Pith. sign in

REVIEW 5 major objections 5 minor 25 references

Evaluating LLMs as Interpretable Controllers for Dynamical Systems

T0 review · 5 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Pith's one-line read This paper shows that large language models can act as interpretable closed-loop controllers for a physical system, provided the model is sufficiently large and the prompt grounds the actuator semantics.

desk verdict A useful scaling probe with an overclaimed grounding result: the prediction tool is the same fitted model as the plant, so the 'domain knowledge' improvement is structurally expected. read the letter →

arxiv 2607.22609 v1 pith:53GC5IMX submitted 2026-06-14 cs.AI

classification cs.AI
keywords LargeLanguageModelsinterpretablecontroldynamicalsystemsphysics-basedpredictionexplainableAIreasoningtaxonomytemperaturetrackingmodelscale
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

The paper asks whether LLMs can do closed-loop control of a physical system and explain their decisions, rather than merely chat about plans. Five models of different scales were asked to track a temperature setpoint in a heater-and-fan enclosure, in plain language, including scenarios with penalties on actuator use and with a physics-based forward-simulation tool. The result is a clear scale dependence: the two largest models track the reference profile with mean errors below about 0.1 °C, use actuators smoothly, and produce explanations aligned with the physics, while smaller models misread which actuator heats and which cools, or give explanations that contradict their own actions. Adding a predictive physics tool makes the larger models more anticipatory, smoother, and more energy-efficient. The authors read this as evidence that sufficiently capable LLMs can serve as transparent controllers, especially in hybrid settings where a model-based predictor grounds their reasoning.

What carries the argument

The central mechanism is the closed-loop LLM controller: a chat agent prompted with the current temperature, ambient, setpoint, and (when applicable) an actuator-usage penalty, which must emit a structured JSON plan (heater float, fan binary) plus a natural-language rationale. On top of this, a forward-simulation tool ('simulate_plan') lets the model run candidate action sequences against the physics-based model before acting. The interpretive lens that carries the scale-dependence argument is the five-dimension reasoning taxonomy used to classify every explanation; it is the lens through which the authors convert control trajectories into evidence about what the models understand.

What would settle it

Re-run the five models over multiple seeds and sampling temperatures in the same enclosure, and have independent raters (or an automated rubric) re-code a sample of explanations. If Qwen-3 14B's MAE varies by more than about 0.5 °C across runs, or if the same model's explanation changes from 'cohesive' to 'misinterpreted' across trials, the claim that high-complexity models maintain stable, physically grounded reasoning would be falsified.

Watch

Extended reading notes

Core claim

On the authors' account, an LLM can act as a viable closed-loop controller for a dynamical system when two conditions hold: the model is large enough to maintain a stable internal representation of the causal actuator effects, and the natural-language prompt plus, optionally, a physics-based forward simulator, grounds the mapping from states to actions. In their testbed—a 50 cm cuboidal enclosure with a discrete-step heater duty cycle and a binary fan, tracking a 360-minute reference profile at 60-second steps—Qwen-3 14B and GPT-4o keep mean absolute error below 0.1–0.2 °C across all penalty scenarios and produce heater cycles that anticipate the reference, whereas Gemma-3 1B and Phi-4 Mini

Load-bearing premise

The load-bearing premise is that each reported trajectory is a representative single run and that the qualitative reasoning-taxonomy labels are reproducible; if LLM outputs are stochastic across runs or the taxonomy labels would not be assigned the same way by independent raters, the scale-dependent conclusion—small models misread dynamics, large models reason cohesively—is unsupported.

Editorial extensions

If this is right

  • If the central claim holds, LLM-based controllers are viable for slow, well-specified physical processes—at least when the model is sufficiently capable and the prompt is unambiguous about actuator semantics.
  • Natural-language penalties (e.g., 'minimize heater use') are effectively communicated only to high-complexity models; mid-scale models may 'understand' the penalty but break the tracking objective, so prompt-language multi-objective control is not yet reliable across scales.
  • Giving a capable LLM a physics-based forward simulator yields smoother trajectories, fewer overshoots, and lower energy use—evidence that hybrid model-based + language-driven control can improve both performance and interpretability.
  • Explanation quality correlates with control performance, so the reasoning taxonomy could serve as a coarse diagnostic: if a controller's rationale is physically coherent and temporally aware, its actions are more likely to be sound.
  • The results suggest that parameter count is not the only driver: Qwen-3 14B outperformed the much larger GPT-4o in both precision and energy efficiency, pointing to architectural and training differences.

Reading between the lines

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

  • A natural next test, not run here, is whether giving the forward-simulation tool to weaker models (e.g., Mistral 7B) rescues their closed-loop performance; if it does not, the bottleneck is causal reasoning rather than missing predictive information, which would sharpen the scale-dependence story.
  • Because each trajectory appears to be a single run, the numerical MAE claims are fragile; repeating the experiment across seeds and sampling temperatures would show whether the 0.086 °C figure for Qwen-3 is a stable property or a lucky draw.
  • The reasoning taxonomy is a qualitative instrument; encoding it as an automated rubric (or measuring inter-rater agreement) would let the claimed progression from 'hallucinated physics' to 'temporally aware reasoning' be tested quantitatively, and would make the interpretability claim more than anecdotal.
  • If LLM controllers are to be used in safety-critical settings, the same prompt-engineering and tool-based grounding could be combined with classical MPC or PID as a fallback—an architecture that keeps the LLM's explanations but adds formal stability guarantees.
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. The paper evaluates five LLMs (Gemma-3 1B, Phi-4 Mini 3.8B, Mistral 7B, Qwen-3 14B, GPT-4o) as closed-loop controllers for a simulated thermal enclosure described by a first-order physics-based model (Eq. 1). Controllers receive state/setpoint information in natural language and return heater duty-cycle and fan commands with textual reasoning. Experiments cover three settings (no penalty, heater penalty, fan penalty), plus a tool-assisted setting in which selected models may call a forward simulator before acting. The paper reports MAE and actuator-effort results, claims that high-complexity models track the reference well and produce physically coherent explanations, and argues that providing the physics-based prediction tool substantially improves control smoothness and energy efficiency. A qualitative reasoning taxonomy is used to support a scale-dependent progression in causal, temporal, and cohesive reasoning.

Significance. If the claims were fully supported, the paper would be a useful step toward LLM-based interpretable control and hybrid LLM+PBM pipelines. The manuscript is transparent about its prompt design, model equation, and tuning procedure, and the trace tables in the appendix give concrete examples of model behavior. However, the current evidence for the central scale-dependency and tool-grounding conclusions is substantially weaker than the abstract suggests: the tool experiment is close to circular, the quantitative results are single-run, and the reasoning taxonomy lacks coding discipline. With additional experiments and more rigorous analysis, the core idea could be publishable, but the present version overstates what has been demonstrated.

major comments (5)
  1. [Section 4.2 / Sections 2.2, 3.1, 3.3] The tool-assisted improvement claim is structurally guaranteed with the current setup. The plant is the fitted physics-based model (Section 3.1: "continuous updates of values from the simulated model"; Sections 2.2 and 3.3 use Eq. 1 with fitted H'_max and F'_max), and the tool is a forward simulator of the same Eq. 1 with the same fitted parameters. When the prediction tool is an exact oracle of the plant, any model that performs even a small amount of candidate-plan search will appear to improve; this does not test whether LLMs can incorporate model-based knowledge under model mismatch, sensor noise, or uncertainty. The conclusion in the abstract and Section 5 that "incorporating a physics-based model significantly improves control smoothness and energy efficiency" is therefore only supported in the oracle setting. Please quantify the tool-on vs tool-off difference (MAE, actuator usage,
  2. [Section 4.1 / Figure 6] All quantitative performance results appear to be single trajectories. The text reports MAE values (e.g., 0.086 °C for Qwen-3, 0.098 °C with heater penalty) without stating sampling temperature, seeds, number of repetitions, or run-to-run variability. LLM generation is stochastic, and the observed differences between mid- and high-complexity models could easily be within the noise of a single run. The central scale-dependent conclusion and the penalty comparisons need repeated runs with confidence intervals, or at minimum a clear statement that these are illustrative single-run observations rather than statistically supported results.
  3. [Section 4.3 / Tables 2-14, Figure 8] The reasoning taxonomy is assigned by the authors without a coding rubric, inter-rater reliability check, or quantitative grounding. The radar chart in Figure 8 has no numerical basis in the text, and the reported "progression from causal misinterpretation in smaller models to cohesive, temporally aware reasoning in larger ones" is supported only by selected examples. This is load-bearing for the interpretability claims. Please provide a explicit rubric, blind coding by multiple annotators with agreement statistics (e.g., Cohen's kappa), and preferably classification counts across all collected explanations rather than illustrative quotes.
  4. [Section 3.2 / System prompt] The system prompt explicitly states the physical semantics and a control rule: "The fan will make T smaller, the heater will make T larger. Let e = T−Tsp and use smaller control values when |e| is small." This hands the models both the actuator causal structure and a proportional-control heuristic. Consequently, the observation that high-complexity models give "coherent explanations aligned with physical principles" may in large part reflect the model reciting or following the prompt, rather than exhibiting emergent physical reasoning. This does not invalidate the engineering value of prompt-grounding, but it weakens the paper's broader claim about model understanding. Please discuss this explicitly and separate 'ability to follow explicitly stated semantics' from 'ability to infer or reason about physics'.
  5. [Section 4.2 / Figure 7] The tool comparison is also presented without quantitative summary. The text describes "smoother and more stable" behavior and lower heater duty cycle, but no MAE, overshoot count, or actuator-usage numbers are given for tool-on vs tool-off for either model. Even under the oracle assumption, the magnitude of the claimed improvement is therefore not established. Please add a small table with these metrics for both conditions.
minor comments (5)
  1. [Throughout] The prompt text repeatedly writes "JASON" instead of "JSON" (e.g., in the system prompt block). This appears to be a typo and should be corrected.
  2. [Table 1] The table has "V ocabulary" and "Unkonwn" (should be "Vocabulary" and "Unknown"). Also, the GPT-4o reference [22] is the GPT-4 Technical Report, not a GPT-4o-specific reference; please cite the appropriate source if one exists.
  3. [Appendix] The "MPC.txt" example in the appendix is presented without context. It appears to be an illustrative prompt with an SQL-style tool trace, but it is not clearly connected to the main experimental workflow. Please either integrate it into the main text or explain its role.
  4. [Section 5 / Conclusion] The bullet "Does incorporating prior model-based knowledge improve control performance?" answers affirmatively, but given the oracle-tool circularity noted in Major Comment 1, the answer should be qualified to the exact-model case. The future-work paragraph already lists imperfect model parameters and sensor noise; these are precisely what the current experiment omits.
  5. [Figures 3-5] The multi-panel figures are dense; the caption does not state which subplot corresponds to which model or condition. Adding clear panel labels and a legend consistent with the text would improve readability.

Circularity Check

1 steps flagged · score 6.0 of 10

Tool-assisted improvement is guaranteed by construction: the plant and the physics-based prediction tool are the same fitted Eq. 1, reducing the third-axis result to oracle access.

  1. self definitional [Sections 3.2–3.3 and 4.2 (tool-assisted experiment; Eq. 1)]
    "In the tool-assisted setting, the models can call a forward simulator to evaluate candidate action sequences over a 10-minute planning horizon ... The physics-based model is based on Equation 1. ... these new values resulted in improved model performance. ... This message feature has been used to provide the model with continuous updates of values from the simulated model ... uses the predictive model to simulate the outcome of candidate control sequences before acting, yielding smoother and more stable temperature tracking."

    The closed-loop plant is the simulated model (Sec. 3.1), which is the fitted PBM of Eq. 1 with H'_max and F'_max estimated from the measured dataset (Sec. 3.3). The 'prediction tool' is a forward simulator of the same Eq. 1 with the same fitted constants. Therefore the tool's predicted temperature profile is the plant's exact trajectory by construction, i.e., an oracle. The reported improvement in smoothness and energy efficiency is a logical consequence of consulting the plant's own equations, not an empirical demonstration that LLMs incorporate domain knowledge under model uncertainty or plant-model mismatch. The third-axis conclusion reduces to the definition of the tool.

full rationale

The core scale-comparison experiments (Sec. 4.1) are self-contained: LLM outputs are fed to the simulated plant and MAE/actuator usage are measured directly; no fitted parameter is relabeled as a prediction there. The reasoning taxonomy is subjective but not circular. However, the tool-assisted axis is partially circular: the plant and the physics-based prediction tool are the same Equation 1 with the same fitted parameters, so the tool is an exact oracle of the plant. The improvement with the tool is therefore guaranteed whenever the LLM consults it and does not test grounding under mismatch, sensor noise, or uncertainty. This reduces one headline claim to a construction. Self-citations [15,16] are background and not load-bearing. No other circular steps were found.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central empirical comparison relies on a fitted ODE simulator used as both plant and predictor; no new physical entities are introduced, but the free parameters and idealizing assumptions limit generality.

free parameters (3)
  • H'_max (heater power) = not reported numerically in text
    Estimated in Section 3.3 from measured data via Eq. 2 (average of rho*V*Cp*(T_t - T_{t-1})/dt during heater-on/fan-off periods); used to simulate the plant and to power the prediction tool.
  • F'_max (fan airflow) = not reported numerically in text
    Estimated in Section 3.3 from measured data via Eq. 3 during heater-off/fan-on periods; used in the plant model and prediction tool.
  • Planning horizon (HORIZON_MIN) = 10 minutes
    Chosen by the authors for the tool-assisted controller (Section 3.2, tool prompt); a longer or shorter horizon would change the anticipatory benefit measured in Section 4.2.
assumptions (3)
  • domain assumption Energy balance ODE (Eq. 1) describes the enclosure temperature dynamics
    Assumes perfectly mixed air, constant rho and Cp, instantaneous actuator response, no wall heat loss, and no plant effects (Section 2.2). These assumptions define both the simulated plant and the prediction tool.
  • domain assumption Fitted PBM parameters from the same dataset are representative of the evaluation scenario
    H'_max and F'_max are calibrated on a measured dataset (Section 3.3) that defines the evaluation; the dataset is not released, so this representativeness cannot be checked.
  • domain assumption A single run per model/scenario is representative of model behavior
    The paper reports one trajectory per model and scenario without seeds or repeated trials, so claims about model-scale effects assume deterministic or stable sampling behavior (Section 4.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluating LLMs as Interpretable Controllers for Dynamical Systems." pith.science (2026). https://pith.science/paper/53GC5IMX

@misc{pith2026260722609,
  author       = {Pith},
  title        = {Pith review of: Evaluating LLMs as Interpretable Controllers for Dynamical Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/53GC5IMX}},
  note         = {Machine review of arXiv:2607.22609}
}
read the original abstract

Large Language Models (LLMs) are increasingly used for decision-making and reasoning tasks, yet their potential as controllers for physical systems remains largely unexplored. This work investigates whether LLMs can function as interpretable controllers for a dynamic thermal environment, examining their ability to follow setpoints, interpret natural-language commands, reason about actuator effects, and incorporate prior model-based knowledge. Five LLMs of varying scales are evaluated under multiple scenarios, including settings with penalties on heater or fan usage and cases where the models have access to a physics-based prediction tool. The results show that control performance depends on model complexity: while low- and mid-scale models frequently misinterpret actuator dynamics or generate inconsistent reasoning, high-complexity models such as Qwen-3~14B and GPT-4o achieve accurate temperature tracking, stable actuator usage, and coherent explanations aligned with physical principles. Incorporating a physics-based model significantly improves control smoothness and energy efficiency by enabling anticipatory decision-making. A detailed reasoning taxonomy further reveals a clear progression from causal misinterpretation in smaller models to cohesive and temporally aware reasoning in larger ones. The findings demonstrate that LLMs can act as interpretable controllers when sufficiently capable and appropriately grounded in domain knowledge, highlighting promising opportunities for hybrid model-based and language-driven control strategies that can provide plausible explanations.

Figures

Figures reproduced from arXiv: 2607.22609 by the authors.

Figure 1
Figure 1. Schematic. The ChatOllama class also supports the use of structured outputs, forcing the model to deliver its response in JASON format. This approach provides a consistent form of output from the LLM, which in turn can automatically be extracted and turned into actionable control inputs without relying on fragile text parsing. The workflow for the models with and without the prediction tool is shown in [PITH_FULL_I… view at source ↗
Figure 2
Figure 2. LLM controller workflows. The blue boxes are nodes that the controller [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 5
Figure 5. Temperature profile and control sequences with fan penalty. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Temperature profile and control sequences with heater penalty. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Histogram with the performance of the models for all the cases. Lower values indicate better performance. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: Radar-chart comparison of model reasoning behaviours across the five [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 11 linked inside Pith

  1. [1]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, I. Polosukhin, Attention Is All You Need, arXiv:1706.03762 (2023)

  2. [2]

    Ichter, A

    B. Ichter, A. Brohan, Y . Chebotar, C. Finn, K. Hausman, A. Herzog, D. Ho, J. Ibarz, A. Irpan, E. Jang, R. Julian, D. Kalashnikov, S. Levine, Y . Lu, C. Parada, K. Rao, P. Sermanet, A. T. Toshev, V . Vanhoucke, F. Xia, T. Xiao, P. Xu, M. Yan, N. Brown, M. Ahn, O. Cortes, N. Sievers, C. Tan, S. Xu, D. Reyes, J. Rettinghouse, J. Quiambao, P. Pastor, L. Luu,...

  3. [3]

    Zitkovich, T

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid, Q. Vuong, V . Vanhoucke, H. Tran, R. Soricut, A. Singh, J. Singh, P. Sermanet, P. R. Sanketi, G. Salazar, M. S. Ryoo, K. Reymann, K. Rao, K. Pertsch, I. Mordatch, H. Michalewski, Y . Lu, S. Levine, L. Lee, T.-W. E. Lee, I. Leal, Y . Kuang, D. Kalashnikov, R. Julia...

  4. [4]

    Huang, F

    W. Huang, F. Xia, T. Xiao, H. Chan, J. Liang, P. Florence, A. Zeng, J. Tompson, I. Mordatch, Y . Chebotar, P. Ser- manet, T. Jackson, N. Brown, L. Luu, S. Levine, K. Haus- man, B. Ichter, Inner Monologue: Embodied Reasoning through Planning with Language Models, inProceedings of The 6th Conference on Robot Learning, PMLR, 2023, pp. 1769–1782, ISSN: 2640-3498

  5. [5]

    M. Chen, J. Tworek, H. Jun, Q. Yuan, H. P. d. O. Pinto, J. Kaplan, H. Edwards, Y . Burda, N. Joseph, G. Brock- man, A. Ray, R. Puri, G. Krueger, M. Petrov, H. Khlaaf, G. Sastry, P. Mishkin, B. Chan, S. Gray, N. Ryder, M. Pavlov, A. Power, L. Kaiser, M. Bavarian, C. Win- ter, P. Tillet, F. P. Such, D. Cummings, M. Plappert, F. Chantzis, E. Barnes, A. Herbe...

  6. [6]

    Schick, J

    T. Schick, J. Dwivedi-Yu, R. Dessì, R. Raileanu, M. Lomeli, L. Zettlemoyer, N. Cancedda, T. Scialom, Tool- former: Language Models Can Teach Themselves to Use Tools, arXiv:2302.04761 [cs] (2023)

  7. [7]

    E. O. Pyzer-Knapp, M. Manica, P. Staar, L. Morin, P. Ruch, T. Laino, J. R. Smith, A. Curioni, Foundation models for materials discovery – current state and future directions, npj Computational Materials, 11:61 (2025), Publisher: Na- ture Publishing Group

  8. [8]

    Singhal, S

    K. Singhal, S. Azizi, T. Tu, S. S. Mahdavi, J. Wei, H. W. Chung, N. Scales, A. Tanwani, H. Cole-Lewis, S. Pfohl, P. Payne, M. Seneviratne, P. Gamble, C. Kelly, A. Babiker, N. Schärli, A. Chowdhery, P. Mansfield, D. Demner- Fushman, B. Agüera y Arcas, D. Webster, G. S. Cor- rado, Y . Matias, K. Chou, J. Gottweis, N. Tomasev, Y . Liu, A. Rajkomar, J. Barral...

Show all 25 references
  1. [9]

    Huang, P

    W. Huang, P. Abbeel, D. Pathak, I. Mordatch, Lan- guage Models as Zero-Shot Planners: Extracting Action- able Knowledge for Embodied Agents, inProceedings of the 39th International Conference on Machine Learning, PMLR, 2022, pp. 9118–9147, ISSN: 2640-3498

  2. [10]

    C. H. Song, B. M. Sadler, J. Wu, W.-L. Chao, C. Washing- ton, Y . Su, LLM-Planner: Few-Shot Grounded Planning for Embodied Agents with Large Language Models, in 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 2986–2997, ISSN: 2380-7504

  3. [11]

    Stiennon, L

    N. Stiennon, L. Ouyang, J. Wu, D. Ziegler, R. Lowe, C. V oss, A. Radford, D. Amodei, P. F. Christiano, Learn- ing to summarize with human feedback, inAdvances in Neural Information Processing Systems, vol. 33, Curran Associates, Inc., 2020, pp. 3008–3021

  4. [12]

    H. Hu, D. Sadigh, Language Instructed Reinforcement Learning for Human-AI Coordination, inProceedings of the 40th International Conference on Machine Learning, PMLR, 2023, pp. 13584–13598, ISSN: 2640-3498

  5. [13]

    J. Lin, Y . Du, O. Watkins, D. Hafner, P. Abbeel, D. Klein, A. Dragan, Learning to Model the World with Language, arXiv:2308.01399 [cs] (2024)

  6. [14]

    J. Wang, E. Shi, H. Hu, C. Ma, Y . Liu, X. Wang, Y . Yao, X. Liu, B. Ge, S. Zhang, Large language models for robotics: Opportunities, challenges, and perspectives,Jour- nal of Automation and Intelligence, 4:52–64 (2025)

  7. [15]

    Rasheed, O

    A. Rasheed, O. Ravik, O. San, Large Language Models for Control, arXiv:2511.00337 [eess] (2025)

  8. [16]

    Rasheed, O

    A. Rasheed, O. Ravik, O. San, Hybrid Modeling, Sim-to- Real Reinforcement Learning, and Large Language Model Driven Control for Digital Twins, arXiv:2510.23882 [cs] (2025)

  9. [17]

    Turpin, J

    M. Turpin, J. Michael, E. Perez, S. R. Bowman, Lan- guage Models Don’t Always Say What They Think: Un- faithful Explanations in Chain-of-Thought Prompting, arXiv:2305.04388 [cs] (2023)

  10. [18]

    G. Team, A. Kamath, J. Ferret, S. Pathak, N. Vieillard, R. Merhej, S. Perrin, T. Matejovicova, A. Ramé, M. Riv- ière, L. Rouillard, T. Mesnard, G. Cideron, J.-b. Grill, S. Ramos, E. Yvinec, M. Casbon, E. Pot, I. Penchev, G. Liu, F. Visin, K. Kenealy, L. Beyer, X. Zhai, A. Tsit...

  11. [19]

    Abouelenin, A

    Microsoft, A. Abouelenin, A. Ashfaq, A. Atkinson, H. Awadalla, N. Bach, J. Bao, A. Benhaim, M. Cai, V . Chaudhary, C. Chen, D. Chen, D. Chen, J. Chen, W. Chen, Y .-C. Chen, Y .-l. Chen, Q. Dai, X. Dai, R. Fan, M. Gao, M. Gao, A. Garg, A. Goswami, J. Hao, A. Hendy, Y . Hu, X. J...

  12. [20]

    A. Q. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. S. Chaplot, D. d. l. Casas, F. Bressand, G. Lengyel, G. Lample, L. Saulnier, L. R. Lavaud, M.-A. Lachaux, P. Stock, T. L. Scao, T. Lavril, T. Wang, T. Lacroix, W. E. Sayed, Mistral 7B (2023)

  13. [21]

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, 12 B. Yu, C. Gao, C. Huang, C. Lv, C. Zheng, D. Liu, F. Zhou, F. Huang, F. Hu, H. Ge, H. Wei, H. Lin, J. Tang, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Zhou, J. Lin, K. Dang, K. Bao, K. Yang, L. Yu, L. Deng, M...

  14. [22]

    OpenAI, GPT-4 Technical Report, arXiv:2303.08774 [cs] (2024)

  15. [23]

    Hinton, O

    G. Hinton, O. Vinyals, J. Dean, Distilling the Knowledge in a Neural Network, arXiv:1503.02531 [cs, stat] (2015)

  16. [24]

    J. Yang, Z. Wang, Y . Lin, Z. Zhao, Problematic Tokens: Tokenizer Bias in Large Language Models, in2024 IEEE International Conference on Big Data (BigData), IEEE Computer Society, Washington, DC, USA, 2024, pp. 6387– 6393, arXiv:2406.11214 [cs] version: 3

  17. [25]

    Virtanen, R

    P. Virtanen, R. Gommers, T. E. Oliphant, M. Haber- land, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, S. J. van der Walt, M. Brett, J. Wil- son, K. J. Millman, N. Mayorov, A. R. J. Nelson, E. Jones, R. Kern, E. Larson, C. J. Carey, I. Polat, Y . ...

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.