (2) (1950) 155–162. URLhttps://onlinelibrary.wiley.com/doi/abs/0012-9682(195004)18:2&lt;155:TBP&gt;2.0. CO;2-H

    1950

  4. [4]

    Rubinstein, Perfect equilibrium in a bargaining model, Econometrica 50 (1) (1982) 97–109.doi: 10.2307/1912531

    A. Rubinstein, Perfect equilibrium in a bargaining model, Econometrica 50 (1) (1982) 97–109.doi: 10.2307/1912531

    doi:10.2307/1912531 1982

  5. [5]

    H. A. Simon, A behavioral model of rational choice, The Quarterly Journal of Economics 69 (1) (1955) 99–118.doi:10.2307/1884852

    doi:10.2307/1884852 1955

  6. [6]

    Tversky, D

    A. Tversky, D. Kahneman, Judgment under uncertainty: Heuristics and biases, Science 185 (4157) (1974) 1124–1131.doi:10.1126/science.185.4157.1124

    arXiv 1974

  7. [7]

    Kahneman, A

    D. Kahneman, A. Tversky, Prospect theory: An analysis of decision under risk, Econometrica 47 (2) (1979) 263–291.doi:10.2307/1914185

    doi:10.2307/1914185 1979

  8. [8]

    Ntoutsi, P

    E. Ntoutsi, P. Fafalios, U. Gadiraju, V. Iosifidis, W. Nejdl, M.-E. Vidal, S. Ruggieri, F. Turini, S. Pa- padopoulos, E. Krasanakis, et al., Bias in data-driven artificial intelligence systems—an introductory survey, Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 10 (3) (2020) e1356. doi:10.1002/widm.1356

    doi:10.1002/widm.1356 2020

  9. [9]

    S. M. Dinnar, C. Dede, E. Johnson, C. Straub, K. Korjus, Artificial intelligence and technology in teaching negotiation, Negotiation Journal 37 (1) (2021) 65–82.doi:10.1111/nejo.12351

    work page doi:10.1111/nejo.12351 2021

  10. [10]

    Kramár, T

    J. Kramár, T. Eccles, I. Gemp, A. Tacchetti, K. R. McKee, M. Malinowski, T. Graepel, Y. Bachrach, Negotiation and honesty in artificial intelligence methods for the board game of diplomacy, Nature Communications 13 (1) (2022) 7214.doi:10.1038/s41467-022-34473-5

    work page doi:10.1038/s41467-022-34473-5 2022

  11. [11]

    Chawla, J

    K. Chawla, J. Ramirez, R. Clever, G. Lucas, J. May, J. Gratch, Casino: A corpus of campsite negotiation dialogues for automatic negotiation systems, in: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Association for Computational Linguistics, Online, 2021, pp....

    doi:10.18653/v1/2021 2021

  12. [12]

    B. M. Renting, T. Moerland, C. M. Jonker, Towards general negotiation strategies with end-to-end reinforcement learning (2024).arXiv:2406.15096

    Pith/arXiv arXiv 2024

  13. [13]

    Brody, U

    S. Brody, U. Alon, E. Yahav, How attentive are graph attention networks?, in: International Conference on Learning Representations (ICLR), 2022. URLhttps://openreview.net/forum?id=F72ximsx7C1 22

    2022

  14. [14]

    J. Song, C. Meng, S. Ermon, Denoising diffusion implicit models, in: International Conference on Learning Representations (ICLR), 2021. URLhttps://openreview.net/forum?id=St1giarCHLP

    2021

  15. [15]

    Lewis, D

    M. Lewis, D. Yarats, Y. N. Dauphin, D. Parikh, D. Batra, Deal or no deal? end-to-end learning of negotiation dialogues, in: Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing (EMNLP), Association for Computational Linguistics, 2017, pp. 2443–2453

    2017

  16. [16]

    Fisher, W

    R. Fisher, W. L. Ury, B. Patton, Getting to Yes: Negotiating Agreement Without Giving In, Houghton Mifflin, Boston, MA, 1981

    1981

  17. [17]

    R. B. Myerson, Two-person bargaining problems with incomplete information, Econometrica 52 (2) (1984) 461–487.doi:10.2307/1911499

    work page doi:10.2307/1911499 1984

  18. [18]

    Baarslag, K

    T. Baarslag, K. Hindriks, M. Hendrikx, A. Dirkzwager, C. Jonker, Decoupling negotiating agents to explore the space of negotiation strategies, in: Novel Insights in Agent-based Complex Automated Negotiation, Springer, 2014, pp. 61–80.doi:10.1007/978-4-431-54758-7_4

    work page doi:10.1007/978-4-431-54758-7_4 2014

  19. [19]

    H. He, D. Chen, A. Balakrishnan, P. Liang, Decoupling strategy and generation in negotiation dia- logues, in: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Process- ing (EMNLP), Association for Computational Linguistics, 2018, pp. 2333–2343.doi:10.18653/v1/ D18-1256

    doi:10.18653/v1/ 2018

  20. [20]

    D. M. Ziegler, N. Stiennon, J. Wu, T. B. Brown, A. Radford, D. Amodei, P. Christiano, G. Irving, Fine-tuning language models from human preferences, arXiv preprint arXiv:1909.08593 (2019). URLhttps://arxiv.org/abs/1909.08593

    Pith/arXiv arXiv 1909

  21. [21]

    K. Cao, A. Lazaridou, M. Lanctot, J. Z. Leibo, K. Tuyls, S. Clark, Emergent communication through negotiation, in: International Conference on Learning Representations (ICLR), 2018. URLhttps://openreview.net/forum?id=Hk6WhagRW

    2018

  22. [22]

    K. Sohn, H. Lee, X. Yan, Learning structured output representation using deep conditional generative models, in: Advances in Neural Information Processing Systems (NeurIPS), Vol. 28, Curran Associates, 2015, pp. 3483–3491. URLhttps://papers.nips.cc/paper/2015/hash/8d55a249e6baa5c06772297520da2051-Abstract. html

    2015

  23. [23]

    Mirza, S

    M. Mirza, S. Osindero, Conditional generative adversarial nets, arXiv preprint arXiv:1411.1784 (2014). URLhttps://arxiv.org/abs/1411.1784

    Pith/arXiv arXiv 2014

  24. [24]

    J. Ho, A. Jain, P. Abbeel, Denoising diffusion probabilistic models, in: Advances in Neural Information Processing Systems (NeurIPS), Vol. 33, Curran Associates, 2020, pp. 6840–6851. URLhttps://arxiv.org/abs/2006.11239

    Pith/arXiv arXiv 2020

  25. [25]

    Dhariwal, A

    P. Dhariwal, A. Q. Nichol, Diffusion models beat GANs on image synthesis, in: Advances in Neural Information Processing Systems (NeurIPS), Vol. 34, Curran Associates, 2021, pp. 8780–8794. URLhttps://arxiv.org/abs/2105.05233

    Pith/arXiv arXiv 2021

  26. [26]

    J. Ho, T. Salimans, Classifier-free diffusion guidance, in: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2022. URLhttps://arxiv.org/abs/2207.12598

    Pith/arXiv arXiv 2021

  27. [27]

    Hoogeboom, V

    E. Hoogeboom, V. G. Satorras, C. Vignac, M. Welling, Equivariant diffusion for molecule generation in 3D, in: Proceedings of the 39th International Conference on Machine Learning (ICML), PMLR, 2022, pp. 8867–8887. URLhttps://arxiv.org/abs/2203.17003

    Pith/arXiv arXiv 2022

  28. [28]

    Dwork, M

    C. Dwork, M. Hardt, T. Pitassi, O. Reingold, R. Zemel, Fairness through awareness, in: Proceedings of the 3rd Innovations in Theoretical Computer Science Conference (ITCS), ACM, 2012, pp. 214–226. doi:10.1145/2090236.2090255. 23

    arXiv 2012

  29. [29]

    Hardt, E

    M. Hardt, E. Price, N. Srebro, Equality of opportunity in supervised learning, in: Advances in Neural Information Processing Systems (NeurIPS), Vol. 29, Curran Associates, 2016, pp. 3315–3323. URLhttps://arxiv.org/abs/1610.02413

    Pith/arXiv arXiv 2016

  30. [30]

    M. J. Kusner, J. R. Loftus, C. Russell, R. Silva, Counterfactual fairness, in: Advances in Neural Information Processing Systems (NeurIPS), Vol. 30, Curran Associates, 2017, pp. 4066–4076

    2017

  31. [31]

    Brody, U

    S. Brody, U. Alon, E. Yahav, How attentive are graph attention networks? (2022).arXiv:2105.14491. URLhttps://arxiv.org/abs/2105.14491

    Pith/arXiv arXiv 2022

  32. [32]

    J. Song, C. Meng, S. Ermon, Denoising diffusion implicit models, arXiv preprint arXiv:2010.02502 (2020)

    Pith/arXiv arXiv 2010

  33. [33]

    Kraft, A software package for sequential quadratic programming, Tech

    D. Kraft, A software package for sequential quadratic programming, Tech. Rep. DFVLR-FB 88-28, Deutsche Forschungs- und Versuchsanstalt für Luft- und Raumfahrt (DFVLR), Köln, Germany (1988)

    1988

  34. [34]

    Virtanen, R

    P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P.Peterson, W.Weckesser, J.Bright, S.J.vanderWalt, M.Brett, J.Wilson, K.J.Millman, N.Mayorov, A.R.J.Nelson, E.Jones, R.Kern, E.Larson, C.J.Carey, İ.Polat, Y.Feng, E.W.Moore, J.VanderPlas, D. Laxalde, J. Perktold, R. Cimrman, I. Henriksen, E. A. Quintero, C. R. Har...

    2020

  35. [35]

    Sandholm, Automated negotiation, Commun

    T. Sandholm, Automated negotiation, Commun. ACM 42 (3) (1999) 84–85.doi:10.1145/295685. 295866. URLhttps://doi.org/10.1145/295685.295866

    arXiv 1999

  36. [36]

    Bellucci, J

    E. Bellucci, J. Zeleznikow, Developing negotiation decision support systems that support mediators: A case study of the Family Winner system, Artificial Intelligence and Law 13 (2) (2006) 233–271. doi:10.1007/s10506-006-9013-1

    work page doi:10.1007/s10506-006-9013-1 2006

  37. [37]

    M. E. Schweitzer, R. Croson, Curtailing deception: The impact of direct questions on lies and omissions, International Journal of Conflict Management 10 (3) (1999) 225–248.doi:10.1108/eb022826. 24

    work page doi:10.1108/eb022826 1999