Pith. sign in

REVIEW 4 major objections 5 minor 34 references

The paper claims that a local LLM's agreement across a handful of prompt-varied responses is enough to decide when to offload to the cloud, making trained routers unnecessary.

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 12:39 UTC pith:YI3HWBQH

load-bearing objection New training-free routing idea worth taking seriously; evaluation protocol and Bayesian overclaim need fixing before the headline results are accepted. the 4 major comments →

arxiv 2607.20481 v1 pith:YI3HWBQH submitted 2026-05-30 cs.AI

Routing Without Training: Controllable-Ratio LLM Offloading via Reliability Gating

Evan Chen , Shiqiang Wang , Kevin S Chan , Su Wang , Christopher Brinton This is my paper
classification cs.AI
keywords training-free routingLLM offloadingresponse agreementBayesian early stoppinglocal-cloud collaborationmode mass estimationprompt variationself-consistency
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.

Using only the local model's own outputs, CARGO decides when to trust local inference and when to call a stronger cloud model. It samples the same query several times with semantically equivalent but stylistically different system prompts, measures how often the answers agree, and stops sampling once a Bayesian estimate of the agreement level is precise enough. Low agreement maps to a high offload probability, and a short warmup calibrates the router to hit any target offload ratio, with no training data, no router model, and no task-specific adaptation. On math, science, and reading benchmarks across several small local backbones, this training-free router consistently beats other training-free heuristics and in several settings outperforms a supervised learned router, supporting the claim that routing can emerge from the local model's intrinsic response behavior.

Core claim

The paper argues that the fraction of times a local model produces the same answer across several prompt-varied deterministic generations, called the agreement level, is a reliable and transferable signal for whether the local answer can be trusted. From this signal, CARGO builds a probabilistic router: generate responses with diverse reasoning-style prompts at temperature zero, track the empirical mode mass, stop early once a Beta-posterior credible interval is sufficiently narrow, then set the offload probability as a decreasing function of agreement. A warmup phase calibrates the routing threshold to any desired collaboration ratio. In experiments with Llama, Qwen, and Phi backbones on ma

What carries the argument

The central mechanism is the empirical mode mass, θ̂_k = max_a c_a/k, estimated from prompt-varied deterministic samples and treated as a Beta posterior. A Bayesian credible-interval stopping rule decides when enough samples have been drawn, and a sigmoid mapping σ(γ(λ−θ̂)) converts agreement into an offload probability, with the intercept λ calibrated by stochastic approximation during a warmup phase to hit a target collaboration ratio or token budget.

Load-bearing premise

Section 3's Bayesian early stopper assumes that, after conditioning on the empirical majority answer, the indicators for whether each sampled response matches that majority are independent Bernoulli draws with a fixed probability; because the majority is chosen from the same sample, the indicators are dependent and the Beta posterior is not the true posterior. If this assumption fails, the credible-interval rule is a heuristic and the contraction guarantee in Lemma 3.1 does n

What would settle it

Take a large set of queries with known correct answers, run CARGO, and record for each stopped query whether the reported 1−δ credible interval for the mode mass actually contains the true mode mass estimated from a very large sample. If the empirical coverage is substantially below 1−δ, the Bayesian guarantee is false. As a second check, replace the Bayesian early stopper with a fixed sample count matched to CARGO's average stopping round; if accuracy and routing behavior change little, the stopping rule is not the source of the reported gains.

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

If this is right

  • A router can be added at inference time to any local LLM without training data or task-specific supervision, making it usable with off-the-shelf and finetuned models alike.
  • Target collaboration ratios (e.g., 10%, 30%, 50%) become controllable at deployment time via a short warmup, so the same system adapts to changing budgets and latency constraints.
  • The Bayesian early stopper reduces the average number of local generations per query, lowering the compute and latency overhead of agreement-based routing.
  • The same agreement principle extends to token-budgeted offloading, allowing control over total cloud token consumption rather than just the fraction of queries offloaded.
  • Because the method operates only through prompts and outputs, it can be applied as an inference-time wrapper to existing LLM or agent-based applications without modifying the underlying models.

Where Pith is reading between the lines

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

  • Inference: The independence assumption behind the Beta posterior is violated in practice because the majority answer is selected from the same sample used to estimate the mode mass; the credible interval should be treated as a heuristic, and its actual coverage deserves empirical measurement rather than presumption.
  • Inference: The agreement signal could plausibly be combined with verifier-based or program-based self-consistency refinements to sharpen routing on math tasks, since those methods improve answer aggregation in the local-only setting.
  • Inference: The same calibration machinery could be adapted to other resource objectives beyond query count or token count, such as latency budgets, energy limits, or monetary cost, by changing the warmup objective function.
  • Inference: The paper's noted capability floor suggests a hybrid practical design: use agreement-based routing only when the local model passes a coarse capability screen on a small held-out set, reserving trained routing for very weak local models.

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 CARGO, a training-free router for local-cloud LLM offloading. CARGO estimates the local model's reliability by sampling responses under varied system prompts, uses a Beta posterior on the empirical agreement (mode mass) with a credible-interval early-stopping rule to limit per-query samples, and calibrates a logistic offloading parameter λ during a warmup batch to meet a target cloud collaboration ratio. The authors evaluate on math and QA benchmarks across several local LLMs (Qwen, Phi, Llama families) using DeepSeek-R1 as the cloud model, comparing against random offloading, self-confidence routing, CoT-step routing, and a supervised DeBERTa learned router. They report that CARGO consistently outperforms training-free baselines and in several settings outperforms the learned router at fixed offload ratios (e.g., ρ=0.3), and they include a token-budget variant, ablations on prompt-varied sampling, and a limitation study on very small local models.

Significance. If the claimed results hold, this is a meaningful contribution: it suggests that an intrinsic, training-free signal—prompt-varied response agreement—can support controllable local-cloud routing without a learned router or collaboration-aware fine-tuning. The paper is commendably thorough in scope: multiple model families and scales, several benchmarks, a token-budget extension, ablation of the sampling mechanism, and an explicit boundary condition for weak local models. The central idea is plausible and the experimental design is extensive. However, the evaluation protocol and the statistical justification of the early-stopping rule contain load-bearing gaps that must be addressed before the headline claims can be accepted.

major comments (4)
  1. [§4, Baselines paragraph] The sentence 'we randomly flip a small number of decisions when needed so that every method matches the same prescribed collaboration ratio exactly' is not innocuous and is not quantified anywhere. If a method's natural offload rate differs from ρ, the reported accuracy is the accuracy of a stochastic mixture of that policy and random routing. For a method with natural rate below ρ, the random flips convert high-agreement/easy queries from local to cloud, which can inflate measured accuracy; for a method with rate above ρ, the flips convert low-agreement/hard queries from cloud to local, which can depress accuracy. Since CARGO's λ is calibrated on a finite warmup batch, its deployment-time natural rate may deviate from ρ, and the paper does not report natural rates or flip counts for any method, including baselines. Without these numbers, the margins in Tables 1–2 and Figure 3 could refl
  2. [§3, 'Bayesian Estimation of Mode Mass', Eq. (2)] The Beta posterior in Eq. (2) is not the posterior of the mode mass θ = P(A=â_k | x) because it conditions on the empirical majority answer â_k, which is selected from the same sample used to estimate θ. The indicators Z_i = 1{A_i=â_k} are not unconditionally independent Bernoulli(θ) with a fixed θ; the selection of â_k introduces a data-dependent target and a multiplicity effect. Consequently, the credible interval [L_k, U_k] is not a valid (1−δ) posterior interval, and Lemma 3.1/Theorem 3.2 do not license the claimed uncertainty guarantee for the stopping rule. The routing procedure may still work as a heuristic, and the empirical agreement estimator is justified by the SLLN, but the paper should either weaken the statistical claims or provide a selection-adjusted uncertainty quantification (e.g., bootstrap or post-selection inference).
  3. [§4, 'Experimental Setup' and Figure 5] The headline comparison is accuracy at a fixed offload ratio, but it ignores local compute and latency. CARGO performs up to K_max local generations per query (with early stopping), while all baselines use a single local generation (plus possibly a confidence token or CoT step count). Figure 5 reports mean stopping round but does not translate it into end-to-end latency or local FLOPs, and Table 1 gives no accuracy-versus-local-cost trade-off. CARGO's advantage could partly be the result of spending more local compute per query. Please report mean samples per query for all methods and provide accuracy as a function of local inference cost or latency, not only as a function of offload ratio.
  4. [Algorithm 1, lines 2–6 and Section 3 'Warmup Calibration'] The warmup batch B is described only as 'a fixed batch of B warmup queries'; the source (training, validation, random deployment queries) and the values of B, T0, and η_λ are not reported. This matters because the calibrated λ is fixed for deployment, and the paper does not report the realized offload ratio before the post-hoc flips. If B is small or unrepresentative, the target-ratio guarantee is not established. Please specify the warmup protocol, report the realized natural offload rates, and show the sensitivity of the final accuracy to B and T0.
minor comments (5)
  1. [§1, Contributions bullet] Typo: 'greater correctness and consistency than than self-reported confidence' should read 'than self-reported confidence'.
  2. [Appendix A, Proof of Lemma 3.1] The proof argues asymptotically with O(k) notation, but the claimed bound Var(θ|A_k) ≤ 1/[4(k+α0+β0+1)] follows directly from ab ≤ (a+b)^2/4. The current asymptotic argument does not establish the stated constant; please replace it with the direct inequality.
  3. [Theorem 3.2] The proof assumes a unique maximizer 'a⋆' with a positive gap Δ to all other answers, but this is not stated in the theorem. If there are ties for the mode, the empirical majority answer may not converge to a single answer, and the claim that 'with probability one there exists a finite K...' needs a different argument or an added uniqueness assumption.
  4. [Algorithm 1, inputs and lines 11–12] The algorithm input lists a temperature T, but prompt-varied sampling is described as deterministic decoding with T=0. Please state this explicitly in the algorithm or remove T from the input list.
  5. [§4.1, 'Performance Across Collaboration Ratios'] The phrase 'Randomvaries from query offloading' is missing a space and a period; it should read 'while Random varies from query offloading.' Also, the three-trial variance band is reported only for CARGO; please add variance or standard errors for baselines in Figure 3.

Circularity Check

0 steps flagged

No significant circularity: the central accuracy claims rest on external benchmark evaluations, not on reductions of fitted inputs or self-citations.

full rationale

CARGO's derivation chain is self-contained. The only fitted quantity is the routing intercept λ, updated by Eq. (3) to match a target offload ratio on a warmup batch; the reported accuracies on held-out benchmarks (Tables 1–2, Figure 3) are not derived from that fit, so the central comparison does not reduce to its inputs. The Bayesian early-stopping analysis (Lemma 3.1, Theorem 3.2) is a heuristic for estimating agreement, and even if the conditional-independence assumption is imperfect after conditioning on the empirical majority answer, that is a statistical-validity caveat, not a case where a prediction is defined as its input. The self-citations ([7], [12]) are used as baselines or related work and are not load-bearing for the paper's core claim. The paper's Appendix C.2 explicitly limits the method to local models with nontrivial reasoning ability, and Section 4's random-flip ratio matching is a legitimate experimental-design concern because reported accuracy can be a stochastic mixture with random routing, but that is a comparison-validity issue rather than circular derivation. No enumerated circularity pattern is present.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The paper's central mechanism is an inference-time signal (agreement) plus a Bayesian early-stopping wrapper. The load-bearing statistical assumption is that the Beta-Binomial posterior is valid for the mode count computed after selecting the majority answer; this is not derived. The hand-designed prompt set is a de facto feature extractor, which is a free design choice. Beyond that, the empirical evaluation rests on standard benchmarks and baselines.

free parameters (7)
  • Beta prior shapes α0, β0
    Hand-chosen pseudo-counts; control posterior concentration and early stopping.
  • Credible interval width threshold ε
    Stopping threshold; smaller ε increases sample count and latency.
  • Logistic slope γ
    Controls sharpness of sigmoid offload probability.
  • Maximum samples K_max
    Sampling budget cap.
  • Warmup batch size B, iterations T0, step size ηλ
    Calibration hyperparameters for stochastic approximation.
  • Hand-designed prompt set S = 9-10 prompt variants
    Semantically equivalent system prompts; acts as a manually engineered feature extractor for the agreement signal.
  • Token weights α, β (token-budget variant)
    Weights for input/output token costs in Algorithm 2.
axioms (5)
  • ad hoc to paper Conditioned on empirical majority answer â_k, indicators Z_i are independent Bernoulli(θ)
    Underlies Beta posterior in Eq. (2); selection of â_k from the data breaks independence and makes θ data-dependent.
  • domain assumption Decoding samples are conditionally independent given query under prompt sampling
    Used in Lemma 3.1/Theorem 3.2; plausible if prompts are sampled i.i.d. and decoding is deterministic.
  • domain assumption The true mode a⋆ is unique (or eventually identified)
    Theorem 3.2 proof requires a unique maximizer; ties are not treated.
  • standard math Stochastic approximation for λ converges to target ratio
    Robbins-Monro style update; assumes stable smooth mapping from λ to offload rate.
  • ad hoc to paper Beta prior is an adequate model for mode mass
    Chosen for conjugacy; no empirical justification provided.

pith-pipeline@v1.3.0-alltime-deepseek · 16424 in / 12542 out tokens · 114604 ms · 2026-08-02T12:39:18.596223+00:00 · methodology

0 comments
read the original abstract

Local-cloud collaboration is a practical way to deploy large language models under resource constraints, but existing methods often rely on trained routers or collaboration-aware finetuning that tie routing behavior to a particular operating regime. In this work, we show that such training may be unnecessary: the local model's own inference-time agreement across sampled responses already provides a strong signal for deciding when to trust local execution and when to offload to a stronger cloud model. We propose CARGO, a training-free routing framework that estimates this agreement through prompt-varied sampling, applies Bayesian early stopping for sample-efficient uncertainty control, and supports arbitrary target collaboration ratios through lightweight deployment-time calibration. Across diverse reasoning and question-answering tasks, multiple local LLM families and scales, and both pretrained and finetuned local models, CARGO consistently outperforms other training-free baselines and in several settings surpasses supervised learned routers. These results suggest that effective and adaptable local-cloud collaboration can emerge directly from the local model's intrinsic response behavior, without requiring an additional trained router.

Figures

Figures reproduced from arXiv: 2607.20481 by Christopher Brinton, Evan Chen, Kevin S Chan, Shiqiang Wang, Su Wang.

Figure 1
Figure 1. Figure 1: Overview of CARGO. The local model estimates agreement under prompt-varied sampling, then applies [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of inference-time routing sig [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Evaluation across multiple target collaboration ratios [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Prompt-varied sampling on MATH-lighteval. For all nonzero temperatures, sampling is performed with a [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Early-exit evaluation of CARGO across datasets and local models. The bars report the mean stopping round and red markers indicate the maximum sampled response in the dataset. The results show that the proposed agreement-estimation procedure usually stops far before the sampling budget is exhausted, while adapting its stopping depth to dataset difficulty. Performance Across Collaboration Ratios [PITH_FULL_… view at source ↗
Figure 6
Figure 6. Figure 6: Token-budgeted offloading results. CARGO can also control cloud-token usage in￾stead of offloading ratio. Extending to Token-Budgeted Offloading. CARGO can support resource-aware control by regulating cloud-token usage rather than only the offloading ratio. We define c(x) = r(x) [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance on 1B-sized LLMs [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance after RLHF-finetuning local LLMs on specific tasks. It suggests that the main role of finetuning [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Offloading evaluation under token-budgeted routing. The x-axis shows the total allowed cloud-token budget. [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

34 extracted references · 19 linked inside Pith

  1. [1]

    Llm cascade with multi-objective optimal consideration,

    K. Zhang, L. Peng, C. Wang, A. Go, and X. Liu, “Llm cascade with multi-objective optimal consideration,” 2024

    2024

  2. [2]

    Llava-phi: Efficient multi-modal assistant with small language model,

    Y. Zhu, M. Zhu, N. Liu, Z. Xu, and Y. Peng, “Llava-phi: Efficient multi-modal assistant with small language model,” inProceedings of the 1st International Workshop on Efficient Multimedia Computing under Limited, 2024, pp. 18–22

    2024

  3. [3]

    Tinyllama: An open-source small language model,

    P. Zhang, G. Zeng, T. Wang, and W. Lu, “Tinyllama: An open-source small language model,”arXiv preprint arXiv:2401.02385, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Mobilellm: Optimizing sub-billion parameter language models for on-device use cases,

    Z. Liu, C. Zhao, F. Iandola, C. Lai, Y. Tian, I. Fedorov, Y. Xiong, E. Chang, Y. Shi, R. Krishnamoorthi et al., “Mobilellm: Optimizing sub-billion parameter language models for on-device use cases,” in Forty-first International Conference on Machine Learning, 2024

    2024

  5. [5]

    On-device language models: A comprehensive review,

    J. Xu, Z. Li, W. Chen, Q. Wang, X. Gao, Q. Cai, and Z. Ling, “On-device language models: A comprehensive review,”arXiv preprint arXiv:2409.00088, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Edgellm: Fast on-device llm inference with speculative decoding,

    D. Xu, W. Yin, H. Zhang, X. Jin, Y. Zhang, S. Wei, M. Xu, and X. Liu, “Edgellm: Fast on-device llm inference with speculative decoding,”IEEE Transactions on Mobile Computing, 2024

    2024

  7. [7]

    Bridging on-device and cloud llms for collaborative reasoning: A unified methodology for local routing and post-training

    W. Fang, D.-J. Han, L. Yuan, E. Chen, and C. G. Brinton, “Bridging on-device and cloud llms for collaborative reasoning: A unified methodology for local routing and post-training.”

  8. [8]

    Hybrid llm: Cost-efficient and quality-aware query routing,

    D. Ding, A. Mallick, C. Wang, R. Sim, S. Mukherjee, V. Ruhle, L. V. Lakshmanan, and A. H. Awadallah, “Hybrid llm: Cost-efficient and quality-aware query routing,”arXiv preprint arXiv:2404.14618, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Routellm: Learning to route llms with preference data,

    I. Ong, A. Almahairi, V. Wu, W.-L. Chiang, T. Wu, J. E. Gonzalez, M. W. Kadous, and I. Stoica, “Routellm: Learning to route llms with preference data,”arXiv preprint arXiv:2406.18665, 2024

    Pith/arXiv arXiv 2024

  10. [10]

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

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

    Pith/arXiv arXiv 2023

  11. [11]

    Repic: Reinforced post-training for personalizing multi-modal language models,

    Y. Oh, D. Chung, J. Shin, S. Park, J. Barthelemy, J. Mok, and S. Yoon, “Repic: Reinforced post-training for personalizing multi-modal language models,” 2025, to appear; arXiv:2506.18369

    arXiv 2025

  12. [12]

    Joint continual learning of local language models and cloud offloading decisions with budget constraints,

    E. Chen, W. Fang, S. Wang, and C. Brinton, “Joint continual learning of local language models and cloud offloading decisions with budget constraints,”arXiv preprint arXiv:2602.00166, 2026

    arXiv 2026

  13. [13]

    Universal self-consistency for large language model generation,

    X. Chen, R. Aksitov, U. Alon, J. Ren, K. Xiao, P. Yin, S. Prakash, C. Sutton, X. Wang, and D. Zhou, “Universal self-consistency for large language model generation,”arXiv preprint arXiv:2311.17311, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    Self-consistency improves chain of thought reasoning in language models,

    X. Wang, J. Wei, D. Schuurmans, Q. Le, E. Chi, S. Narang, A. Chowdhery, and D. Zhou, “Self-consistency improves chain of thought reasoning in language models,”arXiv preprint arXiv:2203.11171, 2022

    Pith/arXiv arXiv 2022

  15. [15]

    Confidence improves self-consistency in llms,

    A. Taubenfeld, T. Sheffer, E. Ofek, A. Feder, A. Goldstein, Z. Gekhman, and G. Yona, “Confidence improves self-consistency in llms,” inFindings of the Association for Computational Linguistics: ACL 2025, 2025, pp. 20090–20111

    2025

  16. [16]

    Not all votes count! programs as verifiers improve self-consistency of language models for math reasoning,

    V. Y. Toh, D. Ghosal, and S. Poria, “Not all votes count! programs as verifiers improve self-consistency of language models for math reasoning,”arXiv preprint arXiv:2410.12608, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms,

    M. Xiong, Z. Hu, X. Lu, Y. Li, J. Fu, J. He, and B. Hooi, “Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms,”arXiv preprint arXiv:2306.13063, 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Fact-and-reflection (far) improves confidence calibration of large language models,

    X. Zhao, H. Zhang, X. Pan, W. Yao, D. Yu, T. Wu, and J. Chen, “Fact-and-reflection (far) improves confidence calibration of large language models,” inFindings of the Association for Computational Linguistics: ACL 2024, 2024, pp. 8702–8718. 11

    2024

  19. [19]

    When to trust llms: Aligning confidence with response quality,

    S. Tao, L. Yao, H. Ding, Y. Xie, Q. Cao, F. Sun, J. Gao, H. Shen, and B. Ding, “When to trust llms: Aligning confidence with response quality,” inFindings of the Association for Computational Linguistics: ACL 2024, 2024, pp. 5984–5996

    2024

  20. [20]

    Measuring mathematical problem solving with the math dataset,

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

    Pith/arXiv arXiv 2021

  21. [21]

    Are nlp models really able to solve simple math word problems?

    A. Patel, S. Bhattamishra, and N. Goyal, “Are nlp models really able to solve simple math word problems?” inProceedings of the 2021 conference of the North American chapter of the association for computational linguistics: human language technologies, 2021, pp. 2080–2094

    2021

  22. [22]

    Training verifiers to solve math word problems,

    K. Cobbe, V. Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakanoet al., “Training verifiers to solve math word problems,”arXiv preprint arXiv:2110.14168, 2021

    Pith/arXiv arXiv 2021

  23. [23]

    Let’s verify step by step,

    H. Lightman, V. Kosaraju, Y. Burda, H. Edwards, B. Baker, T. Lee, J. Leike, J. Schulman, I. Sutskever, and K. Cobbe, “Let’s verify step by step,”arXiv preprint arXiv:2305.20050, 2023

    Pith/arXiv arXiv 2023

  24. [24]

    Agieval: A human-centric benchmark for evaluating foundation models,

    W. Zhong, R. Cui, Y. Guo, Y. Liang, S. Lu, Y. Wang, A. Saied, W. Chen, and N. Duan, “Agieval: A human-centric benchmark for evaluating foundation models,” inFindings of the association for computational linguistics: NAACL 2024, 2024, pp. 2299–2314

    2024

  25. [25]

    A diverse corpus for evaluating and developing english math word problem solvers,

    S.-Y. Miao, C.-C. Liang, and K.-Y. Su, “A diverse corpus for evaluating and developing english math word problem solvers,” inProceedings of the 58th annual meeting of the Association for Computational Linguistics, 2020, pp. 975–984

    2020

  26. [26]

    Think you have solved question answering? try arc, the ai2 reasoning challenge,

    P. Clark, I. Cowhey, O. Etzioni, T. Khot, A. Sabharwal, C. Schoenick, and O. Tafjord, “Think you have solved question answering? try arc, the ai2 reasoning challenge,”arXiv preprint arXiv:1803.05457, 2018

    Pith/arXiv arXiv 2018

  27. [27]

    Measuring massive multitask language understanding,

    D. Hendrycks, C. Burns, S. Basart, A. Zou, M. Mazeika, D. Song, and J. Steinhardt, “Measuring massive multitask language understanding,”arXiv preprint arXiv:2009.03300, 2020

    Pith/arXiv arXiv 2009

  28. [28]

    Solving quantitative reasoning problems with language models,

    A. Lewkowycz, A. Andreassen, D. Dohan, E. Dyer, H. Michalewski, V. Ramasesh, A. Slone, C. Anil, I. Schlag, T. Gutman-Soloet al., “Solving quantitative reasoning problems with language models,” Advances in neural information processing systems, vol. 35, pp. 3843–3857, 2022

    2022

  29. [29]

    Squad: 100,000+ questions for machine comprehension of text,

    P. Rajpurkar, J. Zhang, K. Lopyrev, and P. Liang, “Squad: 100,000+ questions for machine comprehension of text,” inProceedings of the 2016 conference on empirical methods in natural language processing, 2016, pp. 2383–2392

    2016

  30. [30]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azharet al., “Llama: Open and efficient foundation language models,”arXiv preprint arXiv:2302.13971, 2023

    Pith/arXiv arXiv 2023

  31. [31]

    Qwen technical report,

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

    Pith/arXiv arXiv 2023

  32. [32]

    Phi-3 technical report: A highly capable language model locally on your phone,

    M. Abdin, J. Aneja, H. Awadalla, A. Awadalla, A. A. Awan, N. Bach, A. Bahree, A. Bakhtiari, H. Behl, A. Benhaimet al., “Phi-3 technical report: A highly capable language model locally on your phone,” arXiv preprint arXiv:2404.14219, 2024

    Pith/arXiv arXiv 2024

  33. [33]

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

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

    Pith/arXiv arXiv 2025

  34. [34]

    Deberta: Decoding-enhanced bert with disentangled attention,

    P. He, X. Liu, J. Gao, and W. Chen, “Deberta: Decoding-enhanced bert with disentangled attention,” arXiv preprint arXiv:2006.03654, 2020. 12 Appendix A Proof for Lemma 3.1 14 B Proof for Theorem 3.2 14 C Additional Experiments 15 C.1 Fixed Collaboration Ratio on more datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 C.2 Performance of Tr...

    Pith/arXiv arXiv 2006