Pith. sign in

REVIEW 3 major objections 5 minor 23 references

Edge LLM agents can stop thinking early and call the cloud only when still uncertain, with joint statistical guarantees on reward and cloud use.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 18:48 UTC pith:YP7EHI4T

load-bearing objection Solid systems paper: joint TC+deferral with multi-objective LTT actually works on ReAct agents, but the headline token cuts are conditional on certification succeeding. the 3 major comments →

arxiv 2607.26865 v1 pith:YP7EHI4T submitted 2026-07-29 stat.ML cs.AIcs.ITcs.LGmath.IT

Think Short, Defer Smart, Act, and Repeat: Calibrated Reasoning and Uncertainty-Aware Deferral for Edge LLM Agents

Amirmohammad Farzaneh , Osvaldo Simeone This is my paper
classification stat.ML cs.AIcs.ITcs.LGmath.IT
keywords edge LLM agentsReActthought calibrationuncertainty-aware deferralLearn-Then-Testperplexityconvergence probecloud cascade
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.

ReAct-style LLM agents burn most of their compute on long thought traces, yet when run on the edge they still need a way to hand hard steps to a larger cloud model without blowing the budget. This paper argues that those two decisions—when to stop thinking and when to defer—should be calibrated together, because long reasoning can make a weak local model look overconfident and suppress needed cloud calls. TSDS trains a light probe on the model’s hidden states to halt generation once the intended action has stabilized, then uses action perplexity to decide whether to escalate. Both thresholds are chosen by multi-objective Learn-Then-Test on full episode trajectories so that, with high probability over calibration data, expected reward stays above a floor and expected cloud-call rate stays under a cap, while thinking tokens fall sharply versus deferral-only baselines. A sympathetic reader cares because the method turns a resource-reliability trade-off into something you can certify before deployment on multi-step edge agents.

Core claim

Jointly calibrating a thought-convergence threshold and a perplexity deferral threshold with multi-objective Learn-Then-Test yields simultaneous finite-sample guarantees on expected episode reward and cloud-call rate, while cutting per-episode thinking compute by roughly 43%–73% versus calibrated deferral-only baselines on multi-hop QA, code generation, and household robot planning.

What carries the argument

Think Short, Defer Smart (TSDS): a convergence probe that stops on-device reasoning once the intended action stabilizes, paired with a perplexity deferral rule, with both thresholds certified together by multi-objective Learn-Then-Test on end-to-end episode trajectories.

Load-bearing premise

Calibration and deployment episodes must be independent draws from the same distribution, so the certified reward floor and cloud budget still hold in the wild.

What would settle it

On a held-out stream of the same ReAct tasks, check whether the LTT-selected thresholds keep average reward at or above R_min and average cloud-call rate at or below C_max_D while still beating a full-thought calibrated-deferral baseline on thinking tokens; if either constraint fails more often than the claimed 1−δ rate, or the token savings disappear, the central claim fails.

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

If this is right

  • Edge ReAct agents can cut thinking tokens by about half to three-quarters versus full-trace calibrated deferral without abandoning certified reward and cloud budgets.
  • Thought truncation and cloud deferral reinforce each other: stopping early preserves uncertainty signals that full traces tend to wash out.
  • On multi-step tasks, thought calibration alone is often not enough; certified deferral is needed to recover reward.
  • The same joint LTT recipe can pick operating points that also respect episode-length cost when several certified pairs remain.

Where Pith is reading between the lines

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

  • If task mix or tools drift after calibration, the method would need online recalibration or distribution-shift-aware risk control to keep the same certificates.
  • Richer uncertainty scores (for example semantic entropy) could shrink cloud use further under the same LTT shell without changing the stopping probe.
  • Multi-tier cascades—small edge, mid-tier, large cloud—look like a direct extension of the same multi-constraint certification idea.

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

3 major / 5 minor

Summary. The paper proposes TSDS, a ReAct-style edge LLM agent that couples (i) a lightweight convergence probe that early-stops on-device thought generation once the intended action has stabilized and (ii) a perplexity-based rule that defers uncertain actions to a cloud model. The two thresholds (λ_L, λ_D) are selected jointly by multi-objective Learn-Then-Test on end-to-end episode trajectories, yielding a finite-sample guarantee (Prop. 1) that expected reward stays above R_min and expected cloud-call rate stays below C_max_D with probability at least 1−δ whenever a certified pair is returned. Experiments on GSM8K, HotpotQA, MBPP, and a new text household-robot benchmark compare against full-thought edge, always-cloud, thought-calibration-only, and calibrated-deferral-only baselines, reporting large reductions in thinking tokens relative to deferral-only policies while claiming to preserve the certified constraints.

Significance. If the joint claim holds as stated, the work is a useful systems contribution at the intersection of test-time compute control and cascading for multi-step LLM agents: it shows that thought truncation and deferral are synergistic rather than independent knobs, supplies distribution-free simultaneous control of reward and cloud cost via standard LTT/FWER machinery, and releases code plus a reusable household-robot environment. The empirical interplay evidence (Appendix H / Fig. 13) that continued reasoning can suppress PPL and block needed deferrals is a concrete, falsifiable insight for edge deployment. Strengths that should be credited include open implementations, 50-split reporting with error bars, component-isolating baselines, and a correctly invoked multi-objective LTT procedure rather than heuristic threshold picking.

major comments (3)
  1. [Abstract; Prop. 1; §5.2; App. G.4] Abstract and Prop. 1 jointly advertise certified reward, certified cloud-call rate, and large token reductions. LTT only guarantees the constraints conditional on returning a nonempty certified set Λ̂. Appendix G.4 states that on MBPP TSDS certifies in only 22/50 random 60/40 splits, and Fig. 2 / Table 3 averages (including the −64% thinking-cost claim) are taken over that selection event or mix certified and non-certified regimes. The main text never reports certification frequency, never bounds P(Λ̂ ≠ ∅), and never specifies a fallback when certification fails. Without that, the headline “certified + large token cut” claim is not distribution-free over the same measure that generates deployment episodes. Please report certification rates for every benchmark in the main results, define an explicit fallback (e.g., ReDAct-CD or E-ReAct), and either soften the abstract or add a procedure/b
  2. [Prop. 1; §4.3; §6] Prop. 1 and §4.3 assume calibration and test episodes are i.i.d. from the deployment distribution. For edge agents this is strong: prompts, tools, and environments are typically nonstationary. The paper should state this limitation explicitly near the guarantee, and ideally include at least one stress check (e.g., calibrate on one HotpotQA/household task family and test on another, or shift tool APIs) so readers can gauge how brittle the certified (R_min, C_max_D) pair is under mild distribution shift.
  3. [§4.1, Eq. (5); App. G.1] The probe supervision in Eq. (5) labels a position positive only if every later early-stopped action agrees with the full-thought action. That is a reasonable offline proxy, but the online rule is irrevocable and the full-thought action itself can be wrong. The manuscript does not quantify how often early-stopped actions that match the full-thought action still fail the environment, nor how often the probe stops before a later flip that would have changed the action. A short diagnostic (flip rate after the certified stop, agreement with oracle/cloud action) on D_cal would make the central “convergence = safe to stop” modeling assumption checkable rather than axiomatic.
minor comments (5)
  1. [Fig. 1] Figure numbering and panel labels in the main text occasionally refer to “thought convergence threshold λ_T / λ_L” inconsistently (Fig. 1 uses λ_T and λ_L in different panels). Standardize on λ_L everywhere.
  2. [§5.2–5.4; App. G.2] R_min values (0.69 MBPP, 0.22 HotpotQA, 0.52 household, 0.855 GSM8K) appear chosen near edge full-thought performance; a one-sentence justification per benchmark would help readers interpret how tight the floors are.
  3. [§2] Related work cites ReDAct and Thought Calibration appropriately; a short explicit comparison table (what each method calibrates, single- vs multi-step, guarantee type) would clarify the incremental contribution of joint multi-objective LTT.
  4. [Title; throughout] Typographical inconsistencies: “ThinkShort, DeferSmart” spacing in the title block, and occasional missing spaces in math (“λL” vs “λ_L”). Clean for camera-ready.
  5. [§5.3; App. G.3] Appendix G.3 Pareto discussion is useful but under-referenced from §5.3; a pointer in the main HotpotQA subsection would help.

Circularity Check

0 steps flagged

No significant circularity: LTT certifies external reward/deferral metrics; probe labels and thresholds are not the test quantities by construction.

full rationale

The load-bearing derivation is multi-objective Learn-Then-Test selection of (λ_L, λ_D) so that expected episode reward R(λ) and cloud-call rate C_D(λ) meet user floors/budgets with finite-sample FWER control (Prop. 1; Sec. 4.3; App. C–F). Those risks are environment/task metrics and observed deferral indicators, not algebraic rearrangements of the probe loss or of the fitted thresholds. Probe targets (Eq. 5) are standard offline stable-run labels from full-thought rollouts; stopping and PPL deferral are then applied online and scored on held-out splits—supervised probing plus post-hoc threshold testing, not a tautology. Prop. 1 and its proof reduce to external LTT/FWER machinery (Angelopoulos et al., 2021) with Bonferroni; the authors’ related LTT citation is framing, not a uniqueness theorem that forces the result. Token-reduction claims are empirical comparisons to baselines on GSM8K/HotpotQA/MBPP/household, not fitted inputs renamed as predictions. Conditional certification rates (e.g., MBPP 22/50 splits) affect how strongly the headline should be read, but that is a coverage/selection issue, not circularity of the derivation chain.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The reliability claim rests on classical multiple-testing risk control plus modeling choices that hidden states reveal action commitment and that action perplexity is a usable deferral score. Threshold grids, probe layer, and constraint levels are free design choices fixed before test evaluation but still shape which operating points exist. No new physical entities are postulated; the main invented artifacts are the TSDS pipeline and a text household benchmark.

free parameters (4)
  • Thought threshold grid Λ_L and deferral threshold grid Λ_D = e.g. λ_L∈{0.1..0.9}; λ_D from PPL percentiles on D_tr (task-specific)
    Discrete candidate pairs searched by LTT; spacing and quantile construction determine the certified set size and selected cost.
  • Probe layer ℓ and EMA α = ℓ=5 (most tasks), ℓ=10 (household); α=0.5
    Architecture/hyperparameters of φ_θ chosen by ablation (ℓ=5 or 10; α=0.5) and fixed before LTT.
  • R_min, C_max_D, δ = δ=0.10; C_max_D=0.70; R_min task-specific (e.g. 0.69 MBPP, 0.22 HotpotQA)
    User constraint levels that define which hypotheses LTT may accept; directly gate reported success.
  • L_max thought caps and probe stride/window = task-specific (e.g. 2048 MBPP edge; stride 16; W=16)
    Hard compute limits and labeling stride shape both training labels and achievable C_L.
axioms (5)
  • domain assumption Calibration episodes are i.i.d. from the same distribution as deployment episodes, so LTT FWER control yields Prop. 1 guarantees.
    Stated in Prop. 1; standard for distribution-free calibration but strong for live agents.
  • standard math One-sided binomial (or analogous) p-values with Bonferroni control the family-wise error over the finite grid Λ.
    Appendix C / Angelopoulos et al. 2021; underpins joint reward and deferral certification.
  • ad hoc to paper Agreement of early-stopped actions with the full-thought action at all later probe positions (Eq. 5) labels true reasoning convergence for irrevocable stopping.
    Defines supervised targets for φ_θ; ties probe training to the stopping semantics used online.
  • domain assumption Action perplexity (or listed alternatives) is a monotone-enough uncertainty score for selective cloud escalation.
    Sec. 4.2 and App. B; inherited from ReDAct-style deferral literature.
  • domain assumption Terminal scalar reward r∈[0,1] and cloud-call count are the right multi-objective risks to control for edge deployment.
    Problem (4); episode length is only a tie-break, not a hard constraint.
invented entities (2)
  • TSDS joint pipeline (convergence probe + PPL deferral + multi-objective LTT) independent evidence
    purpose: Couple early exit and cloud escalation under simultaneous finite-sample constraints.
    Composition of prior modules into one certified agent loop; primary claimed contribution.
  • Household-robot text benchmark (four rooms, fetch/place/clean/heat) independent evidence
    purpose: Provide a multi-step embodied-style ReAct testbed without vision/hardware.
    New environment released with the paper; useful but simplified relative to ALFWorld/SayCan.

pith-pipeline@v1.2.0-daily-grok45 · 24040 in / 3768 out tokens · 83049 ms · 2026-07-30T18:48:35.124336+00:00 · methodology

0 comments
read the original abstract

LLM agents following the ReAct paradigm are promising enablers of complex multi-step tasks, including multi-hop question answering, code generation, and control of physical AI systems. Yet, when deployed at the edge, they must tightly manage their reasoning budget while remaining reliable and deferring to a cloud-side model only when local uncertainty is too high to act safely. We propose Think Short, Defer Smart (TSDS), a framework that synergistically integrates a lightweight convergence probe, which halts on-device reasoning once the intended action has stabilized, with a perplexity-based deferral rule that escalates uncertain actions to a cloud-side model. Both mechanisms are jointly calibrated on end-to-end episode trajectories via a multi-objective Learn-Then-Test (LTT) procedure, providing simultaneous finite-sample guarantees on expected episode reward and cloud-call rate. We evaluate TSDS on four ReAct benchmarks spanning arithmetic reasoning (GSM8K), multi-hop question answering (HotpotQA), code generation (MBPP), and multi-step embodied planning (household robot), and compare against thought-calibration-only and calibrated-deferral-only standalone baselines. TSDS reduces per-episode thinking compute by 43%-73% over deferral-only baselines across HotpotQA, MBPP, and the household robot task, while maintaining certified reward and cloud-call rate guarantees.

Figures

Figures reproduced from arXiv: 2607.26865 by Amirmohammad Farzaneh, Osvaldo Simeone.

Figure 1
Figure 1. Figure 1: TSDS pipeline as a high-level illustration ( [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Test results for MBPP, shown for five policies and averaged over [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Reward vs. thinking cost on HotpotQA across five λL grid values. The blue dashed curve shows thought calibration only, and the solid green curve shows TSDS at the certified deferral threshold λD = 0.318. All values are means over the full pool of 250 validation episodes. Setup: The cloud model is DeepSeek-R1-Distill-Qwen-14B. Edge and cloud thought caps are set to Lmax = 384 and 1024, respectively. The pro… view at source ↗
Figure 4
Figure 4. Figure 4: Test results for HotpotQA, shown for all five policies and averaged over [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Test results for the household robot task, shown for five policies and averaged over [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Illustration of the thought calibration stopping rule. The probe signal [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Test AUROC of φθ(h (ℓ) t,p) vs. probing layer ℓ on GSM8K under the stable-run label (14). Solid lines: seed-averaged test AUROC; shaded bands: ±1 standard deviation. Dashed line: chance (0.5). The EMA and rolling-mean attention probes dominate at ℓ ≈ 5. Q4 recovery to 0.812 for EMA) reflects class-balance shifts: Q1 has a low positive rate (stability requires all subsequent probe positions to agree, rare f… view at source ↗
Figure 8
Figure 8. Figure 8: Per-quartile AUROC at the best layer ℓ ⋆ for each probe architecture on GSM8K under the stable-run similarity score (14). The EMA probe leads on Q1, Q2, and Q3, while the rolling-mean attention probe attains the highest late-thought AUROC (AUROCQ4 = 0.874). The dashed grey line marks chance performance (AUROC = 0.5). samples having higher positive rates. Only 22% of problems are stable throughout their tho… view at source ↗
Figure 9
Figure 9. Figure 9: GSM8K test-set results (50-seed averages ± std, ntest=195 per split, δ = 0.10). Across 50 independent 85/15 splits of 1300 held-out problems, each method’s operating point λˆ is selected by LTT on the calibration portion; all metrics are measured on the held-out test portion. Left: test-set reward; dashed line marks Rmin = 0.855. A bar above this line reflects empirical test performance, not the LTT guaran… view at source ↗
Figure 10
Figure 10. Figure 10: HotpotQA results under two constraint regimes, averaged over [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Pareto frontiers in (CˆL, Cˆ S) space on HotpotQA for four deferral budgets C max D (Rmin ≈ 0.10 throughout), averaged over 200 random 75/25 calibration/test splits. For each split and each certified λL, we record (CˆL, minλD∈Λ( ˆ λL) Cˆ S); plotted points are means across splits; shaded bands are ±1 std of Cˆ S. Tightening C max D shifts the frontier upward and rightward: fewer allowed deferrals force mo… view at source ↗
Figure 12
Figure 12. Figure 12: Floor plan of the simulated household environment. Each room contains a set of surfaces [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Thought calibration–deferral interplay on the [PITH_FULL_IMAGE:figures/full_fig_p021_13.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

23 extracted references · 16 linked inside Pith

  1. [1]

    Do as I can, not as I say: Grounding language in robotic affordances.arXiv preprint arXiv:2204.01691,

    Michael Ahn, Anthony Brohan, Noah Brown, et al. Do as I can, not as I say: Grounding language in robotic affordances.arXiv preprint arXiv:2204.01691,

    Pith/arXiv arXiv

  2. [3]

    can be any information-theoretic quantity derived from the token-level probabilities of the generated action sequencey= (τ 1, . . . , τL)given contextx, with joint probability p(y|x, θ) = LY i=1 p(τi |x, τ<i, θ).(8) All three measures below are available as free by-products of the autoregressive decoding pass, requiring no additional forward pass (Malinin...

    2021

  3. [10]

    Cited via Kram ´ar et al. (2026). J´anos Kram ´ar, Joseph Engels, Senthooran Wang, Bilal Chughtai, Rohin Shah, Neel Nanda, and Arthur Conmy. Building production-ready probes for Gemini.arXiv preprint arXiv:2601.11516,

    arXiv 2026

  4. [11]

    Uncertainty estimation in autoregressive structured prediction

    Andrey Malinin and Mark Gales. Uncertainty estimation in autoregressive structured prediction. arXiv preprint arXiv:2002.07650,

    Pith/arXiv arXiv 2002

  5. [12]

    Gonzalez, M

    Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M. Waleed Kadous, and Ion Stoica. RouteLLM: Learning to route LLMs with preference data. arXiv preprint arXiv:2406.18665,

    Pith/arXiv arXiv

  6. [13]

    ReDAct: Reasoning-driven action selection under uncertainty for language model agents.arXiv preprint arXiv:2604.07036,

    10 Aliaksei Piatrashyn et al. ReDAct: Reasoning-driven action selection under uncertainty for language model agents.arXiv preprint arXiv:2604.07036,

    Pith/arXiv arXiv

  7. [14]

    V oyager: An open-ended embodied agent with large language models

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. V oyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291,

    Pith/arXiv arXiv

  8. [15]

    EAT: Efficient adaptive thinking for LLMs via early exit.arXiv preprint arXiv:2503.05061,

    Hao Wang et al. EAT: Efficient adaptive thinking for LLMs via early exit.arXiv preprint arXiv:2503.05061,

    arXiv

  9. [16]

    Thought calibration.arXiv preprint arXiv:2505.18404,

    Ruixuan Wu, Weijia Shi, and Luke Zettlemoyer. Thought calibration.arXiv preprint arXiv:2505.18404,

    Pith/arXiv arXiv

  10. [17]

    On-device language models: A comprehensive review.arXiv preprint arXiv:2409.00088,

    Jianing Xu et al. On-device language models: A comprehensive review.arXiv preprint arXiv:2409.00088,

    Pith/arXiv arXiv

  11. [18]

    Large language model cascades with mixture of thought representations for cost-efficient reasoning.arXiv preprint arXiv:2310.03094,

    Murong Yue, Jie Zhao, Min Zhang, Liang Du, and Ziyu Yao. Large language model cascades with mixture of thought representations for cost-efficient reasoning.arXiv preprint arXiv:2310.03094,

    Pith/arXiv arXiv

  12. [19]

    Language agent tree search unifies reasoning, acting, and planning in language models.arXiv preprint arXiv:2310.04406,

    Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. Language agent tree search unifies reasoning, acting, and planning in language models.arXiv preprint arXiv:2310.04406,

    Pith/arXiv arXiv

  13. [21]

    Mean Token Entropy (MTE): UMTE(y|x) = 1 L LX i=1 H(τ i |x, τ<i, θ),(11) whereH(τ i | ·)is the Shannon entropy of the token distribution at positioni

    for deferral and is used in all our experiments. Mean Token Entropy (MTE): UMTE(y|x) = 1 L LX i=1 H(τ i |x, τ<i, θ),(11) whereH(τ i | ·)is the Shannon entropy of the token distribution at positioni. While SP and PPL ex- amine the probability of thegeneratedtoken, MTE measures the spread of the full distribution, cap- turing uncertainty that SP and PPL may...

    2021

  14. [22]

    D THOUGHTCALIBRATION: STOPPINGRULEILLUSTRATION Fig

    as the default choice forAas it requires no assump- tions on the structure ofΛ, and provides unconditional FWER control in any setting. D THOUGHTCALIBRATION: STOPPINGRULEILLUSTRATION Fig. 6 illustrates the stopping rule from Section 4.1 on a single interaction step. The edge model generates thought tokensτ t,1, τt,2, . . .sequentially; at each strided pos...

    2026

  15. [1961]

    Therefore: Actions are encoded in hidden representations before text generation.arXiv preprint arXiv:2604.01202,

    Nishant Esakkiraja et al. Therefore: Actions are encoded in hidden representations before text generation.arXiv preprint arXiv:2604.01202,

    Pith/arXiv arXiv

  16. [2020]

    Conformal prediction for large language model inference.arXiv preprint arXiv:2502.09535,

    Max Frankel et al. Conformal prediction for large language model inference.arXiv preprint arXiv:2502.09535,

    arXiv

  17. [2021]

    Cai, Michael Terry, Quoc V

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie J. Cai, Michael Terry, Quoc V . Le, and Charles Sutton. Program synthesis with large language models.arXiv preprint arXiv:2108.07732,

    Pith/arXiv arXiv

  18. [2022]

    Angelopoulos, Stephen Bates, Emmanuel J

    9 Anastasios N. Angelopoulos, Stephen Bates, Emmanuel J. Cand `es, Michael I. Jordan, and Lihua Lei. Learn then test: Calibrating predictive algorithms to achieve risk control.arXiv preprint arXiv:2110.01052,

    Pith/arXiv arXiv

  19. [2023]

    Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168,

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168,

    Pith/arXiv arXiv

  20. [2024]

    FrugalGPT: How to use large language models while reducing cost and improving performance.arXiv preprint arXiv:2305.05176,

    Lingjiao Chen, Matei Zaharia, and James Zou. FrugalGPT: How to use large language models while reducing cost and improving performance.arXiv preprint arXiv:2305.05176,

    Pith/arXiv arXiv

  21. [2025]

    DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Pith/arXiv arXiv

  22. [2026]

    Statistically valid hyperparameter selection: From tuning to guarantees.arXiv preprint arXiv:2606.25601,

    Amirmohammad Farzaneh and Osvaldo Simeone. Statistically valid hyperparameter selection: From tuning to guarantees.arXiv preprint arXiv:2606.25601,

    Pith/arXiv arXiv

  23. [3584]

    only-large-right

    For each of the100GSM8K train-split problems, we decode the thought trace greedily until</think>orL max = 384tokens. A single forward pass with hidden-state output providesh (ℓ) t,p at every thought position and every layer at zero marginal cost. For each strided positionp∈ Pt, we splice</think>afterτ t,1:p and decode to obtainA p,t. The convergence label...

    arXiv 2021