Towards Network-Aware Operation of Integrated Energy Systems: A Comprehensive Review

Alessandra Parisio

Network constraints and topology, often overlooked, decide what integrated multi-energy systems can actually do.

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

2026-07-15 13:21 UTC pith:UY4Y3PME

load-bearing objection Wrong full text is attached to this arXiv ID; only the IES abstract is usable, so we cannot honestly evaluate the review’s novelty or methods yet. the 2 major comments →

arxiv 2603.07266 v2 pith:UY4Y3PME submitted 2026-03-07 eess.SY cs.SY

Towards Network-Aware Operation of Integrated Energy Systems: A Comprehensive Review

Alessandra Parisio This is my paper

classification eess.SY cs.SY

keywords integrated energy systemsnetwork-aware operationmulti-energy systemssector couplingdistributed optimizationenergy network topologyfeasibility guaranteeslow-carbon energy systems

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.

Integrated energy systems couple electricity, gas, heating, and cooling networks (and sometimes hydrogen, transport, or water) so the carriers depend on one another. In practice the networks are still planned and run largely in isolation, which wastes flexibility and efficiency, especially as distributed resources tighten local coupling. The review argues that prior work has underplayed the explicit role of network constraints and topology, even though these shape feasible operating regions, affect scalability, and determine how uncertainty and formal guarantees can be handled. It surveys network-aware modeling, optimization, and control methods, diagnoses limits in tractability, feasibility guarantees, and scalability, and points toward distributed optimization with theoretical guarantees and control informed by operational data. A reader cares because low-carbon cities will need coordinated multi-carrier operation that respects real network physics rather than treating carriers as abstract balances.

Core claim

This review provides what it presents as the first comprehensive analysis of network-aware modeling, optimization, and control methods for integrated energy systems. It claims that network constraints and topology are a key point the literature often overlooks, yet they shape feasible operating regions, affect scalability, and determine how uncertainty and formal guarantees can be addressed. From that analysis it identifies methodological limits in tractability, feasibility guarantees, and scalability, and outlines research directions including distributed optimization with theoretical guarantees and data-informed control as foundations for scalable, network-aware operation of future low-car

What carries the argument

Network constraints and topology. These are the physical limits and structural layout of the coupled electricity, gas, heating, and cooling networks. They carry the argument by defining the true feasible set of multi-carrier operating points, by controlling whether optimization and control methods remain tractable at scale, and by determining what kinds of uncertainty handling and formal performance guarantees are possible.

Load-bearing premise

The paper rests on the premise that earlier work has overlooked network constraints and topology enough for this to count as the first comprehensive network-aware treatment, and that the dominant barriers are methodological rather than data, markets, or regulation.

What would settle it

A systematic literature survey with an explicit inclusion protocol that finds a substantial prior body of reviews or methods already treating topology and network constraints as first-class objects in multi-carrier optimization and control, with comparable coverage of tractability and guarantees, would undermine the central "first comprehensive" claim.

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

If this is right

Coordinated multi-carrier schedules that ignore network topology will systematically under-use flexibility and can produce infeasible operating points.
Scalable IES operation will require distributed optimization methods that come with theoretical guarantees rather than purely heuristic coordination.
Control designs informed by operational data can help manage multi-scale dynamics that pure model-based methods struggle with.
Future low-carbon operational frameworks must treat network physics as first-class objects rather than secondary constraints.
Progress on tractability and feasibility guarantees is a prerequisite for reliable deployment of tightly coupled multi-energy systems.

Where Pith is reading between the lines

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

Market and regulatory designs that settle only on energy balances without network-level rights may lock in the isolation the paper criticizes.
The same topology-driven feasible-region arguments likely extend to hydrogen and transport networks once they are tightly coupled, even if current literature is thinner there.
Data-informed control across separately owned networks will probably need privacy-preserving or federated forms that the review only hints at.
A shared benchmark suite of multi-network test cases with known topology would make the claimed gaps in guarantees easier to measure and close.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Wrong full text is attached to this arXiv ID; only the IES abstract is usable, so we cannot honestly evaluate the review’s novelty or methods yet.

read the letter

Colleague — stop before you dig in. The cacheable “full manuscript” for 2603.07266 is not the energy-systems review. It is a completely different paper (Trace Inversion / reasoning-trace stealing, arXiv 2603.07267, cs.CR). So we have only the abstract of the IES piece.

From that abstract alone: the pitch is a first comprehensive, network-aware survey of modeling, optimization, and control for integrated electricity–gas–heat–cooling (plus hydrogen/transport/water) systems. The claimed angle is that topology and network constraints are routinely under-weighted, and that this shapes feasible regions, scalability, uncertainty handling, and formal guarantees. The agenda points toward distributed optimization with guarantees and data-informed control. That framing is useful if the body actually delivers a clean taxonomy and gap analysis; multi-energy surveys already exist, so the “first comprehensive network-aware” claim is the load-bearing novelty assertion and is currently uncheckable.

What we cannot assess: inclusion protocol, comparison tables, how they treat non-convex multi-scale dynamics, whether the “overlooked topology” claim is fair to the literature, or whether the research directions are more than standard wishlist items. Circularity risk is low for a review; soundness and novelty scores have to stay provisional. No code, data, or formal proofs are claimed in the abstract.

Who it is for: people working on multi-energy operation, DER coupling, and distributed control who want a map of network-aware methods. A serious editor should still send a real full text to referees if the body matches the abstract’s ambition — the topic matters for low-carbon city operation — but desk-level judgment is impossible on the wrong PDF.

Recommendation: do not cite or discuss the substance until the correct manuscript is in hand. Flag the ID/text mismatch, request the right file, then re-read. Until then, treat this as abstract-only and unreviewed.

Referee Report

2 major / 0 minor

Summary. The manuscript claims to be a first comprehensive review of network-aware modeling, optimization, and control for Integrated Energy Systems (IES), arguing that electricity, gas, heating/cooling (and optionally hydrogen, transport, water) networks are still planned and operated in isolation despite growing DER-driven coupling. It asserts that explicit network constraints and topology—often overlooked—shape feasible operating regions, scalability, and the treatment of uncertainty and formal guarantees, and that the review identifies limitations in tractability, feasibility guarantees, and scalability while outlining directions such as distributed optimization with theoretical guarantees and data-informed control.

Significance. If the review delivers a rigorous, protocol-based taxonomy of network-aware methods and a defensible gap analysis, it would be a useful foundation for low-carbon multi-energy operation research. The abstract correctly flags non-convexity, multi-scale dynamics, and interdependencies as central difficulties. However, the supplied full-text body is an unrelated paper on LLM reasoning-trace inversion (arXiv:2603.07267), so none of the promised literature synthesis, comparison tables, or methodological claims can be verified. Significance therefore remains conditional on the correct manuscript being provided.

major comments (2)

The CACHEABLE full manuscript text is not the IES review (arXiv:2603.07266). It is instead the complete paper “How to Steal Reasoning Without Reasoning Traces” (arXiv:2603.07267, cs.CR). No sections, equations, tables, inclusion criteria, or gap taxonomy for network-aware IES operation are present. The central claim of a “first comprehensive analysis” and the asserted methodological limitations cannot be assessed from the supplied body.
Because the wrong full text was provided, load-bearing elements required of a review—search protocol, inclusion/exclusion criteria, classification of network-aware vs. network-agnostic formulations, and evidence that topology/constraints are systematically overlooked—cannot be checked. A revised submission must supply the actual IES manuscript body before technical review is possible.

Circularity Check

0 steps flagged

No circular derivation: review abstract has no equation-level reduction; supplied full text is a different paper and also non-circular.

full rationale

The target abstract (IES network-aware review, arXiv:2603.07266) is a synthesis claim, not a first-principles derivation. It asserts that network constraints/topology are often overlooked and that the review is a first comprehensive network-aware analysis, then flags tractability, feasibility guarantees, and scalability as open issues. There are no fitted parameters renamed as predictions, no self-definitional identities (X defined via Y then used to predict Y), and no uniqueness theorem imported from the same authors that forces the result. The only soft framing risk is the unquantified “first comprehensive” branding, which is not circular math under the stated patterns. Independently, the CACHEABLE full-text block is a completely different manuscript (Trace Inversion / “How to Steal Reasoning Without Reasoning Traces,” arXiv:2603.07267). That paper trains inversion models on inputs/answers/(optional) summaries and evaluates token overlap and student fine-tuning gains on external benchmarks; its claims are empirical and do not reduce by construction to their inputs. With no load-bearing step that equals its own definition or fit, circularity score is 0 and steps are empty.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

Abstract-only review of multi-carrier energy systems. Load-bearing content rests on standard domain definitions of IES and on the unproven meta-claim that network topology/constraints are systematically under-treated. No free parameters or invented physical entities; axioms are domain assumptions about system structure and literature gaps.

axioms (4)

domain assumption Integrated Energy Systems are defined as interconnected electricity, gas, heating, and cooling networks (optionally hydrogen, transport, water) with cross-vector dependence.
Definitional premise of the entire review; stated in the opening of the abstract.
domain assumption Modern multi-energy networks are still planned and operated largely in isolation, causing inefficiency and unused flexibility.
Empirical/institutional premise that motivates the survey; not evidenced in the abstract.
ad hoc to paper Network constraints and topology are a key factor often overlooked in the IES literature and shape feasibility, scalability, uncertainty handling, and formal guarantees.
Central organizing claim of this review; treated as established gap without citation evidence in the provided text.
domain assumption IES operation is hard due to interdependencies, non-convex behaviors, and multi-scale network dynamics.
Standard systems-engineering characterization used to justify optimization/control focus.

reviewed 2026-07-15 · how reviews work

0 comments

read the original abstract

Integrated Energy Systems (IES) are systems of interconnected electricity, gas, heating, and cooling networks, where the carriers interact and depend on one another. Beyond these core vectors, IES may also incorporate additional infrastructures, such as hydrogen, transportation and water networks, whenever sector coupling or cross-vector exchanges are relevant. Although modern cities already function as multi-energy systems, these networks are still planned and operated in isolation, which leads to inefficiencies and unused flexibility. As distributed energy resources (DERs) grow, local coupling among electricity, heating, and gas networks becomes stronger, so coordinated operation across carriers and infrastructures is essential. IES can improve efficiency, flexibility, and renewable integration, yet operation is challenging because of complex interdependencies, non-convex behaviors, and multi-scale dynamics of the energy networks. A key point that the literature often overlooks is the explicit role of network constraints and topology, which shape feasible operating regions, affect scalability, and determine how uncertainty and formal guarantees can be addressed. This review provides a first comprehensive analysis of network-aware modeling, optimization, and control methods for IES. We identify methodological limitations related to tractability, feasibility guarantees, and scalability. Building on these insights, we outline research directions that include distributed optimization with theoretical guarantees and control approaches informed by operational data. The review offers a foundation for scalable, network-aware operational frameworks for future low-carbon energy systems.

discussion (0)

Reference graph

Works this paper leans on

45 extracted references · 10 linked inside Pith

[1]

Building with extended thinking, 2025

Anthropic. Building with extended thinking, 2025. Accessed: 2025-12-12

2025
[2]

Detecting and preventing distillation attacks

Anthropic. Detecting and preventing distillation attacks. https://www.anthropic.com/ news/detecting-and-preventing-distillation-attacks, Feb. 2026

2026
[3]

Arora, H

D. Arora, H. Singh, et al. Have LLMs advanced enough? a challenging problem solving benchmark for large language models. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7527–7543, 2023

2023
[4]

Y . Bai, A. Jones, K. Ndousse, A. Askell, A. Chen, N. DasSarma, D. Drain, S. Fort, D. Ganguli, T. Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback.arXiv:2204.05862, 2022

Pith/arXiv arXiv 2022
[5]

Carlini, F

N. Carlini, F. Tramer, E. Wallace, M. Jagielski, A. Herbert-V oss, K. Lee, A. Roberts, T. Brown, D. Song, U. Erlingsson, et al. Extracting training data from large language models. In Proceedings of the 30th USENIX Security Symposium, pages 2633–2650, 2021

2021
[6]

H. Chen, T. Zhu, L. Zhang, B. Liu, D. Wang, W. Zhou, and M. Xue. QUEEN: Query unlearning against model extraction.IEEE Transactions on Information Forensics and Security, 20:2143– 2156, 2025

2025
[7]

X. Chen, S. Zhou, K. Liang, X. Sun, and X. Liu. Skip-thinking: Chunk-wise chain-of-thought distillation enable smaller language models to reason better and faster. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 12153–12168, 2025

2025
[8]

Gemini thinking: Thought summaries and internal reasoning, 2025

Google. Gemini thinking: Thought summaries and internal reasoning, 2025. Accessed: 2025- 12-12

2025
[9]

J. Gou, B. Yu, S. J. Maybank, and D. Tao. Knowledge distillation: A survey.International Journal of Computer Vision, 129(6):1789–1819, 2021

2021
[10]

Green, M

T. Green, M. Gubri, H. Puerto, S. Yun, and S. J. Oh. Leaky thoughts: Large reasoning models are not private thinkers. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 26518–26540, 2025

2025
[11]

C. Gu, X. L. Li, P. Liang, and T. Hashimoto. On the learnability of watermarks for language models. InProceedings of the Twelfth International Conference on Learning Representations, 2024

2024
[12]

E. Guha, R. Marten, S. Keh, N. Raoof, G. Smyrnis, H. Bansal, M. Nezhurina, J. Mercat, T. Vu, Z. Sprague, et al. Openthoughts: Data recipes for reasoning models.arXiv:2506.04178, 2025

Pith/arXiv arXiv 2025
[13]

D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, et al. DeepSeek-R1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv:2501.12948, 2025

Pith/arXiv arXiv 2025
[14]

Hendrycks, C

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:2103.03874, 2021

Pith/arXiv arXiv 2021
[15]

Hinton, O

G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv:1503.02531, 2015

Pith/arXiv arXiv 2015
[16]

Holmes and W

T. Holmes and W. Gooderham. Exploiting DeepSeek-R1: Breaking down chain of thought security. https://www.trendmicro.com/en_us/research/25/c/ exploiting-deepseek-r1.html, 2025. 10

2025
[17]

N. Jain, K. Han, A. Gu, W.-D. Li, F. Yan, T. Zhang, S. Wang, A. Solar-Lezama, K. Sen, and I. Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. InProceedings of the Thirteenth International Conference on Learning Representations, 2025

2025
[18]

Kirchenbauer, J

J. Kirchenbauer, J. Geiping, Y . Wen, J. Katz, I. Miers, and T. Goldstein. A watermark for large language models. InProceedings of the International Conference on Machine Learning, pages 17061–17084. PMLR, 2023

2023
[19]

Kojima, S

T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa. Large language models are zero-shot reasoners.Advances in Neural Information Processing Systems, 35:22199–22213, 2022

2022
[20]

Krishna, G

K. Krishna, G. S. Tomar, A. P. Parikh, N. Papernot, and M. Iyyer. Thieves on sesame street! model extraction of bert-based apis.arXiv:1910.12366, 2019

Pith/arXiv arXiv 1910
[21]

P. Li, Z. Tan, M. Zhang, H. Qu, H. Liu, and T. Chen. DOGE: Defensive output generation for llm protection against knowledge distillation.arXiv:2505.19504, 2025

arXiv 2025
[22]

J. Liu, C. S. Xia, Y . Wang, and L. Zhang. Is your code generated by chatgpt really correct? rig- orous evaluation of large language models for code generation.Advances in neural information processing systems, 36:21558–21572, 2023

2023
[23]

H. Ma, T. Chen, T.-K. Hu, C. You, X. Xie, and Z. Wang. Undistillable: Making a nasty teacher that cannot teach students.arXiv:2105.07381, 2021

Pith/arXiv arXiv 2021
[24]

H. Ma, Y . Huang, H. Tang, C. You, D. Kong, and X. Xie. Sparse logits suffice to fail knowledge distillation. InProceedings of the ICLR 2022 Workshop on PAIR^2Struct: Privacy, Accountabil- ity, Interpretability, Robustness, Reasoning on Structured Data, 2022

2022
[25]

Mattern, F

J. Mattern, F. Mireshghallah, Z. Jin, B. Schölkopf, M. Sachan, and T. Berg-Kirkpatrick. Membership inference attacks against language models via neighbourhood comparison. arXiv:2305.18462, 2023

Pith/arXiv arXiv 2023
[26]

J. X. Morris, W. Zhao, J. T. Chiu, V . Shmatikov, and A. M. Rush. Language model inversion. InProceedings of the Twelfth International Conference on Learning Representations, 2024

2024
[27]

Muennighoff, Z

N. Muennighoff, Z. Yang, W. Shi, X. L. Li, L. Fei-Fei, H. Hajishirzi, L. Zettlemoyer, P. Liang, E. Candès, and T. B. Hashimoto. s1: Simple test-time scaling. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 20286–20332, 2025

2025
[28]

Openai o1 system card

OpenAI. Openai o1 system card. https://openai.com/index/ openai-o1-system-card/, 2024. Accessed: 2025-12-25

2024
[29]

Openai reasoning models documentation, 2025

OpenAI. Openai reasoning models documentation, 2025. Accessed: 2025-12-12

2025
[30]

Orekondy, B

T. Orekondy, B. Schiele, and M. Fritz. Knockoff nets: Stealing functionality of black-box models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4954–4963, 2019

2019
[31]

Ouyang, J

L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730–27744, 2022

2022
[32]

Sander, P

T. Sander, P. Fernandez, A. Durmus, M. Douze, and T. Furon. Watermarking makes language models radioactive.Advances in Neural Information Processing Systems, 37:21079–21113, 2024

2024
[33]

Savani, A

Y . Savani, A. Trockman, Z. Feng, Y . E. Xu, A. Schwarzschild, A. Robey, M. Finzi, and J. Z. Kolter. Antidistillation sampling.arXiv:2504.13146, 2025

arXiv 2025
[34]

Shokri, M

R. Shokri, M. Stronati, C. Song, and V . Shmatikov. Membership inference attacks against machine learning models. InProceedings of the 2017 IEEE symposium on Security and Privacy, pages 3–18. IEEE, 2017. 11

2017
[35]

Shridhar, A

K. Shridhar, A. Stolfo, and M. Sachan. Distilling reasoning capabilities into smaller language models. InProceedings of the Findings of the Association for Computational Linguistics: ACL 2023, pages 7059–7073, 2023

2023
[36]

Qwq: Reflect deeply on the boundaries of the unknown

Team Qwen. Qwq: Reflect deeply on the boundaries of the unknown. https://qwenlm. github.io/blog/qwq-32b-preview/, 2024

2024
[37]

Tramèr, F

F. Tramèr, F. Zhang, A. Juels, M. K. Reiter, and T. Ristenpart. Stealing machine learning models via prediction APIs. InProceedings of the 25th USENIX Security Symposium, pages 601–618, 2016

2016
[38]

Wallace, M

E. Wallace, M. Stern, and D. Song. Imitation attacks and defenses for black-box machine translation systems. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 5531–5546, 2020

2020
[39]

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:2203.11171, 2022

Pith/arXiv arXiv 2022
[40]

J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhou, et al. Chain-of- thought prompting elicits reasoning in large language models.Advances in Neural Information Processing Systems, 35:24824–24837, 2022

2022
[41]

R. Wen, Z. Li, M. Backes, and Y . Zhang. Membership inference attacks against in-context learn- ing. InProceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, pages 3481–3495, 2024

2024
[42]

Y . Yang, X. Zhang, Y . Jiang, X. Chen, H. Wang, S. Ji, and Z. Wang. PRSA: Prompt reverse stealing attacks against large language models.arXiv:2402.19200, 2024

Pith/arXiv arXiv 2024
[43]

Zhang, J

C. Zhang, J. X. Morris, and V . Shmatikov. Extracting prompts by inverting LLM outputs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14753–14777, 2024

2024
[44]

Zhang, N

Y . Zhang, N. Carlini, and D. Ippolito. Effective prompt extraction from language models. In Proceedings of the First Conference on Language Modeling, 2024

2024
[45]

K. Zhou, C. Liu, X. Zhao, S. Jangam, J. Srinivasa, G. Liu, D. Song, and X. E. Wang. The hidden risks of large reasoning models: A safety assessment of R1.arXiv:2502.12659, 2025. 12 Zero-shot Inversion Prompt (With Summaries) You are a language model that reconstructs full internal reasoning traces from high-level bubble summaries. You will be given: - A p...

arXiv 2025

This paper was first reviewed by grok-4.5 on July 15, 2026.

[1] [1]

Building with extended thinking, 2025

Anthropic. Building with extended thinking, 2025. Accessed: 2025-12-12

2025

[2] [2]

Detecting and preventing distillation attacks

Anthropic. Detecting and preventing distillation attacks. https://www.anthropic.com/ news/detecting-and-preventing-distillation-attacks, Feb. 2026

2026

[3] [3]

Arora, H

D. Arora, H. Singh, et al. Have LLMs advanced enough? a challenging problem solving benchmark for large language models. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7527–7543, 2023

2023

[4] [4]

Y . Bai, A. Jones, K. Ndousse, A. Askell, A. Chen, N. DasSarma, D. Drain, S. Fort, D. Ganguli, T. Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback.arXiv:2204.05862, 2022

Pith/arXiv arXiv 2022

[5] [5]

Carlini, F

N. Carlini, F. Tramer, E. Wallace, M. Jagielski, A. Herbert-V oss, K. Lee, A. Roberts, T. Brown, D. Song, U. Erlingsson, et al. Extracting training data from large language models. In Proceedings of the 30th USENIX Security Symposium, pages 2633–2650, 2021

2021

[6] [6]

H. Chen, T. Zhu, L. Zhang, B. Liu, D. Wang, W. Zhou, and M. Xue. QUEEN: Query unlearning against model extraction.IEEE Transactions on Information Forensics and Security, 20:2143– 2156, 2025

2025

[7] [7]

X. Chen, S. Zhou, K. Liang, X. Sun, and X. Liu. Skip-thinking: Chunk-wise chain-of-thought distillation enable smaller language models to reason better and faster. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 12153–12168, 2025

2025

[8] [8]

Gemini thinking: Thought summaries and internal reasoning, 2025

Google. Gemini thinking: Thought summaries and internal reasoning, 2025. Accessed: 2025- 12-12

2025

[9] [9]

J. Gou, B. Yu, S. J. Maybank, and D. Tao. Knowledge distillation: A survey.International Journal of Computer Vision, 129(6):1789–1819, 2021

2021

[10] [10]

Green, M

T. Green, M. Gubri, H. Puerto, S. Yun, and S. J. Oh. Leaky thoughts: Large reasoning models are not private thinkers. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 26518–26540, 2025

2025

[11] [11]

C. Gu, X. L. Li, P. Liang, and T. Hashimoto. On the learnability of watermarks for language models. InProceedings of the Twelfth International Conference on Learning Representations, 2024

2024

[12] [12]

E. Guha, R. Marten, S. Keh, N. Raoof, G. Smyrnis, H. Bansal, M. Nezhurina, J. Mercat, T. Vu, Z. Sprague, et al. Openthoughts: Data recipes for reasoning models.arXiv:2506.04178, 2025

Pith/arXiv arXiv 2025

[13] [13]

D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, et al. DeepSeek-R1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv:2501.12948, 2025

Pith/arXiv arXiv 2025

[14] [14]

Hendrycks, C

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:2103.03874, 2021

Pith/arXiv arXiv 2021

[15] [15]

Hinton, O

G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv:1503.02531, 2015

Pith/arXiv arXiv 2015

[16] [16]

Holmes and W

T. Holmes and W. Gooderham. Exploiting DeepSeek-R1: Breaking down chain of thought security. https://www.trendmicro.com/en_us/research/25/c/ exploiting-deepseek-r1.html, 2025. 10

2025

[17] [17]

N. Jain, K. Han, A. Gu, W.-D. Li, F. Yan, T. Zhang, S. Wang, A. Solar-Lezama, K. Sen, and I. Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. InProceedings of the Thirteenth International Conference on Learning Representations, 2025

2025

[18] [18]

Kirchenbauer, J

J. Kirchenbauer, J. Geiping, Y . Wen, J. Katz, I. Miers, and T. Goldstein. A watermark for large language models. InProceedings of the International Conference on Machine Learning, pages 17061–17084. PMLR, 2023

2023

[19] [19]

Kojima, S

T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa. Large language models are zero-shot reasoners.Advances in Neural Information Processing Systems, 35:22199–22213, 2022

2022

[20] [20]

Krishna, G

K. Krishna, G. S. Tomar, A. P. Parikh, N. Papernot, and M. Iyyer. Thieves on sesame street! model extraction of bert-based apis.arXiv:1910.12366, 2019

Pith/arXiv arXiv 1910

[21] [21]

P. Li, Z. Tan, M. Zhang, H. Qu, H. Liu, and T. Chen. DOGE: Defensive output generation for llm protection against knowledge distillation.arXiv:2505.19504, 2025

arXiv 2025

[22] [22]

J. Liu, C. S. Xia, Y . Wang, and L. Zhang. Is your code generated by chatgpt really correct? rig- orous evaluation of large language models for code generation.Advances in neural information processing systems, 36:21558–21572, 2023

2023

[23] [23]

H. Ma, T. Chen, T.-K. Hu, C. You, X. Xie, and Z. Wang. Undistillable: Making a nasty teacher that cannot teach students.arXiv:2105.07381, 2021

Pith/arXiv arXiv 2021

[24] [24]

H. Ma, Y . Huang, H. Tang, C. You, D. Kong, and X. Xie. Sparse logits suffice to fail knowledge distillation. InProceedings of the ICLR 2022 Workshop on PAIR^2Struct: Privacy, Accountabil- ity, Interpretability, Robustness, Reasoning on Structured Data, 2022

2022

[25] [25]

Mattern, F

J. Mattern, F. Mireshghallah, Z. Jin, B. Schölkopf, M. Sachan, and T. Berg-Kirkpatrick. Membership inference attacks against language models via neighbourhood comparison. arXiv:2305.18462, 2023

Pith/arXiv arXiv 2023

[26] [26]

J. X. Morris, W. Zhao, J. T. Chiu, V . Shmatikov, and A. M. Rush. Language model inversion. InProceedings of the Twelfth International Conference on Learning Representations, 2024

2024

[27] [27]

Muennighoff, Z

N. Muennighoff, Z. Yang, W. Shi, X. L. Li, L. Fei-Fei, H. Hajishirzi, L. Zettlemoyer, P. Liang, E. Candès, and T. B. Hashimoto. s1: Simple test-time scaling. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 20286–20332, 2025

2025

[28] [28]

Openai o1 system card

OpenAI. Openai o1 system card. https://openai.com/index/ openai-o1-system-card/, 2024. Accessed: 2025-12-25

2024

[29] [29]

Openai reasoning models documentation, 2025

OpenAI. Openai reasoning models documentation, 2025. Accessed: 2025-12-12

2025

[30] [30]

Orekondy, B

T. Orekondy, B. Schiele, and M. Fritz. Knockoff nets: Stealing functionality of black-box models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4954–4963, 2019

2019

[31] [31]

Ouyang, J

L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730–27744, 2022

2022

[32] [32]

Sander, P

T. Sander, P. Fernandez, A. Durmus, M. Douze, and T. Furon. Watermarking makes language models radioactive.Advances in Neural Information Processing Systems, 37:21079–21113, 2024

2024

[33] [33]

Savani, A

Y . Savani, A. Trockman, Z. Feng, Y . E. Xu, A. Schwarzschild, A. Robey, M. Finzi, and J. Z. Kolter. Antidistillation sampling.arXiv:2504.13146, 2025

arXiv 2025

[34] [34]

Shokri, M

R. Shokri, M. Stronati, C. Song, and V . Shmatikov. Membership inference attacks against machine learning models. InProceedings of the 2017 IEEE symposium on Security and Privacy, pages 3–18. IEEE, 2017. 11

2017

[35] [35]

Shridhar, A

K. Shridhar, A. Stolfo, and M. Sachan. Distilling reasoning capabilities into smaller language models. InProceedings of the Findings of the Association for Computational Linguistics: ACL 2023, pages 7059–7073, 2023

2023

[36] [36]

Qwq: Reflect deeply on the boundaries of the unknown

Team Qwen. Qwq: Reflect deeply on the boundaries of the unknown. https://qwenlm. github.io/blog/qwq-32b-preview/, 2024

2024

[37] [37]

Tramèr, F

F. Tramèr, F. Zhang, A. Juels, M. K. Reiter, and T. Ristenpart. Stealing machine learning models via prediction APIs. InProceedings of the 25th USENIX Security Symposium, pages 601–618, 2016

2016

[38] [38]

Wallace, M

E. Wallace, M. Stern, and D. Song. Imitation attacks and defenses for black-box machine translation systems. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 5531–5546, 2020

2020

[39] [39]

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:2203.11171, 2022

Pith/arXiv arXiv 2022

[40] [40]

J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhou, et al. Chain-of- thought prompting elicits reasoning in large language models.Advances in Neural Information Processing Systems, 35:24824–24837, 2022

2022

[41] [41]

R. Wen, Z. Li, M. Backes, and Y . Zhang. Membership inference attacks against in-context learn- ing. InProceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, pages 3481–3495, 2024

2024

[42] [42]

Y . Yang, X. Zhang, Y . Jiang, X. Chen, H. Wang, S. Ji, and Z. Wang. PRSA: Prompt reverse stealing attacks against large language models.arXiv:2402.19200, 2024

Pith/arXiv arXiv 2024

[43] [43]

Zhang, J

C. Zhang, J. X. Morris, and V . Shmatikov. Extracting prompts by inverting LLM outputs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14753–14777, 2024

2024

[44] [44]

Zhang, N

Y . Zhang, N. Carlini, and D. Ippolito. Effective prompt extraction from language models. In Proceedings of the First Conference on Language Modeling, 2024

2024

[45] [45]

K. Zhou, C. Liu, X. Zhao, S. Jangam, J. Srinivasa, G. Liu, D. Song, and X. E. Wang. The hidden risks of large reasoning models: A safety assessment of R1.arXiv:2502.12659, 2025. 12 Zero-shot Inversion Prompt (With Summaries) You are a language model that reconstructs full internal reasoning traces from high-level bubble summaries. You will be given: - A p...

arXiv 2025