Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

KABB: Knowledge-Aware Bayesian Bandits for Dynamic Expert Coordination in Multi-Agent Systems

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

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

Pith's one-line read KABB: knowledge-aware Bayesian bandits route each query to 2–3 experts and report a 77.9% length-controlled win rate on AlpacaEval 2.0, 9.8 points above the six-proposer MoA baseline at lower cost.

desk verdict Plausible routing idea, but the headline results can't be attributed to KABB until the knowledge graph is specified and the numbers stop contradicting each other. read the letter →

arxiv 2502.07350 v2 pith:OKBBFBGB submitted 2025-02-11 cs.AI

Jusheng Zhang , Zimeng Huang , Yijia Fan , Ningyuan Liu , Mingyan Li , Zhuojie Yang , Jiawei Yao , Jian Wang
show 1 more author
Keze Wang
This is my paper · ORCID
classification cs.AI
keywords multi-agentsystemsknowledge-awareBayesianbanditsThompsonsamplingknowledgedistanceexpertroutingcost-performancefrontierAlpacaEval2.0largelanguagemodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that the coordination problem in multi-agent LLM systems—who should answer a given query and how many experts are enough—can be formulated as a knowledge-aware Bayesian bandit. The proposed KABB framework represents tasks and expert profiles as vectors over a fixed concept set, measures the distance between an expert team and a task across semantic overlap, dependency structure, historical success, and team synergy, then selects the top-k experts with a Thompson-sampling rule that injects this distance into a Beta-distribution confidence score. The authors report that this dynamic selection reaches a 77.9% length-controlled win rate on AlpacaEval 2.0, about 9.8 points above the mixture-of-agents baseline under the same model configuration, while calling only 2–3 experts instead of 6 proposers, and that it reaches 9.65 on MT-Bench. The payoff, if true, is that instruction-following quality can be bought with far fewer inference calls by exploiting semantic structure and past performance rather than running every expert on every query.

What carries the argument

The load-bearing object is the knowledge distance metric $Dist(S,t)$: $Dist(S,t)=\log(1+d_t)\,[\omega_1(1-\rho_{\mathrm{overlap}}(S,t))+\omega_2|R_{dep}(S,t)|/K+\omega_3(1-\bar H_S(t))+\omega_4(1-\mathrm{Synergy}(S))]$, a weighted blend of task difficulty, concept-overlap mismatch, dependency-edge complexity, historical failure rate, and team complementarity. It enters the Thompson-sampling confidence value $\tilde\theta_S^{(t)}=(\alpha_S^{(t)}/(\alpha_S^{(t)}+\beta_S^{(t)}))\,e^{-\lambda Dist(S,t)}e^{-\kappa\Delta t}\,\mathrm{Synergy}(S)^\eta$, so exploration is biased toward teams that are semantically close to the task, historically strong, and mutually complementary. The companion dual-adaptation update $\alpha_S^{(t+1)}=\gamma^{\Delta t}\alpha_S^{(t)}+r_S^{(t)}+\delta\,KM(S,t)$ (with the symmetric $\beta$ update) lets the Beta parameters forget stale experience and absorb a knowledge-matching reward, converting subset selection from a combinatorial search into a distance-penalized Bayesian ranking.

What would settle it

Re-run the main AlpacaEval 2.0 comparison with the knowledge graph's concept-to-expert assignments randomly permuted, keeping the same prompts, aggregator, and bandit hyperparameters. If the 77.9% LC win rate and the 9.8-point gap over MoA survive the permutation, the knowledge-distance term is not the source of the gain; a companion run with $\lambda=0$, which turns off the distance penalty in Eq. (6), would isolate how much of the reported routing quality comes from the graph versus from the Beta prior and the aggregator.

Watch

Extended reading notes

Core claim

The central claim is that a multi-agent LLM system can be coordinated by a bandit whose exploration-exploitation trade-off is modulated by a knowledge-distance metric, and that this yields better instruction-following quality at a fraction of the inference cost of running every expert. Concretely, KABB represents tasks by concept vectors $d_t$, experts by capability vectors $v_e$, defines a team-task distance $Dist(S,t)$ that combines semantic overlap, dependency structure, historical success, and team synergy, and uses a Beta-distribution Thompson sampler with exponential time decay and a knowledge-matching reward to pick the top-$k$ experts. The reported evidence is that the resulting system reaches a 77.9% length-controlled win rate on AlpacaEval 2.0 versus 68.1% for the MoA baseline, a 9.65 average on MT-Bench versus 9.41, and per-instruction costs near the Pareto frontier while selecting only 2 experts (or 3 in the variant analysis) instead of 6 proposers. The paper also argues, via ablation, that the knowledge-aware routing component and the MAB optimizer each contribute to these gains beyond a classifier-based router and beyond PPO, MCTS, and A2C alternatives.

Load-bearing premise

The load-bearing premise is that the hand-authored mapping of tasks and experts onto 12 knowledge concepts, 24 experts, and graph dependencies genuinely captures which LLMs are good at which tasks; if those encodings are arbitrary or noisy, the knowledge-distance penalty in Eq. (6) would not encode real expertise and the reported win rates could come from prompt specialization and the aggregator rather than from the bandit mechanism.

Editorial extensions

If this is right

  • If the central claim holds, instruction-following systems can cut inference calls by roughly two-thirds to three-quarters: top-2 or top-3 experts replace all six proposers, with a reported 9.8-point LC win-rate gain over MoA.
  • The reported MT-Bench score of 9.65 places the routed ensemble above every single model in the comparison table, implying that routing plus aggregation can add capability over the strongest member model even when that model is not routed to every query.
  • The ablation results (KA versus classifier routing, MAB versus PPO/MCTS/A2C) imply that both the knowledge graph and the bandit's exploration-exploitation balance contribute to the routing gains, not the prompt engineering alone.
  • The cost analysis implies that dynamic expert count creates a deployment dial: fewer experts for cost-sensitive settings and more experts to approach the strongest single-model win rate, with reported per-instruction costs near the Pareto frontier.

Reading between the lines

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

  • Beyond the paper: if the concept vectors and graph edges were learned from logged routing outcomes instead of hand-assigned, the knowledge signal would become directly testable; the current paper does not specify how its 12 concepts, 24 experts, and dependencies are built.
  • Beyond the paper: the cost-per-instruction figure counts selected experts' API prices; including the router's concept-vector computation and the aggregator's synthesis calls would move the frontier modestly, since those calls are cheaper than the six proposers they replace.
  • Beyond the paper: the appendix's note that multi-agent systems can interfere on closed deterministic tasks suggests a hybrid policy—single model for well-specified queries, bandit ensemble for open-ended instructions—as a natural deployment extension.
  • Beyond the paper: re-running the Routing Alignment Score protocol on prompts outside AlpacaEval's distribution would test whether the knowledge-graph concept layer generalizes or is tuned to the benchmark's prompt styles.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. KABB introduces a multi-agent expert-selection framework that combines a knowledge graph over concepts and experts, a five-term knowledge distance metric, a dynamic Bayesian bandit (Thompson sampling), and an aggregator. Tasks are parsed into concept vectors, experts into capability vectors, and a knowledge-distance penalty plus a knowledge-matching reward are injected into Beta-distribution updates and into the sampling score. The authors report a 77.9% LC win rate on AlpacaEval 2.0, 9.65 on MT-Bench, and lower cost than MoA, and they provide ablations (knowledge-aware vs classifier routing, MAB vs PPO/MCTS/A2C), human-annotated routing metrics, and theoretical convergence claims. The appendices contain proofs, parameter sensitivity studies, and case studies.

Significance. If the central claim were substantiated, this would be a useful contribution: integrating semantic structure into bandit-based expert selection for LLM ensembles is timely, and the reported cost-performance trade-off (Figures 4-5) is attractive. The paper deserves credit for attempting to validate routing quality with human-annotated RAS and PWRS, for covering several benchmarks, and for including case studies that show both successful and unsuccessful expert configurations. However, the current manuscript does not allow the central claim to be verified: the knowledge graph that defines the core routing signal is never specified, the experimental configurations are internally inconsistent, and the theoretical guarantees are asserted rather than proved. The contribution is therefore not yet reproducible from the text.

major comments (5)
  1. [§3.1-3.2, §4.1, Appendix F] The knowledge graph on which Eqs. (4)-(6) are based is never concretely specified. Section 4.1 states only that "12 knowledge concepts and 24 experts are defined" and that models are "evenly distributed across these experts using tailored prompts," but the concepts, the expert-concept assignment, the graph edges, the dependency relation R_dep, and the learned weights ω are not given. Appendix F provides only generic persona prompts for "Analysis Expert," "Strategy Expert," and the aggregator, with no mapping to the 12 concepts. Because Dist(S,t) and KM(S,t) are computed entirely from this unspecified graph, a reader cannot determine whether the reported 77.9% LC win rate reflects the knowledge-aware bandit mechanism or simply prompt specialization plus the aggregator. The authors must provide the full graph construction, the hyperparameters (λ, δ, η, κ, threshold, initial ω), and the code or data to make the claimed attribution testable.
  2. [§4.1, §4.2, Table 2, Figure 2] The number of selected experts/concepts is inconsistent across the paper. Section 4.1 states that the system "dynamically routes queries to top-3 experts from top-2 knowledge concepts," while §4.2 says "KABB selects only 2 experts to respond to instruction," Table 2 says the system "dynamically routes queries to the top-2 experts derived from the top-2 knowledge concepts," and the Figure 2 caption describes "top-4 relevant experts." Since the paper's cost-effectiveness claim depends directly on the number of selected experts, the authors must specify exactly which configuration produced each result in Table 1, Table 2, and Figures 4-5. Relatedly, MT-Bench is reported as 9.65 in the abstract and Table 1 but 9.60 in §4.2; these numbers must be reconciled.
  3. [Theorem 3.3, Appendix G.2, G.8, G.9] Theorem 3.3 asserts, for any ε > 0, existence of a parameter configuration (λ*, η*, γ*) such that R(T) ≤ εT + O(√(T log T)), but no construction or explicit dependence of the O-term on ε is provided. The proof sketch in Appendix G.2 lists a Lyapunov function and UCB-style analysis without deriving the claimed bound, and the supplementary bounds in G.8-G.9 (Eq. (7), Lemma G.6, Theorem G.7) contain conflicting forms of the regret and do not state the needed assumptions (stationarity of rewards, boundedness, number of arms versus subsets). As written, the theorem is a non-constructive existence claim rather than a usable convergence guarantee. Please either give a complete proof with explicit parameter choices and assumptions, or state the result as a heuristic motivation and remove the word "Theorem."
  4. [Table 2, §4.3] The central routing ablation is presented without any measure of uncertainty. Table 2 reports a 1.5-point LC win-rate gap between KA (MAB) at 62.4 and CL (MAB) at 60.9, with no error bars, no number of repeated runs, and no significance test. The RAS and PWRS metrics also rely on human annotations with only 10% overlap between annotators, and no variance is reported for the preference scores. Please provide run-to-run variability (e.g., standard errors across multiple evaluations or bootstrap intervals) for all headline numbers, at least for the KABB vs MoA and KA vs CL comparisons, so the reader can judge whether the reported gaps are statistically meaningful.
  5. [Eqs. (5) and (6)] There is a potential feedback loop between the knowledge-matching reward and the sampling score. In Eq. (5), KM(S,t) is added to α_S and β_S, while in Eq. (6) the same Beta expectation α/(α+β) is multiplied by a knowledge-distance penalty that includes ρ_overlap and Synergy(S). Since KM itself contains ρ_overlap·Synergy(S), the "knowledge" correction partially reflects the algorithm's own accumulating statistics rather than an externally fixed semantic signal. The authors should provide an experiment with a fixed, pre-specified knowledge graph (no online updates of the graph weights and no knowledge reward derived from bandit parameters) to isolate the contribution of the knowledge-aware mechanism.
minor comments (5)
  1. [§4.2, Figure 3] The benchmark name is misspelled as "FLAS-Hard" in the main text, and the Figure 3 caption refers to "Qwen2-70B-Instruct" while the text and Table 1 use "Qwen2-72B-Instruct."
  2. [§4.3] The paragraph describing the classifier-based routing contains a duplicated sentence: "We replaced our Knowledge-Aware (KA) routing mechanism with a classifier-based routing (CL) approach. To be specific, We replaced our Knowledge-Aware (KA) routing mechanism..." Please remove the duplicate.
  3. [Appendix E] In the text following Table 5, "MABB succeeded in selecting qualified experts" should read "KABB."
  4. [Appendix D.2] The sentence "with one concept and one expect selected for instruction" should read "one concept and one expert selected."
  5. [General] The abstract promises that source code will be released, but no code, knowledge-graph data, or hyperparameter configuration file accompanies the submission. Please either provide these artifacts or state explicitly that they are omitted from the review version.

Circularity Check

2 steps flagged · score 6.0 of 10

Partial circularity: the knowledge-distance sampling score re-uses the algorithm's own Beta posterior mean, and the appendix validates KABB with metrics defined from that same internal state; the headline AlpacaEval win rate is external and not itself circular.

  1. self definitional [Eq. (4) and Eq. (6), Sections 3.2 and 3.3]
    "Dist(S, t) = log(1 + d_t) · [ω1 (1−ρ_overlap(S,t)) + ω2 |R_dep(S,t)|/K + ω3 (1−\bar H_S(t)) + ω4 (1−Synergy(S))] (4) ... where \bar H_S(t) is average historical success rate of expert subset. ... ˜θ(t)_S = E[θ(t)_S] · exp(−λ·Dist(S,t))·γ^{Δt}·Synergy(S)^η = (α(t)_S/(α(t)_S+β(t)_S))·exp(−λ·[log(1+d_t)·Σ_{i=1}^4 ω_i Ψ_i])·... (6)"

    The sampling score in Eq. (6) already contains the Beta posterior mean h = α/(α+β) as the 'historical expectation'. Eq. (4) places the same historical success rate inside the 'knowledge distance' penalty through ω3(1−\bar H_S(t)). Thus the distance is not an independent semantic signal: the score is h·exp(λω3h) up to the other terms, so the knowledge-aware correction partially re-weights the algorithm's own posterior estimate. Because the CL baseline uses only Sentence-BERT cosine similarity and no posterior, the KA-vs-CL gap in Table 2 conflates this self-referential double-counting with genuine knowledge-graph information.

  2. fitted input called prediction [Appendix B.1, Supplementary Performance Evaluation]
    "Historical performance prediction accuracy, using the dynamic weight α/(α+β) (where α and β represent the number of successful and failed tasks, respectively), with a prediction error ≤ 0.1 considered correct."

    The quantities α and β are exactly the parameters that KABB itself updates in Eq. (5) from observed successes/failures and the knowledge-matching pseudo-count. Calling α/(α+β) a 'historical performance prediction' and scoring its prediction error means the evaluation measures the model's own internal posterior state rather than an externally specified predictive target. MoA and MoA-lite do not maintain this Bernoulli posterior, so the reported 85.2% versus 70.1% versus 45.5% advantage is partly guaranteed by the fact that the metric is defined from KABB's own components.

full rationale

The headline result on AlpacaEval 2.0 is a measured benchmark outcome against GPT-4 judges, so it is not itself circular; the 77.9% LC win rate is externally observable and does not reduce to the method's equations. However, the paper's mechanism-level claims are partially circular. Eq. (4) defines the knowledge distance using the same historical success rate that Eq. (6) already uses as the Beta posterior mean, so the 'knowledge-aware' penalty is partly a function of the algorithm's own estimate rather than an independent semantic signal. The supplementary validation in Appendix B compounds this by defining 'historical performance prediction accuracy' directly on α/(α+β), the very state variable KABB maintains, while the baselines have no corresponding quantity. The knowledge-graph construction being unspecified (12 concepts, 24 experts, graph edges, and learned weights are never listed) is a reproducibility and correctness risk, but it is not circularity by itself and was not counted as a circular step. There is no load-bearing self-citation chain: the authors do not rely on their own prior theorems to justify the framework. The empirical material, internal contradictions (top-2 vs top-3 experts, MT-Bench 9.60 vs 9.65), and absence of error bars belong to correctness risk rather than circularity. Overall, the central external benchmark is independent, but the claimed advantage of the knowledge-aware mechanism over classifier routing is partly built from the method's own posterior, giving a partial-circularity score of 6.

Assumptions & free parameters 6 free parameters · 6 assumptions · 3 invented entities

The KABB model rests on a hand-authored knowledge graph, assumed concept and capability encodings, and a postulated Beta evolution. The theoretical analysis adds ad hoc stability and uniform-gap assumptions, while the empirical claims depend on threshold and hyperparameter choices that are selected from sensitivity scans without a held-out split. No code or data artifacts are provided for independent verification.

free parameters (6)
  • omega_1..omega_4 (learnable distance weights) = not reported (sum to 1)
    Equation (4) balances semantic mismatch, dependency, historical effectiveness, and synergy; trained via deep learning with no training details.
  • lambda (knowledge distance penalty) = not reported
    Used in the Thompson sampling score (Eq 6) and in the UCB appendix; no value or learning schedule is given.
  • delta (knowledge-matching reward strength) = not reported
    Equation (5) controls how much KM(S,t) alters the Beta parameters; no value is reported.
  • eta (synergy exponent) = not reported
    Equation (6) includes Synergy(S)^eta; Appendix G.12 asserts eta > 1 but gives no value.
  • kappa (time decay rate in gamma_dt = e^{-kappa dt}) = not reported
    Equation (5) uses exponential decay; Appendix B.2 varies a time decay factor between 0.2 and 1.0 but does not report kappa directly.
  • knowledge distance threshold = 0.75 in Appendix B.2
    The sensitivity scan in Appendix B.2 selects 0.75 as the best threshold and this value is also used for the F1 computation; the selection appears post hoc.
assumptions (6)
  • domain assumption There exists a knowledge graph whose concept nodes, expert nodes, and dependency edges accurately represent the semantic structure of tasks and experts.
    Section 3.1 and Eq (4) rely on graph topology for concept vectors and dependency paths, but the paper does not specify how the 12 concepts, 24 experts, or edges are constructed.
  • domain assumption Task concept vectors d_t and expert capability vectors v_e can be reliably extracted from natural language.
    Section 3.1 Steps 1 and 2 assume NLP-based concept extraction and capability mapping are available and accurate; no implementation or error analysis is given.
  • domain assumption Expert success probabilities follow a time-varying Beta distribution with exponential decay updates (Eq 5).
    The Bayesian evolution is postulated as the model of expert performance; it is not derived from observations and no goodness-of-fit is reported.
  • ad hoc to paper The system reaches zero selection error asymptotically: lim_{t->infinity} P_e^(t) = 0.
    Appendix G.4 Step 4 assumes stability of the selection error to drop the error-entropy term; this is close to assuming the convergence that the theorem is meant to establish.
  • ad hoc to paper Reward gaps satisfy the uniform exploration hypothesis epsilon = Theta(1/sqrt(K)).
    Appendix G.6 Step 2 uses this to convert classical O(K log T / epsilon^2) complexity into the claimed O(sqrt(K log|C|)) form; no argument supports the scaling.
  • ad hoc to paper The knowledge graph G is fully determined by the expert set E and concept layer C.
    Appendix G.5 Step 3 asserts this to decompose mutual information; the paper gives no formal definition of G as a function of E and C.
invented entities (3)
  • Synergy(S) and pairwise synergy coefficients C_syn(e_i,e_j)
    purpose: Quantify internal team complementarity and modulate both the distance metric (Eq 4) and the Thompson sampling score (Eq 6).
    C_syn is never defined, and no external data or ablation validates that higher Synergy corresponds to better team performance.
  • Knowledge matching index KM(S,t) = rho_overlap * Synergy(S)
    purpose: Adds a synthetic reward term to the Beta parameter updates in Eq (5).
    KM is built from the same overlap and synergy quantities already in the sampling score, so it carries no independent information about future expert success.
  • Predefined concept set C with 12 knowledge concepts
    purpose: Discretizes the task space for concept extraction and routing in the experiments.
    The 12 concepts are hand-picked; no procedure is given for choosing them or for verifying they cover the benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KABB: Knowledge-Aware Bayesian Bandits for Dynamic Expert Coordination in Multi-Agent Systems." pith.science (2026). https://pith.science/paper/OKBBFBGB

@misc{pith2026250207350,
  author       = {Pith},
  title        = {Pith review of: KABB: Knowledge-Aware Bayesian Bandits for Dynamic Expert Coordination in Multi-Agent Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OKBBFBGB}},
  note         = {Machine review of arXiv:2502.07350}
}
read the original abstract

As scaling large language models faces prohibitive costs, multi-agent systems emerge as a promising alternative, though challenged by static knowledge assumptions and coordination inefficiencies. We introduces Knowledge-Aware Bayesian Bandits (KABB), a novel framework that enhances multi-agent system coordination through semantic understanding and dynamic adaptation. The framework features three key innovations: a three-dimensional knowledge distance model for deep semantic understanding, a dual-adaptation mechanism for continuous expert optimization, and a knowledge-aware Thompson Sampling strategy for efficient expert selection. Extensive evaluation demonstrates KABB achieves an optimal cost-performance balance, maintaining high performance while keeping computational demands relatively low in multi-agent coordination.

Figures

Figures reproduced from arXiv: 2502.07350 by the authors.

Figure 1
Figure 1. Comparison of MoA and KABB (Ours) on solv [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The KABB framework combines knowledge graph embeddings, team synergy metrics, and dynamic Bayesian [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Results on FLASK-Hard where we use the default [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: The trade-off between performance and compu [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Relationship between the number of selected experts and selected concepts, and the AlpacaEval 2.0 LC Win Rate. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. A Heuristic Perspective on Debiasing Language Models

    cs.CL 2026-08 conditional novelty 5.0 of 10

    HEIMAT debiases language models by generating heuristic prompts, building substitution sets, and fine-tuning the model with a Jensen-Shannon divergence loss to align predictions across demographic groups, with no fixe...

Reference graph

Works this paper leans on

44 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  3. [3]

    C., Truhn, D., Busch, F., Dorfner, F., Nawabi, J., Makowski, M

    Adams, L. C., Truhn, D., Busch, F., Dorfner, F., Nawabi, J., Makowski, M. R., and Bressem, K. K. Llama 3 challenges proprietary state-of-the-art large language models in radiology board--style examination questions. Radiology, 312 0 (2): 0 e241191, 2024

    work page 2024

  4. [4]

    A., Cao, Y., and Wilson, K

    Alamdari, P. A., Cao, Y., and Wilson, K. H. Jump starting bandits with llm-generated prior knowledge. arXiv preprint arXiv:2406.19317, 2024

    arXiv 2024

  5. [5]

    Qwen technical report

    Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023

    arXiv 2023

  6. [6]

    A decision-language model (dlm) for dynamic restless multi-armed bandit tasks in public health, 2024

    Behari, N., Zhang, E., Zhao, Y., Taneja, A., Nagaraj, D., and Tambe, M. A decision-language model (dlm) for dynamic restless multi-armed bandit tasks in public health, 2024. URL https://arxiv.org/abs/2402.14807

    arXiv 2024

  7. [7]

    Representation learning: A review and new perspectives, 2014

    Bengio, Y., Courville, A., and Vincent, P. Representation learning: A review and new perspectives, 2014. URL https://arxiv.org/abs/1206.5538

    arXiv 2014

  8. [8]

    A survey on mixture of experts, 2024

    Cai, W., Jiang, J., Wang, F., Tang, J., Kim, S., and Huang, J. A survey on mixture of experts, 2024. URL https://arxiv.org/abs/2407.06204

    arXiv 2024

Show all 44 references
  1. [9]

    Frugalgpt: How to use large language models while reducing cost and improving performance

    Chen, L., Zaharia, M., and Zou, J. Frugalgpt: How to use large language models while reducing cost and improving performance. arXiv preprint arXiv:2305.05176, 2023

    2023 arXiv

  2. [10]

    Concept extraction and requirement analysis for progressive knowledge learning

    Chen, Z., Liu, B., Hsu, M., Castellanos, M., and Ghosh, R. Concept extraction and requirement analysis for progressive knowledge learning. In Proceedings of the 27th International Conference on Computational Linguistics, pp.\ 1046--1056, 2018

    2018

  3. [11]

    Dubois, Y., Galambosi, B., Liang, P., and Hashimoto, T. B. Length-controlled alpacaeval: A simple way to debias automatic evaluators. arXiv preprint arXiv:2404.04475, 2024

    2024 arXiv

  4. [12]

    C., Wang, B., Kuo, C.-C

    Ge, X., Wang, Y. C., Wang, B., Kuo, C.-C. J., et al. Knowledge graph embedding: An overview. APSIPA Transactions on Signal and Information Processing, 13 0 (1), 2024

    2024

  5. [13]

    Large language models with knowledge domain partitioning for specialized domain knowledge concentration

    Gong, X., Liu, M., and Chen, X. Large language models with knowledge domain partitioning for specialized domain knowledge concentration. 2024

    2024

  6. [14]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  7. [16]

    V., Wiest, O., and Zhang, X

    Guo, T., Chen, X., Wang, Y., Chang, R., Pei, S., Chawla, N. V., Wiest, O., and Zhang, X. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680, 2024 b

    2024 arXiv

  8. [17]

    and Yang, X

    Guo, Y. and Yang, X. J. Modeling and predicting trust dynamics in human-robot teaming: A bayesian inference approach, 2020. URL https://arxiv.org/abs/2007.13151

    2020 arXiv

  9. [18]

    Measuring mathematical problem solving with the math dataset

    Hendrycks, D., Burns, C., Kadavath, S., Arora, A., Basart, S., Tang, E., Song, D., and Steinhardt, J. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874, 2021

    2021 arXiv

  10. [19]

    Enabling ensemble learning for heterogeneous large language models with deep parallel collaboration

    Huang, Y., Feng, X., Li, B., Xiang, Y., Wang, H., Qin, B., and Liu, T. Enabling ensemble learning for heterogeneous large language models with deep parallel collaboration. CoRR, abs/2404.12715, 2024. doi:10.48550/ARXIV.2404.12715. URL https://doi.org/10.48550/arXiv.2404.12715

    2024 doi arXiv

  11. [20]

    Jiang, D., Ren, X., and Lin, B. Y. Llm-blender: Ensembling large language models with pairwise ranking and generative fusion. arXiv preprint arXiv:2306.02561, 2023

    2023 arXiv

  12. [21]

    Li, L., Chu, W., Langford, J., and Schapire, R. E. A contextual-bandit approach to personalized news article recommendation. In Proceedings of the 19th international conference on World wide web, pp.\ 661--670, 2010

    2010

  13. [22]

    Task-oriented dialog systems that consider multiple appropriate responses under the same context

    Li, Y., Su, H., Shen, X., Li, W., Cao, Z., and Niu, S. Task-oriented dialog systems that consider multiple appropriate responses under the same context. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pp.\ 6493--6500, 2019

    2019

  14. [23]

    Deepseek-v3 technical report

    Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024

    2024 arXiv

  15. [24]

    Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing, 2021

    Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., and Neubig, G. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing, 2021. URL https://arxiv.org/abs/2107.13586

    2021 arXiv

  16. [25]

    Merge, ensemble, and cooperate! a survey on collaborative strategies in the era of large language models

    Lu, J., Pang, Z., Xiao, M., Zhu, Y., Xia, R., and Zhang, J. Merge, ensemble, and cooperate! a survey on collaborative strategies in the era of large language models. arXiv preprint arXiv:2407.06089, 2024

    2024 arXiv

  17. [26]

    Routing to the expert: Efficient reward-guided ensemble of large language models

    Lu, K., Yuan, H., Lin, R., Lin, J., Yuan, Z., Zhou, C., and Zhou, J. Routing to the expert: Efficient reward-guided ensemble of large language models. arXiv preprint arXiv:2311.08692, 2023

    2023 arXiv

  18. [27]

    and Teneketzis, D

    Mahajan, A. and Teneketzis, D. Multi-armed bandit problems. In Foundations and applications of sensor management, pp.\ 121--151. Springer, 2008

    2008

  19. [28]

    P., Mirza, M., Graves, A., Lillicrap, T

    Mnih, V., Badia, A. P., Mirza, M., Graves, A., Lillicrap, T. P., Harley, T., Silver, D., and Kavukcuoglu, K. Asynchronous methods for deep reinforcement learning. CoRR, abs/1602.01783, 2016. URL http://arxiv.org/abs/1602.01783

    2016 arXiv

  20. [29]

    Graph feedback bandits on similar arms: With and without graph structures, 2025

    Qi, H., Guo, F., Zhu, L., Zhang, Q., and Li, X. Graph feedback bandits on similar arms: With and without graph structures, 2025. URL https://arxiv.org/abs/2501.14314

    2025

  21. [30]

    Language models are unsupervised multitask learners

    Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., and Sutskever, I. Language models are unsupervised multitask learners. OpenAI Blog, 1 0 (8): 0 9, 2019

    2019

  22. [31]

    Proximal policy optimization algorithms, 2017

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. Proximal policy optimization algorithms, 2017. URL https://arxiv.org/abs/1707.06347

    2017 arXiv

  23. [32]

    and Suggala, A

    Sharma, D. and Suggala, A. S. Offline-to-online hyperparameter transfer for stochastic bandits, 2025. URL https://arxiv.org/abs/2501.02926

    2025 arXiv

  24. [33]

    Large language model routing with benchmark datasets

    Shnitzer, T., Ou, A., Silva, M., Soule, K., Sun, Y., Solomon, J., Thompson, N., and Yurochkin, M. Large language model routing with benchmark datasets. arXiv preprint arXiv:2309.15789, 2023

    2023 arXiv

  25. [34]

    W., Chowdhery, A., Le, Q

    Suzgun, M., Scales, N., Sch \"a rli, N., Gehrmann, S., Tay, Y., Chung, H. W., Chowdhery, A., Le, Q. V., Chi, E. H., Zhou, D., et al. Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261, 2022

    arXiv

  26. [35]

    Medagents: Large language models as collaborators for zero-shot medical reasoning

    Tang, X., Zou, A., Zhang, Z., Li, Z., Zhao, Y., Zhang, X., Cohan, A., and Gerstein, M. Medagents: Large language models as collaborators for zero-shot medical reasoning. arXiv preprint arXiv:2311.10537, 2023

    2023 arXiv

  27. [36]

    Mba-rag: a bandit approach for adaptive retrieval-augmented generation through question complexity

    Tang, X., Gao, Q., Li, J., Du, N., Li, Q., and Xie, S. Mba-rag: a bandit approach for adaptive retrieval-augmented generation through question complexity. arXiv preprint arXiv:2412.01572, 2024

    2024 arXiv

  28. [37]

    G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ram \'e , A., et al

    Team, G., Riviere, M., Pathak, S., Sessa, P. G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ram \'e , A., et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024

    2024 arXiv

  29. [38]

    M., Sun, Y., Kundu, S., Xing, E., and Yurochkin, M

    Wang, H., Polo, F. M., Sun, Y., Kundu, S., Xing, E., and Yurochkin, M. Fusing models with complementary expertise. arXiv preprint arXiv:2310.01542, 2023

    2023 arXiv

  30. [39]

    Mixture-of-agents enhances large language model capabilities

    Wang, J., Wang, J., Athiwaratkun, B., Zhang, C., and Zou, J. Mixture-of-agents enhances large language model capabilities. arXiv preprint arXiv:2406.04692, 2024

    2024 arXiv

  31. [40]

    Concept-based interactive task learning

    Wang, S., Yu, M., Jiang, J., Zhang, W., Guo, X., Chang, S., Wang, Z., Klinger, T., Tesauro, G., and Campbell, M. Concept-based interactive task learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42 0 (9): 0 2290--2304, 2020

    2020

  32. [41]

    Wizardlm: Empowering large pre-trained language models to follow complex instructions

    Xu, C., Sun, Q., Zheng, K., Geng, X., Zhao, P., Feng, J., Tao, C., Lin, Q., and Jiang, D. Wizardlm: Empowering large pre-trained language models to follow complex instructions. In The Twelfth International Conference on Learning Representations, 2024

    2024

  33. [42]

    Flask: Fine-grained language model evaluation based on alignment skill sets

    Ye, S., Kim, D., Kim, S., Hwang, H., Kim, S., Jo, Y., Thorne, J., Kim, J., and Seo, M. Flask: Fine-grained language model evaluation based on alignment skill sets. arXiv preprint arXiv:2307.10928, 2023

    2023 arXiv

  34. [43]

    Zhang, R., Du, H., Niyato, D., Kang, J., Xiong, Z., Zhang, P., and Kim, D. I. Optimizing generative ai networking: A dual perspective with multi-agent systems and mixture of experts. arXiv preprint arXiv:2405.12472, 2024

    2024 arXiv

  35. [44]

    Judging llm-as-a-judge with mt-bench and chatbot arena

    Zheng, L., Chiang, W.-L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36: 0 46595--46623, 2023

    2023

  36. [45]

    Monte carlo tree search: a review of recent modifications and applications

    Świechowski, M., Godlewski, K., Sawicki, B., and Mańdziuk, J. Monte carlo tree search: a review of recent modifications and applications. Artificial Intelligence Review, 56 0 (3): 0 2497–2562, July 2022. ISSN 1573-7462. doi:10.1007/s10462-022-10228-y. URL http://dx.doi.org/10....

    2022 doi

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.