arxiv: 2604.08276 · v1 · submitted 2026-04-09 · 💻 cs.AI · cs.CR

Recognition: 2 theorem links

· Lean Theorem

ACF: A Collaborative Framework for Agent Covert Communication under Cognitive Asymmetry

Wansheng Wu , Kaibo Huang , Yukun Wei , Zhongliang Yang , Linna Zhou

Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:24 UTC · model grok-4.3

classification 💻 cs.AI cs.CR

keywords covert communicationcognitive asymmetryagent networkssteganographyAI agentsprefix-independent decodingcollaborative framework

0 comments

The pith

ACF enables reliable covert communication in AI agent networks despite cognitive asymmetry between agents.

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

The paper addresses the problem of covert communication in autonomous agent networks where agents update their internal memories through interactions, leading to cognitive asymmetry. Traditional methods fail because they require identical sequence prefixes for encoding and decoding secrets. The proposed Asymmetric Collaborative Framework (ACF) decouples the covert communication from semantic reasoning using separate statistical and cognitive layers. It introduces a prefix-independent decoding method controlled by a shared steganographic configuration. This allows for reliable secret extraction with provable error bounds even under severe asymmetry, as shown in evaluations on memory-augmented workflows.

Core claim

The Asymmetric Collaborative Framework (ACF) structurally decouples covert communication from semantic reasoning via orthogonal statistical and cognitive layers. By deploying a prefix-independent decoding paradigm governed by a shared steganographic configuration, ACF eliminates the reliance on cognitive symmetry. This enables reliable secret extraction with provable error bounds under severe cognitive asymmetry in dynamic agent deployments, while maintaining computational indistinguishability and semantic fidelity.

What carries the argument

The prefix-independent decoding paradigm governed by a shared steganographic configuration that decouples covert communication from semantic reasoning in agent networks.

If this is right

Reliable secret extraction is possible even with prefix discrepancies caused by dynamic memory updates.
The framework provides Effective Information Capacity guarantees for agent networks.
Computational indistinguishability is preserved, avoiding detection.
Semantic fidelity and covert capabilities coexist under asymmetry conditions.

Where Pith is reading between the lines

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

This could facilitate hidden coordination in large-scale agent swarms where synchronization of states is impractical.
The shared configuration approach might apply to other forms of asymmetric communication channels in distributed systems.
Testing ACF in real-world agent platforms could reveal practical overheads not captured in simulated workflows.

Load-bearing premise

A shared steganographic configuration can be set up and sustained among agents without depending on cognitive symmetry or adding overhead that compromises the covert nature of the communication.

What would settle it

An observation that the ACF decoding fails to recover secrets accurately when agent memories diverge significantly, or that the outputs become distinguishable from normal agent communications.

Figures

Figures reproduced from arXiv: 2604.08276 by Kaibo Huang, Linna Zhou, Wansheng Wu, Yukun Wei, Zhongliang Yang.

**Figure 2.** Figure 2: Impact of controlled cognitive asymmetry on Bit Error Rate (BER). [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Trade-off analysis between semantic utility and channel reliability [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

read the original abstract

As generative artificial intelligence evolves, autonomous agent networks present a powerful paradigm for interactive covert communication. However, because agents dynamically update internal memories via environmental interactions, existing methods face a critical structural vulnerability: cognitive asymmetry. Conventional approaches demand strict cognitive symmetry, requiring identical sequence prefixes between the encoder and decoder. In dynamic deployments, inevitable prefix discrepancies destroy synchronization, inducing severe channel degradation. To address this core challenge of cognitive asymmetry, we propose the Asymmetric Collaborative Framework (ACF), which structurally decouples covert communication from semantic reasoning via orthogonal statistical and cognitive layers. By deploying a prefix-independent decoding paradigm governed by a shared steganographic configuration, ACF eliminates the reliance on cognitive symmetry. Evaluations on realistic memory-augmented workflows demonstrate that under severe cognitive asymmetry, symmetric baselines suffer severe channel degradation, whereas ACF uniquely excels across both semantic fidelity and covert communication. It maintains computational indistinguishability, enabling reliable secret extraction with provable error bounds, and providing robust Effective Information Capacity guarantees for modern agent networks.

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

ACF gives a workable structural split for covert agent comms under memory mismatch, but the shared config setup is the part that still needs concrete proof it doesn't recreate the original problem.

read the letter

The paper's core move is to treat cognitive asymmetry—different agent memories and prefixes—as a structural limit rather than something to patch with more symmetry. It splits the channel into an orthogonal statistical layer and a cognitive layer, then uses a prefix-independent decoder driven by a shared steganographic configuration. That framing looks new relative to the symmetric steganography baselines it cites, and the reported experiments on memory-augmented workflows show the symmetric methods degrade sharply while ACF holds semantic fidelity and covertness together. The claim of provable error bounds and Effective Information Capacity guarantees is the part that stands out as potentially useful if the derivations actually deliver independent checks rather than fitted parameters. The work engages the practical deployment issue in autonomous agent networks without obvious overclaiming in the abstract. The main soft spot is the bootstrap and maintenance of that shared configuration. The abstract does not spell out how agents initialize or update it without either assuming some common starting state or adding detectable traffic, which would undercut the covert goal. If the full paper supplies a mechanism plus analysis showing the overhead stays below detection thresholds, the central claim strengthens; otherwise the asymmetry problem is only moved, not removed. The citation pattern is standard and the experimental setup on realistic workflows is a plus. This is for researchers working on secure coordination in multi-agent systems. A reader who cares about real-world agent covert channels would find the framework idea worth testing. I would bring it to a reading group to walk through the config protocol and the bound derivations. I would not cite it yet. It deserves peer review because it targets a deployment barrier that existing methods actually hit.

Referee Report

3 major / 2 minor

Summary. The paper proposes the Asymmetric Collaborative Framework (ACF) to enable covert communication among autonomous agents under cognitive asymmetry arising from dynamic memory updates and differing sequence prefixes. It decouples covert communication from semantic reasoning using orthogonal statistical and cognitive layers, introduces a prefix-independent decoding paradigm governed by a shared steganographic configuration, and claims to achieve reliable secret extraction with provable error bounds, Effective Information Capacity guarantees, computational indistinguishability, and superior performance over symmetric baselines in evaluations on memory-augmented workflows.

Significance. If the provable bounds and covertness claims hold without reintroducing detectable overhead or symmetry assumptions, ACF could meaningfully advance steganographic methods for dynamic AI agent networks by addressing a structural vulnerability in prefix-dependent approaches. The orthogonal-layer design and emphasis on prefix independence represent a potentially useful structural contribution if supported by concrete derivations and reproducible evaluations.

major comments (3)

[Abstract and §3] The central claim (abstract and §3) that a shared steganographic configuration enables prefix-independent decoding and eliminates reliance on cognitive symmetry is load-bearing, yet the manuscript provides no explicit mechanism, protocol, or initialization procedure for establishing, synchronizing, or updating this configuration across agents with mismatched prefixes and memories. This leaves open whether the bootstrap step itself requires symmetry or introduces detectable communication that undermines the covert goal.
[§4] §4 (evaluations) reports that ACF maintains performance under severe cognitive asymmetry while symmetric baselines degrade, but the description lacks quantitative details on how asymmetry is induced (e.g., prefix discrepancy distributions, memory-update rates), error-bar analysis, or direct comparison to the claimed provable bounds. Without these, it is unclear whether the empirical results corroborate the theoretical guarantees.
[Abstract and §2] The abstract and §2 assert 'provable error bounds' and 'Effective Information Capacity guarantees,' but the manuscript does not show the derivation or state the assumptions under which these bounds hold (e.g., whether they depend on the shared configuration remaining secret and synchronized). This makes it difficult to assess whether the bounds are non-vacuous or reduce to fitted parameters.

minor comments (2)

[§3] Notation for the steganographic configuration and the prefix-independent decoder should be introduced with explicit symbols and a small illustrative example early in §3 to improve readability.
[§2] The paper should include a brief related-work subsection contrasting ACF with prior agent steganography methods that assume prefix symmetry, citing specific limitations addressed.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback, which helps clarify key aspects of the ACF framework. We address each major comment below and will revise the manuscript to incorporate additional details, derivations, and experimental specifications as outlined. These changes will strengthen the presentation of the shared configuration, theoretical guarantees, and empirical validation without altering the core contributions.

read point-by-point responses

Referee: [Abstract and §3] The central claim (abstract and §3) that a shared steganographic configuration enables prefix-independent decoding and eliminates reliance on cognitive symmetry is load-bearing, yet the manuscript provides no explicit mechanism, protocol, or initialization procedure for establishing, synchronizing, or updating this configuration across agents with mismatched prefixes and memories. This leaves open whether the bootstrap step itself requires symmetry or introduces detectable communication that undermines the covert goal.

Authors: We agree that an explicit description of the initialization and synchronization procedure is necessary to fully support the load-bearing claim. The manuscript in §3 presents the shared steganographic configuration as a pre-deployed orthogonal statistical layer that operates independently of cognitive states and prefix content, enabling prefix-independent decoding. However, we acknowledge the absence of a detailed bootstrap protocol. In the revised manuscript, we will add a new subsection in §3 specifying the initialization procedure: an initial secure setup phase (e.g., during agent deployment or via an out-of-band trusted channel) for distributing the configuration, followed by runtime updates that leverage the statistical layer itself for synchronization without semantic alignment or detectable overhead. This avoids reintroducing symmetry assumptions or covert-channel leakage, as the bootstrap occurs outside the dynamic interaction phase. revision: yes
Referee: [§4] §4 (evaluations) reports that ACF maintains performance under severe cognitive asymmetry while symmetric baselines degrade, but the description lacks quantitative details on how asymmetry is induced (e.g., prefix discrepancy distributions, memory-update rates), error-bar analysis, or direct comparison to the claimed provable bounds. Without these, it is unclear whether the empirical results corroborate the theoretical guarantees.

Authors: We appreciate this observation on the experimental reporting. The evaluations in §4 induce cognitive asymmetry through controlled variations in prefix lengths and memory-update frequencies within memory-augmented agent workflows, using mismatch rates sampled to reflect realistic dynamic environments. We agree that additional quantitative rigor is required. In the revision, we will expand §4 to include: explicit prefix discrepancy distributions (e.g., uniform sampling over 20-80% mismatch) and memory-update rates; error bars computed from 10 independent runs with standard deviations; and a direct comparison of observed error rates against the theoretical bounds from §2. This will provide clearer evidence that the empirical results align with and support the provable guarantees. revision: yes
Referee: [Abstract and §2] The abstract and §2 assert 'provable error bounds' and 'Effective Information Capacity guarantees,' but the manuscript does not show the derivation or state the assumptions under which these bounds hold (e.g., whether they depend on the shared configuration remaining secret and synchronized). This makes it difficult to assess whether the bounds are non-vacuous or reduce to fitted parameters.

Authors: We acknowledge that the full derivations and explicit assumptions for the provable error bounds and Effective Information Capacity (EIC) guarantees are not sufficiently detailed in the main text of §2. These bounds are derived analytically from the statistical distance between the steganographic embedding distribution and the cover distribution under the orthogonal-layer model, yielding an error probability of O(1/sqrt(n)) for sequence length n, contingent on the shared configuration remaining secret and synchronized. In the revised version, we will expand §2 with a dedicated derivation subsection (or move supporting material to an appendix) that states all assumptions explicitly—including configuration secrecy, prefix independence via statistical orthogonality, and independent memory updates—and provides the step-by-step proof. This will clarify that the bounds are analytically derived rather than fitted parameters. revision: yes

Circularity Check

1 steps flagged

Central claim reduces to unshown bootstrap of shared steganographic configuration

specific steps

self definitional [Abstract]
"By deploying a prefix-independent decoding paradigm governed by a shared steganographic configuration, ACF eliminates the reliance on cognitive symmetry."

The elimination of cognitive symmetry is asserted as a direct consequence of using the shared configuration, yet the configuration itself must be established and synchronized without identical prefixes or detectable communication; this bootstrap step is not derived or shown, rendering the core claim equivalent to assuming the shared state whose feasibility under asymmetry is the problem being solved.

full rationale

The abstract presents ACF as eliminating cognitive symmetry via a prefix-independent decoding paradigm governed by a shared steganographic configuration, with provable error bounds. However, no derivation, equations, or mechanism is shown for initializing or maintaining this shared configuration across agents with differing memories/prefixes without reintroducing symmetry or detectable overhead. This makes the elimination of symmetry and the error bounds claims reduce to an input assumption by construction, as the framework's functionality depends on the very shared state whose establishment is not independently justified in the provided text.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, so the ledger is populated from stated claims. The shared steganographic configuration is treated as an external input whose establishment cost is not analyzed.

free parameters (1)

shared steganographic configuration
The framework relies on this pre-shared element to enable prefix-independent decoding; its generation and distribution method is not detailed.

axioms (1)

domain assumption Agents can maintain a shared steganographic configuration without cognitive symmetry or detectable overhead.
Invoked when the abstract states that ACF 'eliminates the reliance on cognitive symmetry' via the shared configuration.

pith-pipeline@v0.9.0 · 5477 in / 1299 out tokens · 28997 ms · 2026-05-10T18:24:33.637674+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/AbsoluteFloorClosure.lean reality_from_one_distinction unclear

?

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

By deploying a prefix-independent decoding paradigm governed by a shared steganographic configuration, ACF eliminates the reliance on cognitive symmetry.
IndisputableMonolith/Foundation/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

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

ACF structurally decouples the covert channel from the agent’s dynamic reasoning via orthogonal statistical and cognitive layers

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

36 extracted references · 8 canonical work pages · 5 internal anchors

[1]

Provably secure robust image steganography,

Z. Yang, K. Chen, K. Zenget al., “Provably secure robust image steganography,”IEEE Transactions on Multimedia, vol. 26, pp. 5040– 5053, 2023

2023
[2]

Efficient audio steganography using generalized audio intrinsic energy with micro-amplitude modification suppression,

W. Su, J. Ni, X. Huet al., “Efficient audio steganography using generalized audio intrinsic energy with micro-amplitude modification suppression,”IEEE Transactions on Information Forensics and Security, vol. 19, pp. 6559–6572, 2024

2024
[3]

Large-capacity and flexible video steganography via invertible neural network,

C. Mou, Y . Xu, J. Songet al., “Large-capacity and flexible video steganography via invertible neural network,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 22 606–22 615

2023
[4]

Rnn-stega: Linguistic steganography based on recurrent neural networks,

Z.-L. Yang, X.-Q. Guo, Z.-M. Chenet al., “Rnn-stega: Linguistic steganography based on recurrent neural networks,”IEEE Transactions on Information Forensics and Security, vol. 14, no. 5, pp. 1280–1295, 2018

2018
[5]

Linguistic generative steganogra- phy with enhanced cognitive-imperceptibility,

Z. Yang, L. Xiang, S. Zhanget al., “Linguistic generative steganogra- phy with enhanced cognitive-imperceptibility,”IEEE Signal Processing Letters, vol. 28, pp. 409–413, 2021

2021
[6]

RNN-Stega: Linguistic steganography based on recurrent neural networks,

Z.-L. Yang, X.-Q. Guo, Z.-M. Chenet al., “RNN-Stega: Linguistic steganography based on recurrent neural networks,”IEEE Transactions on Information Forensics and Security, vol. 14, no. 5, pp. 1280–1295, May 2019

2019
[7]

Neural linguistic steganogra- phy,

Z. M. Ziegler, Y . Deng, and A. M. Rush, “Neural linguistic steganogra- phy,” inProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). Hong Kong, China: Association for Computational Linguistics, 2019, pp. 1210–1215

2019
[8]

Toolllm: Facilitating large language models to master 16000+ real-world apis,

Y . Qin, S. Liang, Y . Yeet al., “Toolllm: Facilitating large language models to master 16000+ real-world apis,”arXiv (Cornell University), 2023

2023
[9]

Generative agents: Interactive simulacra of human behavior,

J. S. Park, J. C. O’Brien, C. J. Caiet al., “Generative agents: Interactive simulacra of human behavior,” inProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST), 2023, pp. 1–22

2023
[10]

A survey on large language model based autonomous agents,

L. Wang, C. Ma, X. Fenget al., “A survey on large language model based autonomous agents,”Frontiers of Computer Science, 2024

2024
[11]

Agentbench: Evaluating llms as agents,

X. Liu, H. Yu, H. Zhanget al., “Agentbench: Evaluating llms as agents,” arXiv (Cornell University), 2023

2023
[12]

Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face,

Y . Shen, K. Song, X. Tanet al., “Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face,”arXiv (Cornell University), 2023

2023
[13]

Metagpt: Meta programming for a multi-agent collaborative framework,

S. Hong, M. Zhuge, J. Chenet al., “Metagpt: Meta programming for a multi-agent collaborative framework,”arXiv (Cornell University), 2023

2023
[14]

Communicative agents for software development,

Q. Chen, C. Xin, Y . Chenget al., “Communicative agents for software development,”Proceedings of the 62nd annual meeting of the associa- tion for computational linguistics (ACL), 2024

2024
[15]

Whispering Agents: An event-driven covert communication protocol for the Internet of Agents,

K. Huang, Y . Wei, W. Wuet al., “Whispering Agents: An event-driven covert communication protocol for the Internet of Agents,” Aug. 2025

2025
[16]

LLSM: Generative lin- guistic steganography with large language model,

Y . Wang, R. Song, R. Zhanget al., “LLSM: Generative lin- guistic steganography with large language model,”arXiv preprint arXiv:2401.15656, 2024

work page arXiv 2024
[17]

Generative text steganography with large language model,

J. Wu, Z. Wu, Y . Xueet al., “Generative text steganography with large language model,”Proceedings of the 32nd ACM International Conference on Multimedia, 2024

2024
[18]

Not what you’ve signed up for: Compromising real-world llm-integrated applications with indirect prompt injection,

K. Greshake, S. Abdelnabi, S. Mishraet al., “Not what you’ve signed up for: Compromising real-world llm-integrated applications with indirect prompt injection,” inProceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security (CCS), 2023, pp. 79–93

2023
[19]

Retrieval-Augmented Generation for Large Language Models: A Survey

Y . Gao, Y . Xiong, X. Gaoet al., “Retrieval-augmented generation for large language models: A survey,”arXiv preprint arXiv:2312.10997, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[20]

MemGPT: Towards LLMs as Operating Systems

C. Packer, S. Wooders, K. Linet al., “MemGPT: Towards LLMs as operating systems,”arXiv preprint arXiv:2310.08560, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[21]

MemoryBank: Enhancing Large Language Models with Long-Term Memory

W. Zhong, L. Guo, Q. Gaoet al., “MemoryBank: Enhancing large language models with long-term memory,”arXiv preprint arXiv:2305.10250, 2023

work page internal anchor Pith review arXiv 2023
[22]

LongMemEval: Benchmarking chat assistants on long-term interactive memory,

D. Wu, H. Wang, W. Yuet al., “LongMemEval: Benchmarking chat assistants on long-term interactive memory,”arXiv preprint, 2024

2024
[23]

MemBench: Towards more comprehen- sive evaluation on the memory of LLM-based agents,

H. Tan, Z. Zhang, C. Maet al., “MemBench: Towards more comprehen- sive evaluation on the memory of LLM-based agents,” inFindings of the Association for Computational Linguistics: ACL 2025. Vienna, Austria: Association for Computational Linguistics, 2025, pp. 19 336–19 352

2025
[24]

Discop: Provably secure steganogra- phy in practice based on

J. Ding, K. Chen, Y . Wanget al., “Discop: Provably secure steganogra- phy in practice based on ”distribution copies”,” in2023 IEEE Symposium on Security and Privacy (SP). San Francisco, CA, USA: IEEE, May 2023, pp. 2238–2255

2023
[25]

Provably secure disambiguating neural linguistic steganography,

Y . Qi, K. Chen, K. Zenget al., “Provably secure disambiguating neural linguistic steganography,”IEEE Transactions on Dependable and Secure Computing, 2024

2024
[26]

Provably robust and secure steganog- raphy in asymmetric resource scenario,

M. Bai, J. Yang, K. Panget al., “Provably robust and secure steganog- raphy in asymmetric resource scenario,” in2025 IEEE Symposium on Security and Privacy (SP). San Francisco, CA, USA: IEEE, May 2025, pp. 1438–1456

2025
[27]

Meteor: Cryptographically secure steganography for realistic distributions,

G. Kaptchuk, T. M. Jois, M. Greenet al., “Meteor: Cryptographically secure steganography for realistic distributions,” inProceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security. Virtual Event Republic of Korea: ACM, Nov. 2021, pp. 1529– 1548

2021
[28]

A framework for designing provably secure steganography,

G. Liao, J. Yang, W. Shaoet al., “A framework for designing provably secure steganography,” in34th USENIX Security Symposium (USENIX Security 25). USENIX Association, 2025, pp. 6837–6856

2025
[29]

On the structural memory of llm agents.arXiv preprint arXiv:2412.15266, 2024

R. Zeng, J. Fang, S. Liuet al., “On the structural memory of LLM agents,”arXiv preprint arXiv:2412.15266, 2024

work page arXiv 2024
[30]

A-MEM: Agentic Memory for LLM Agents

W. Xu, Z. Liang, K. Meiet al., “A-MEM: Agentic memory for LLM agents,”arXiv preprint arXiv:2502.12110, 2025

work page internal anchor Pith review arXiv 2025
[31]

ReAct: Synergizing reasoning and acting in language models,

S. Yao, J. Zhao, D. Yuet al., “ReAct: Synergizing reasoning and acting in language models,” inThe Eleventh International Conference on Learning Representations (ICLR 2023), 2023

2023
[32]

CAMEL: Communicative agents for

G. Li, H. Hammoud, H. Itaniet al., “CAMEL: Communicative agents for ”mind” exploration of large language model society,” inAdvances in Neural Information Processing Systems 36 (NeurIPS 2023), 2023

2023
[33]

Qwen technical report,

J. Bai, S. Bai, Y . Chuet al., “Qwen technical report,”arXiv preprint, 2023

2023
[34]

Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

Gemini Team, “Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities,”arXiv preprint arXiv:2507.06261, 2025. [Online]. Available: https://arxiv.org/abs/2507.06261

work page internal anchor Pith review Pith/arXiv arXiv 2025
[35]

A mathematical theory of communication,

C. E. Shannon, “A mathematical theory of communication,”The Bell system technical journal, vol. 27, no. 3, pp. 379–423, 1948

1948
[36]

Idiosyncrasies in large language models,

M. Sun, Y . Yin, Z. Xuet al., “Idiosyncrasies in large language models,” arXiv preprint arXiv:2502.12150, 2025

work page arXiv 2025