CyberCorrect: A Cybernetic Framework for Closed-Loop Self-Correction in Large Language Models

Yang Shu; Yingmin Liu; Yuning Wu

arxiv: 2605.17305 · v1 · pith:6GW4G3A3new · submitted 2026-05-17 · 💻 cs.AI · cs.CL

CyberCorrect: A Cybernetic Framework for Closed-Loop Self-Correction in Large Language Models

Yuning Wu , Yingmin Liu , Yang Shu This is my paper

Pith reviewed 2026-05-20 13:37 UTC · model grok-4.3

classification 💻 cs.AI cs.CL

keywords LLM self-correctioncybernetic frameworkclosed-loop controlerror detectionconvergence controlreasoning taskscontrol theory metrics

0 comments

The pith

CyberCorrect models LLM self-correction as a closed-loop cybernetic control system to raise accuracy and reduce over-corrections.

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

The paper seeks to move beyond generic prompts that ask an LLM to reconsider its output by instead casting the entire correction process as an engineered feedback loop. It positions the language model itself as the plant under control, installs a sensor that combines consistency checks, stated confidence, and logic verification to classify errors, and adds a controller that issues fixes matched to the diagnosed error type plus a judge that halts iterations once stability criteria are met. This structure is evaluated on a new set of 440 annotated reasoning tasks using control-inspired metrics that track not only the final answer but also how the corrections behave across rounds. A reader would care because unstructured self-correction often produces loops or new mistakes, and a principled loop could make iterative improvement more dependable.

Core claim

By treating the LLM generator as the plant in a closed-loop control system, supplying a tri-modal Error Detector that fuses self-consistency, verbalized confidence, and logic-chain verification to diagnose error categories, deploying a type-directed Correction Controller to issue targeted repair instructions, and using a Convergence Judge to terminate iterations according to adapted stability criteria, the CyberCorrect framework achieves 79.8 percent final accuracy on reasoning tasks while cutting the overshoot rate by 41 percent relative to prior self-correction baselines.

What carries the argument

The closed-loop control architecture that uses a tri-modal Error Detector (self-consistency, verbalized confidence, and logic-chain verification) as the sensor, a type-directed Correction Controller for category-specific repairs, and a Convergence Judge that applies stability criteria drawn from control theory.

If this is right

Produces 79.8 percent final accuracy on the 440-task CyberCorrect-Bench, exceeding the strongest prior self-correction method by 6.2 percentage points.
Lowers the overshoot rate of erroneous over-corrections by 41 percent through the added convergence control.
Supplies three control-theoretic metrics—convergence rate, overshoot rate, and oscillation rate—that evaluate correction dynamics across iterations rather than final accuracy alone.
Replaces generic reconsideration prompts with repairs explicitly matched to diagnosed error categories.

Where Pith is reading between the lines

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

The same loop structure could be reused to stabilize other iterative LLM behaviors such as multi-step planning or tool-calling sequences once the detector is retuned for those error patterns.
Control-theoretic ideas might allow automatic adjustment of the judge's stability thresholds when moving to larger models or new task distributions.
The benchmark's annotated error types and correction paths provide a reusable template for building test suites that measure dynamic correction performance in future work.

Load-bearing premise

The tri-modal Error Detector must correctly identify the category of each mistake so the Correction Controller can produce useful targeted repairs and the Convergence Judge can stop the process without introducing new errors.

What would settle it

Applying the full loop to a collection of problems in which the error detector systematically misclassifies error types, such as treating reasoning flaws as factual errors, and finding that final accuracy then drops below the level achieved by simple self-correction prompts would falsify the benefit of the structured approach.

Figures

Figures reproduced from arXiv: 2605.17305 by Yang Shu, Yingmin Liu, Yuning Wu.

**Figure 2.** Figure 2: Accuracy across correction iterations. CyberCorrect converges mono [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Correction success rate by error type. Arithmetic errors are most [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

read the original abstract

Large language model (LLM) self-correction -- the ability to detect and fix errors in generated outputs -- remains largely ad hoc, relying on generic prompts such as "please reconsider your answer" without systematic error analysis or convergence guarantees. We propose CyberCorrect, a framework that formalizes LLM self-correction as a closed-loop control system grounded in cybernetic theory. The framework models the LLM generator as the plant and introduces a tri-modal Error Detector (combining self-consistency, verbalized confidence, and logic-chain verification) as the sensor. A type-directed Correction Controller generates targeted repair instructions based on diagnosed error categories, while a Convergence Judge determines iteration termination using stability criteria adapted from control theory. We further introduce three control-theoretic evaluation metrics -- convergence rate, overshoot rate, and oscillation rate -- that capture correction dynamics beyond final accuracy. Experiments on our constructed CyberCorrect-Bench (440 reasoning tasks with annotated error types and correction paths) show that CyberCorrect achieves 79.8% final accuracy, improving upon the best existing self-correction method by 6.2 percentage points, while reducing overshoot (erroneous over-correction) by 41% through its convergence control mechanism.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper gives a control-theoretic structure to LLM self-correction and new dynamic metrics, but its gains on a self-annotated benchmark need closer checks for circularity.

read the letter

The main thing to know is that this paper casts LLM self-correction as a cybernetic control system and introduces metrics that track the dynamics of the correction process. They get 79.8 percent accuracy on their benchmark, which is 6.2 points above the prior best, and cut overshoot by 41 percent. What is new is the overall framing plus the specific components: the tri-modal Error Detector using self-consistency, verbalized confidence, and logic-chain verification; the type-directed Correction Controller; and the Convergence Judge that uses adapted stability criteria to stop iterating. The three control-theoretic metrics are also presented as fresh. The paper does well in giving a coherent structure to what has been mostly prompt-based fixes and in defining ways to measure not just end accuracy but how the loop behaves. The soft spots are in the experiments. The results come from CyberCorrect-Bench, a 440-task set the authors created with annotated error types and correction paths. Since the controller relies on the detector producing categories that match those annotations, and no detector agreement rates or ablations disabling the type routing are mentioned, the improvements might trace back to the benchmark design more than to the cybernetic principles. Details on experimental controls and significance are also missing from the abstract. This paper is for people focused on making LLM reasoning more reliable through iterative self-fix methods. A reader who wants structured ways to evaluate correction loops could pick up the metrics. It has enough formal grounding and a clear empirical claim to warrant a serious referee. I would recommend putting it through peer review, with the referee asked to examine the benchmark construction and detector validation closely.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes CyberCorrect, a cybernetic framework for closed-loop self-correction in LLMs. It models the LLM generator as the plant, introduces a tri-modal Error Detector (self-consistency, verbalized confidence, and logic-chain verification) as the sensor, a type-directed Correction Controller for targeted repairs based on diagnosed error categories, and a Convergence Judge that applies stability criteria from control theory to terminate iterations. The work introduces three new metrics—convergence rate, overshoot rate, and oscillation rate—and evaluates on a new CyberCorrect-Bench of 440 reasoning tasks with author-annotated error types and correction paths, reporting 79.8% final accuracy (+6.2 pp over the best existing self-correction baseline) and a 41% reduction in overshoot.

Significance. If the central empirical claims hold after addressing evaluation concerns, the framework could offer a more systematic, control-theoretic alternative to ad-hoc self-correction prompting in LLMs, with the dynamics-focused metrics providing a useful extension beyond final accuracy. The explicit grounding in cybernetic principles and the introduction of a Convergence Judge represent a conceptual contribution, though the overall impact depends on demonstrating that gains arise from the closed-loop mechanisms rather than benchmark-specific features.

major comments (3)

[Experiments / CyberCorrect-Bench] CyberCorrect-Bench construction (abstract and Experiments section): The benchmark supplies ground-truth error-type annotations and correction paths that directly inform the type-directed Correction Controller and Convergence Judge. No detector-vs-annotation agreement rate, confusion matrix, or ablation disabling type-directed routing is described. This is load-bearing for the headline claims of 79.8% accuracy and 41% overshoot reduction, as it leaves open the possibility that observed improvements reflect alignment with the authors' annotations rather than independent diagnostic power of the tri-modal Error Detector.
[Evaluation Metrics] Evaluation metrics and controls (abstract and Results): The control-theoretic metrics (convergence rate, overshoot rate, oscillation rate) are defined from the framework itself and used to quantify improvements, yet the manuscript provides no details on experimental controls, statistical significance testing, variance across runs, or whether the metrics are computed independently of internal framework states. This undermines the ability to confirm that the reported gains are robust and not fitted to the custom benchmark.
[§3 / Tri-modal Error Detector] Tri-modal Error Detector reliability (abstract and §3): The central assumption that the combination of self-consistency, verbalized confidence, and logic-chain verification reliably diagnoses error categories for the Correction Controller is not supported by any quantitative validation against external or held-out error labels. Without this, the 6.2 pp accuracy improvement cannot be confidently attributed to the cybernetic control principles.

minor comments (2)

[Abstract] The abstract would benefit from explicitly naming the strongest baseline method and the LLM backbone(s) used in experiments.
[Convergence Judge] Notation for the stability criteria adapted from control theory in the Convergence Judge could be formalized with equations for reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and describe the revisions we will incorporate to strengthen the empirical support for our claims.

read point-by-point responses

Referee: [Experiments / CyberCorrect-Bench] CyberCorrect-Bench construction (abstract and Experiments section): The benchmark supplies ground-truth error-type annotations and correction paths that directly inform the type-directed Correction Controller and Convergence Judge. No detector-vs-annotation agreement rate, confusion matrix, or ablation disabling type-directed routing is described. This is load-bearing for the headline claims of 79.8% accuracy and 41% overshoot reduction, as it leaves open the possibility that observed improvements reflect alignment with the authors' annotations rather than independent diagnostic power of the tri-modal Error Detector.

Authors: We agree this is an important omission. The error-type annotations were produced by the authors solely to define the benchmark tasks and to enable evaluation of correction paths; they were not used to train or tune the tri-modal Error Detector, which operates exclusively on self-consistency, verbalized confidence, and logic-chain verification. In the revised manuscript we will report detector-annotation agreement rates, include a confusion matrix, and add an ablation that disables type-directed routing while keeping the rest of the framework fixed. These additions will clarify that performance gains arise from the detector's independent diagnostic power rather than annotation alignment. revision: yes
Referee: [Evaluation Metrics] Evaluation metrics and controls (abstract and Results): The control-theoretic metrics (convergence rate, overshoot rate, oscillation rate) are defined from the framework itself and used to quantify improvements, yet the manuscript provides no details on experimental controls, statistical significance testing, variance across runs, or whether the metrics are computed independently of internal framework states. This undermines the ability to confirm that the reported gains are robust and not fitted to the custom benchmark.

Authors: We will revise the Results section to specify that the three control-theoretic metrics are computed exclusively from observable iteration traces (final output, number of iterations, and whether an erroneous correction occurred), independent of any internal controller state. We will report standard deviations across five runs with different random seeds, include paired statistical significance tests against baselines, and describe the fixed experimental controls (temperature, prompt templates, and stopping criteria). These details will be added to demonstrate robustness. revision: yes
Referee: [§3 / Tri-modal Error Detector] Tri-modal Error Detector reliability (abstract and §3): The central assumption that the combination of self-consistency, verbalized confidence, and logic-chain verification reliably diagnoses error categories for the Correction Controller is not supported by any quantitative validation against external or held-out error labels. Without this, the 6.2 pp accuracy improvement cannot be confidently attributed to the cybernetic control principles.

Authors: We acknowledge that the current manuscript lacks an explicit quantitative validation of the detector against external labels. The tri-modal detector is intentionally unsupervised and draws on well-studied individual techniques. In revision we will add a new experiment that compares detector outputs against a held-out subset of 100 tasks whose error categories were re-annotated by an independent annotator, reporting precision, recall, and agreement statistics. This will provide direct evidence linking detector reliability to the observed accuracy gains. revision: partial

Circularity Check

2 steps flagged

Author-defined control metrics and self-constructed benchmark with annotated error types create partial circularity in claimed gains

specific steps

renaming known result [Abstract]
"We further introduce three control-theoretic evaluation metrics -- convergence rate, overshoot rate, and oscillation rate -- that capture correction dynamics beyond final accuracy."

Overshoot is explicitly glossed as 'erroneous over-correction' and the metrics are presented as novel control-theoretic quantities; they re-label standard self-correction error rates using cybernetic terminology and then claim quantitative improvements on those re-labeled quantities.
self definitional [Abstract (CyberCorrect-Bench description)]
"Experiments on our constructed CyberCorrect-Bench (440 reasoning tasks with annotated error types and correction paths)"

The tri-modal Error Detector diagnoses error categories that are presupposed to align with the author-supplied annotations in the benchmark; the type-directed Correction Controller and Convergence Judge are then shown to improve performance on the same annotated data, making the reported gains partly dependent on the benchmark's construction rather than an independent test.

full rationale

The paper introduces a cybernetic framing and three new metrics (convergence rate, overshoot rate, oscillation rate) to evaluate its tri-modal detector and type-directed controller. Experiments are run exclusively on the authors' CyberCorrect-Bench, which supplies the ground-truth error-type annotations that the detector is designed to recover. While final accuracy is compared to external baselines, the reported 41% overshoot reduction and convergence improvements are measured using quantities defined inside the framework itself and validated against the same author annotations. This produces moderate circularity in the evaluation loop without reducing the entire result to a pure tautology or self-citation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that control-theoretic stability criteria can be directly adapted to discrete LLM token generation and that the tri-modal detector provides independent error signals; no free parameters, axioms, or invented entities are explicitly quantified in the abstract.

axioms (1)

domain assumption LLM generation can be modeled as a controllable plant whose outputs respond to targeted correction instructions in a stable, measurable way.
Invoked when the framework treats the LLM as the plant and applies convergence criteria from control theory.

invented entities (1)

Tri-modal Error Detector no independent evidence
purpose: Combines self-consistency, verbalized confidence, and logic-chain verification to diagnose error categories.
New component introduced by the framework; no independent evidence outside the paper is mentioned in the abstract.

pith-pipeline@v0.9.0 · 5746 in / 1462 out tokens · 39715 ms · 2026-05-20T13:37:12.665273+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We model LLM self-correction as a discrete-time closed-loop control system... yt+1 = G(x, yt, ut) ... ut = C(τ(et), s(et), ℓ(et))
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

three control-theoretic evaluation metrics—convergence rate, overshoot rate, and oscillation rate

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages · 3 internal anchors

[1]

Language models are few-shot learners,

T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askellet al., “Language models are few-shot learners,” inAdvances in Neural Information Processing Systems (NeurIPS), 2020

work page 2020
[2]

GPT-4 Technical Report

J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “GPT-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[3]

Self-refine: Iterative refinement with self-feedback,

A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y . Yanget al., “Self-refine: Iterative refinement with self-feedback,” inAdvances in Neural Information Processing Systems (NeurIPS), 2023

work page 2023
[4]

Reflexion: Language agents with verbal reinforcement learning,

N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Yao, “Reflexion: Language agents with verbal reinforcement learning,” in Advances in Neural Information Processing Systems (NeurIPS), 2023

work page 2023
[5]

Chain-of-verification reduces hallucination in large language models,

S. Dhuliawala, M. Komeili, J. Xu, R. Raileanu, X. Li, A. Celikyilmaz, and J. Weston, “Chain-of-verification reduces hallucination in large language models,” inFindings of the Association for Computational Linguistics (ACL), 2024

work page 2024
[6]

Large language models cannot self-correct reasoning yet,

J. Huang, X. Chen, S. Mishra, H. S. Zheng, A. W. Yu, X. Song, and D. Zhou, “Large language models cannot self-correct reasoning yet,” in International Conference on Learning Representations (ICLR), 2024

work page 2024
[7]

Wiener,Cybernetics: Or Control and Communication in the Animal and the Machine

N. Wiener,Cybernetics: Or Control and Communication in the Animal and the Machine. MIT Press, 1948

work page 1948
[8]

Ogata,Modern Control Engineering, 5th ed

K. Ogata,Modern Control Engineering, 5th ed. Prentice Hall, 2010

work page 2010
[9]

Auto- matically correcting large language models: Surveying the landscape of diverse automated correction strategies,

L. Pan, M. Saxon, W. Xu, D. Nathani, X. Wang, and W. Y . Wang, “Auto- matically correcting large language models: Surveying the landscape of diverse automated correction strategies,”Transactions of the Association for Computational Linguistics, vol. 12, pp. 484–506, 2024

work page 2024
[10]

REFINER: Reasoning feedback on intermediate representa- tions,

D. Paul, M. Ismayilzada, M. Peyrard, B. Borberás, J. Larrosa, and R. West, “REFINER: Reasoning feedback on intermediate representa- tions,” inConference of the European Chapter of the Association for Computational Linguistics (EACL), 2024

work page 2024
[11]

Generating sequences by learning to self-correct,

S. Welleck, X. Lu, P. West, F. Brahman, T. Shen, D. Khashabi, and Y . Choi, “Generating sequences by learning to self-correct,” in International Conference on Learning Representations (ICLR), 2023

work page 2023
[12]

Language Models (Mostly) Know What They Know

S. Kadavath, T. Conerly, A. Askell, T. Henighan, D. Drain, E. Perez, N. Schiefer, Z. Hatfield-Dodds, N. DasSarma, E. Tran-Johnsonet al., “Language models (mostly) know what they know,”arXiv preprint arXiv:2207.05221, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022
[13]

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, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[14]

K. J. Åström and B. Wittenmark,Adaptive Control, 2nd ed. Addison- Wesley, 1995

work page 1995
[15]

Tree of thoughts: Deliberate problem solving with large language models,

S. Yao, D. Yu, J. Zhao, I. Shafran, T. Griffiths, Y . Cao, and K. Narasimhan, “Tree of thoughts: Deliberate problem solving with large language models,” inAdvances in Neural Information Processing Systems (NeurIPS), 2023

work page 2023
[16]

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

X. Wang, J. Wei, D. Schuurmans, Q. Le, E. Chi, S. Narang, A. Chowdh- ery, and D. Zhou, “Self-consistency improves chain of thought reasoning in language models,” inInternational Conference on Learning Repre- sentations (ICLR), 2023

work page 2023

[1] [1]

Language models are few-shot learners,

T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askellet al., “Language models are few-shot learners,” inAdvances in Neural Information Processing Systems (NeurIPS), 2020

work page 2020

[2] [2]

GPT-4 Technical Report

J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “GPT-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023

[3] [3]

Self-refine: Iterative refinement with self-feedback,

A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y . Yanget al., “Self-refine: Iterative refinement with self-feedback,” inAdvances in Neural Information Processing Systems (NeurIPS), 2023

work page 2023

[4] [4]

Reflexion: Language agents with verbal reinforcement learning,

N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Yao, “Reflexion: Language agents with verbal reinforcement learning,” in Advances in Neural Information Processing Systems (NeurIPS), 2023

work page 2023

[5] [5]

Chain-of-verification reduces hallucination in large language models,

S. Dhuliawala, M. Komeili, J. Xu, R. Raileanu, X. Li, A. Celikyilmaz, and J. Weston, “Chain-of-verification reduces hallucination in large language models,” inFindings of the Association for Computational Linguistics (ACL), 2024

work page 2024

[6] [6]

Large language models cannot self-correct reasoning yet,

J. Huang, X. Chen, S. Mishra, H. S. Zheng, A. W. Yu, X. Song, and D. Zhou, “Large language models cannot self-correct reasoning yet,” in International Conference on Learning Representations (ICLR), 2024

work page 2024

[7] [7]

Wiener,Cybernetics: Or Control and Communication in the Animal and the Machine

N. Wiener,Cybernetics: Or Control and Communication in the Animal and the Machine. MIT Press, 1948

work page 1948

[8] [8]

Ogata,Modern Control Engineering, 5th ed

K. Ogata,Modern Control Engineering, 5th ed. Prentice Hall, 2010

work page 2010

[9] [9]

Auto- matically correcting large language models: Surveying the landscape of diverse automated correction strategies,

L. Pan, M. Saxon, W. Xu, D. Nathani, X. Wang, and W. Y . Wang, “Auto- matically correcting large language models: Surveying the landscape of diverse automated correction strategies,”Transactions of the Association for Computational Linguistics, vol. 12, pp. 484–506, 2024

work page 2024

[10] [10]

REFINER: Reasoning feedback on intermediate representa- tions,

D. Paul, M. Ismayilzada, M. Peyrard, B. Borberás, J. Larrosa, and R. West, “REFINER: Reasoning feedback on intermediate representa- tions,” inConference of the European Chapter of the Association for Computational Linguistics (EACL), 2024

work page 2024

[11] [11]

Generating sequences by learning to self-correct,

S. Welleck, X. Lu, P. West, F. Brahman, T. Shen, D. Khashabi, and Y . Choi, “Generating sequences by learning to self-correct,” in International Conference on Learning Representations (ICLR), 2023

work page 2023

[12] [12]

Language Models (Mostly) Know What They Know

S. Kadavath, T. Conerly, A. Askell, T. Henighan, D. Drain, E. Perez, N. Schiefer, Z. Hatfield-Dodds, N. DasSarma, E. Tran-Johnsonet al., “Language models (mostly) know what they know,”arXiv preprint arXiv:2207.05221, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022

[13] [13]

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, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024

[14] [14]

K. J. Åström and B. Wittenmark,Adaptive Control, 2nd ed. Addison- Wesley, 1995

work page 1995

[15] [15]

Tree of thoughts: Deliberate problem solving with large language models,

S. Yao, D. Yu, J. Zhao, I. Shafran, T. Griffiths, Y . Cao, and K. Narasimhan, “Tree of thoughts: Deliberate problem solving with large language models,” inAdvances in Neural Information Processing Systems (NeurIPS), 2023

work page 2023

[16] [16]

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

X. Wang, J. Wei, D. Schuurmans, Q. Le, E. Chi, S. Narang, A. Chowdh- ery, and D. Zhou, “Self-consistency improves chain of thought reasoning in language models,” inInternational Conference on Learning Repre- sentations (ICLR), 2023

work page 2023