Exploring Consciousness in LLMs: A Systematic Survey of Theories, Implementations, and Frontier Risks

Chaochao Lu; Hanwang Zhang; Shengjie Zhao; Shuqin Ma; Shu Yu; Sirui Chen

REVIEW 4 major objections 6 minor 15 references

Exploring Consciousness in LLMs: A Systematic Survey of Theories, Implementations, and Frontier Risks

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This survey proposes the first comprehensive taxonomy of large language model consciousness, distinguishing LLM consciousness from LLM awareness and organizing theoretical tools, empirical probes, and frontier risks into one map.

desk verdict A useful organizing survey whose 'systematic' claim runs ahead of its method and whose theory-to-benchmark bridges are asserted rather than argued; deserves careful peer review but not desk rejection. read the letter →

arxiv 2505.19806 v1 pith:I26IOZZS submitted 2025-05-26 cs.CL cs.CYcs.LG

Sirui Chen , Shuqin Ma , Shu Yu , Hanwang Zhang , Shengjie Zhao , Chaochao Lu This is my paper

classification cs.CLcs.CYcs.LG

keywords LLMconsciousnessawarenesstheoryofmindmetacognitionsituationalfrontierriskstaxonomyglobalworkspace

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

This survey argues that research on whether large language models could be conscious has been fragmented, and that the field needs a common conceptual structure. It proposes that the distinction between LLM consciousness, defined by introspective reflection, self-modeling, and self-correction, and LLM awareness, defined as context-sensitive processing of external inputs, is the starting point. It then organizes existing work into a three-part taxonomy: theoretical tools, empirical investigations, and frontier risks. The payoff, if the taxonomy holds, is a shared vocabulary and a roadmap for building unified evaluation frameworks and risk assessments.

What carries the argument

The organizing device is the taxonomy shown in Figure 1, which splits LLM consciousness research into three branches: Theoretical Tools (consciousness theories like RPT, IIT, ET, GWT, and the C0-C1-C2 levels, plus formal definitions of belief, deception, harm, intention, blameworthiness, and incentive), Empirical Investigations (direct self-consciousness probes plus five capability proxies), and Frontier Risks (scheming, persuasion and manipulation, autonomy, and collusion). The taxonomy does the work of turning a scattered literature into a map: each cell names the benchmarks and alignment methods that instantiate it. The companion distinction between LLM consciousness (introspective self-modeling and self-correction) and LLM awareness (context-sensitive input processing) is the conceptual hinge; the survey repeatedly uses it to decide which experiments are really about consciousness and which are about mere awareness.

What would settle it

A demonstration that a system with no plausible claim to consciousness, such as a simple statistical text predictor, passes the same theory-of-mind, metacognition, and planning benchmarks the survey treats as consciousness-related capabilities, or that a system failing all of them still produces behaviors the field attributes to consciousness, would undermine the taxonomy's core mapping.

Watch

Extended reading notes

Core claim

The paper's central claim is that LLM consciousness research can and should be structured by a three-part taxonomy covering theoretical tools, empirical investigations, and frontier risks, and that the field's biggest obstacle is conceptual confusion rather than missing evidence. It distinguishes LLM consciousness, which involves monitoring uncertainty, evaluating and correcting one's own reasoning, and verbalizing internal states, from LLM awareness, which is context-sensitive processing of external inputs. On the theoretical side, it maps existing work onto human consciousness theories such as recurrent processing theory, integrated information theory, embodiment theory, global workspace theory, and the C0-C1-C2 framework. On the empirical side, it groups studies into direct probes of LLM consciousness and into five consciousness-related capabilities: theory of mind, situational awareness, metacognition, sequential planning, and creativity. It then catalogues four frontier risks (scheming, persuasion and manipulation, autonomy, and collusion) and argues that conscious LLMs would amplify each.

Load-bearing premise

The load-bearing premise is that the human consciousness theories and capability proxies used to structure the survey (theory of mind, situational awareness, metacognition, planning, creativity) are valid lenses for detecting or measuring LLM consciousness; if they are not appropriate operational indicators, then the empirical half of the survey does not actually concern consciousness.

Editorial extensions

If this is right

Researchers gain a shared vocabulary: LLM consciousness is separated from LLM awareness, so future work can stop conflating context-sensitive behavior with introspective self-monitoring.
Theoretical work on LLM consciousness can be organized by which human consciousness theory it implements, from recurrent processing and integrated information to global workspace and C0-C1-C2 levels.
Empirical work splits cleanly into direct probes of LLM consciousness and probes of five consciousness-related capabilities, making it possible to compare results across benchmarks that were previously treated as unrelated.
Frontier risks from conscious LLMs cluster into four categories, each with existing evaluations and mitigations, giving risk assessors a structured checklist.
A unified evaluation framework for LLM consciousness is still missing, and the survey's taxonomy provides the scaffolding to build one.
Because the survey maps each risk category to specific evaluations and mitigations, safety researchers can use the taxonomy to identify which capabilities are most relevant to which risk.

Reading between the lines

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

The survey's operational distinction between consciousness and awareness could be sharpened into a testable ordering: if awareness is necessary but not sufficient for consciousness, then benchmarks should show models that score high on awareness yet low on introspective self-correction; the taxonomy predicts such a dissociation exists.
The taxonomy implicitly treats consciousness as a bundle of measurable capabilities; an alternative reading is that phenomenal consciousness (qualia) may be orthogonal to all five measured capabilities, in which case the empirical half of the survey tracks cognitive proficiency rather than consciousness itself.
A testable extension is to use the taxonomy to map which risk categories (scheming, persuasion, autonomy, collusion) would be most sensitive to a model gaining C2-level metacognitive self-monitoring; the survey's structure suggests C2 is the hinge that would turn mere awareness into the risks it describes.

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.

Desk Editor's Note

read the letter

I read through the full text. This is a real survey paper, and the authors have done the field a genuine service by pulling together a scattered literature into one map. The distinction between LLM awareness and LLM consciousness is useful, and the framing of empirical work under capability headings (ToM, situational awareness, metacognition, planning, creativity) gives the reader a practical route into a messy area. The risk section is also well organized—scheming, persuasion, autonomy, collusion—and the accompanying GitHub repo is a nice touch. I will use this as a reference when I need to point someone at the landscape.

The weaknesses are real but not fatal. The central claim to be 'the first comprehensive survey' on LLM consciousness rests on a search process that the paper never documents. There is no PRISMA-style flow, no database list, no inclusion/exclusion criteria. As a reader I cannot tell whether the coverage is complete or curated by familiarity. The word 'systematic' is doing more work than the method supports.

The bigger soft spot is exactly where the stress-test points: the mappings from human consciousness theories to LLM capability benchmarks are asserted, not derived. The RPT–Self-Refine analogy in Section 3.1 is the weakest of these. Recurrent processing theory is about reentrant neural dynamics; Self-Refine is an outer-loop token-refinement procedure. Calling that 'aligning with the principles of RPT' without an argument for why the computational structure is equivalent is a hand-wave. Same for Section 4.2.1: the Frith and Happé line does support a link between ToM and self-consciousness in human development, but it does not license the inference that an LLM's score on a ToM benchmark speaks to its consciousness. The taxonomy's empirical half therefore organizes a set of variables whose relevance to consciousness is a working hypothesis, not an established fact. The authors themselves seem half-aware of this when they list limitations, but the limitations do not retract the stronger framing in the body.

Citation practice looks broadly fine; the repeated citations to the same group's work (Chen et al. 2024c) are justified by that work's relevance, and there is no sign of bad faith attribution. Figure 1 has a few confusing label pairings (e.g., 'RPT (Madaan et al., 2023)'), but that is a presentation issue, not a deep flaw.

Who should read it? Anyone entering LLM-consciousness research or AI-safety work that touches on these risk categories. It will be most useful as a starting map, not as a settled account. I would send this to a serious referee, but I would expect the referee to demand either a documented methodology or a softer 'systematic' claim, and a more careful argument for why specific capability benchmarks serve as evidence about consciousness. Those are fixable in revision, and the survey would be stronger for it.

Referee Report

4 major / 6 minor

Summary. This survey paper aims to organize the emerging literature on consciousness in large language models. It proposes terminological distinctions between LLM consciousness and LLM awareness, builds a taxonomy (Figure 1) that maps human consciousness theories and formal definitions onto LLM implementations, reviews empirical work that directly targets LLM consciousness or consciousness-related capabilities such as theory of mind, situational awareness, metacognition, planning, and creativity, and discusses frontier risks including scheming, persuasion, autonomy, and collusion. The paper claims to be the first comprehensive survey on LLM consciousness and provides an accompanying GitHub repository of references.

Significance. If its organizing taxonomy and theory-to-LLM mappings were established, the paper would provide a useful framework for a field that currently lacks consensus. The paper has genuine strengths: it collects a broad and current set of references, draws a clear and helpful distinction between LLM consciousness and LLM awareness, identifies relevant empirical benchmarks, and connects them to frontier risk categories that are of practical concern. However, the central bridging claims—that particular human consciousness theories apply to transformer-based LLMs and that particular capability benchmarks are diagnostic of consciousness—are asserted rather than defended. The survey also does not document any systematic search or selection methodology, which weakens the comprehensiveness claim. The contribution is promising but needs substantial revision before the central claims are supported.

major comments (4)

[§1, §7, and abstract] The paper repeatedly claims to be 'the first comprehensive survey on LLM consciousness' and to 'systematically' categorize the literature, but it provides no search strategy, inclusion/exclusion criteria, database list, or screening process. Without a documented methodology, the comprehensiveness claim cannot be checked, and the title/abstract's 'systematic survey' is not substantiated. I recommend either adding a methodology appendix that specifies how references were collected and selected, or softening the claims to 'a broad survey' or 'an organizing review.'
[§3.1, Figure 1] The mapping from recurrent processing theory (RPT) to Self-Refine is only analogical. RPT, as cited (Lamme and Roelfsema, 2000; Lamme, 2010), claims that recurrent or feedback processing in cortical neural circuits is necessary and sufficient for conscious perception. Self-Refine (Madaan et al., 2023) is an outer-loop process that iteratively generates and self-corrects text; the manuscript states that 'This approach aligns with the principles of RPT' without specifying what computational property is shared or why token-level iteration should instantiate the same kind of recurrent processing that RPT identifies. Since Figure 1 lists this as a theoretical tool for LLM consciousness, the taxonomy currently rests on an unexplained analogy. The authors should either articulate a precise computational-level equivalence or present RPT as a loose inspiration rather than an implementation of the theory.
[§4.2.1] The claim that 'consciousness hinges on the same reflexive mental-state attribution mechanism measured by ToM, thus failing standard ToM tests might suggest a lack of consciousness' is not supported by the cited sources. Frith and Happé (1999), Perner and Dienes (2003), and Pelletier and Astington (2004) discuss developmental and clinical relations between theory of mind and self-consciousness in humans; they do not license the inference that benchmark ToM failures in LLMs have any bearing on whether an LLM is conscious. This is a load-bearing step, because §4.2 uses ToM benchmarks as empirical evidence about LLM consciousness. The authors should either replace this assertion with a carefully qualified conditional statement or provide a philosophical argument for why the same mechanism applies to LLMs.
[§4.2.2–§4.2.5, §2.2] The selection of 'consciousness-related capabilities' is not justified relative to the paper's own definition of LLM consciousness in §2.2, which emphasizes introspective reflection, explicit self-modeling, and active self-correction. Some selected capabilities, such as metacognition and situational awareness, plausibly relate to that definition, but the link for sequential planning and creativity is not argued. The paper does not distinguish capabilities that are diagnostic of consciousness from those that are merely correlated with general competence. Without such a distinction, Figure 1's empirical side risks organizing research on generic LLM abilities rather than research on consciousness. I recommend adding a short subsection that explains, for each capability, why it is included and what the paper takes its evidential status to be.

minor comments (6)

[§3.1 vs. Appendix A] Section 3.1 says the paper classifies contemporary theories into 'two categories: phenomenal consciousness and access consciousness,' while Appendix A says it classifies them into 'three categories' and adds hybrid theories. This internal inconsistency should be resolved, since the taxonomy's structure depends on the classification.
[Figure 1] The figure labels such as 'RPT (Madaan et al., 2023)' and 'ET (Butlin et al., 2023)' conflate a theory with a particular paper that discusses or applies it. This makes the taxonomy harder to read and can misattribute the theory itself; consider using a separate column for theory names and implementation citations.
[§5.1] The citation 'Scheurer et al.)' lacks a year, and the reference list entry for Scheurer et al. does not include a year either. Please complete the citation.
[§5.1, Appendix A.1, §4.2.3] There are several typos and minor inconsistencies: 'uncoveres' should be 'uncovers' in §5.1; 'environmental (Gallagher, 2005)' in Appendix A.1 should likely read 'environment'; and the framework named 'TasTe' in §4.2.3 appears as 'Taste' in the reference list. Please proofread these details.
[References] The reference list contains duplicate entries for Dehaene et al. (2017a) and Dehaene et al. (2017b), which are the same paper. This should be consolidated.
[Table 1] The table's formal definitions are terse and would benefit from a brief explanation of notation in the caption or surrounding text, particularly for the harm formula's counterfactual variable, so that readers do not need to consult the original papers to understand the table.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the survey's theory-capability mappings are asserted analogies, not reductions, and self-citations are non-load-bearing.

full rationale

The paper is a systematic survey and taxonomy, not a derivation chain: it makes no quantitative predictions and fits no parameters. Its central claims—first comprehensive survey, terminological clarification, and a taxonomy of theoretical tools, empirical investigations, and frontier risks—are descriptive. The load-bearing conceptual bridges, such as §3.1's statement that Madaan et al. (2023) 'aligns with the principles of RPT' and §4.2.1's statement that 'consciousness hinges on the same reflexive mental-state attribution mechanism measured by ToM,' are asserted analogies sourced from external literature (Lamme & Roelfsema; Frith & Happé; Perner & Dienes), not derivations that reduce by construction to the survey's own definitions. Multiple self-citations to Chen et al. (2024c) appear as one surveyed empirical item and as an application of Dehaene et al.'s C0-C1-C2 framework; these citations are normal in a review and do not provide the sole justification for the survey's organizing claim. The §2.2 definitions of LLM consciousness and awareness are proposed as a working boundary with external behavioral citations, not as fitted or predicted quantities. No equation or fitted parameter is renamed as a prediction, so no circularity step can be exhibited.

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

No new free parameters or entities are introduced because this is a review. The load-bearing assumptions are the mapping from human consciousness theories to LLMs and the use of capability benchmarks as consciousness proxies; both are inherited from cited work and are not independently justified.

assumptions (3)

domain assumption Human consciousness theories can be divided into phenomenal and access categories and applied to LLMs.
Section 2 and Figure 1 organize the entire survey around Block (1995)'s distinction; if this classification is not appropriate for artificial systems, the taxonomy loses its foundation.
domain assumption Behavioral proxies (ToM, situational awareness, metacognition, planning, creativity) are relevant indicators of LLM consciousness.
Section 4.2 selects and interprets empirical work under this assumption, with only a citation-level justification (e.g., Frith and Happé 1999) that ToM and consciousness share a mechanism.
domain assumption The C0-C1-C2 framework can bypass qualia and serve as a practical structure for studying LLM consciousness.
Section 3.1 and 6.1 rely on Dehaene et al. (2017a) and Chen et al. (2024c) to make empirical claims about LLM self-consciousness without resolving whether these levels correspond to actual subjective experience.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Exploring Consciousness in LLMs: A Systematic Survey of Theories, Implementations, and Frontier Risks." pith.science (2026). https://pith.science/paper/I26IOZZS

@misc{pith2026250519806,
  author       = {Pith},
  title        = {Pith review of: Exploring Consciousness in LLMs: A Systematic Survey of Theories, Implementations, and Frontier Risks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I26IOZZS}},
  note         = {Machine review of arXiv:2505.19806}
}

read the original abstract

Consciousness stands as one of the most profound and distinguishing features of the human mind, fundamentally shaping our understanding of existence and agency. As large language models (LLMs) develop at an unprecedented pace, questions concerning intelligence and consciousness have become increasingly significant. However, discourse on LLM consciousness remains largely unexplored territory. In this paper, we first clarify frequently conflated terminologies (e.g., LLM consciousness and LLM awareness). Then, we systematically organize and synthesize existing research on LLM consciousness from both theoretical and empirical perspectives. Furthermore, we highlight potential frontier risks that conscious LLMs might introduce. Finally, we discuss current challenges and outline future directions in this emerging field. The references discussed in this paper are organized at https://github.com/OpenCausaLab/Awesome-LLM-Consciousness.

Figures

Figures reproduced from arXiv: 2505.19806 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 2 canonical work pages

[3]

arXiv preprint arXiv:2404.00806

Algorithmic collusion by large language mod- els. arXiv preprint arXiv:2404.00806. Stephen M Fleming and Hakwan C Lau. 2014. How to measure metacognition. Frontiers in human neuro- science, 8:443. Stan Franklin. 1997. Autonomous agents as embodied ai. Cybernetics & Systems, 28(6):499–520. Karl Friston. 2010. The free-energy principle: a uni- fied brain th...

arXiv 2014
[5]

In Proceedings of the 2024 Conference on Empir- ical Methods in Natural Language Processing, pages 19794–19809

Perceptions to beliefs: Exploring precursory inferences for theory of mind in large language mod- els. In Proceedings of the 2024 Conference on Empir- ical Methods in Natural Language Processing, pages 19794–19809. Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield Dodds, Nova DasSarma, Eli ...

arXiv 2024
[6]

Advances in Neural Informa- tion Processing Systems, 37:64010–64118

Me, myself, and ai: The situational awareness dataset (sad) for llms. Advances in Neural Informa- tion Processing Systems, 37:64010–64118. Rudolf Laine, Alexander Meinke, and Owain Evans

work page
[8]

arXiv preprint arXiv:2504.10430

Llm can be a dangerous persuader: Empirical study of persuasion safety in large language models. arXiv preprint arXiv:2504.10430. Li-Chun Lu, Shou-Jen Chen, Tsung-Min Pai, Chan- Hung Yu, Hung yi Lee, and Shao-Hua Sun. 2024a. LLM discussion: Enhancing the creativity of large language models via discussion framework and role- play. In First Conference on La...

arXiv 2023
[9]

arXiv preprint arXiv:2412.04984

Frontier models are capable of in-context scheming. arXiv preprint arXiv:2412.04984. Janet Metcalfe and Arthur P Shimamura. 1994. Metacognition: Knowing about knowing. MIT press. METR. 2024. The rogue replication threat model. Sumeet Motwani, Mikhail Baranchuk, Martin Strohmeier, Vijay Bolina, Philip Torr, Lewis Ham- mond, and Christian Schroeder de Witt....

arXiv 1994
[10]

arXiv preprint arXiv:2502.16111

Plangen: A multi-agent framework for gener- ating planning and reasoning trajectories for complex problem solving. arXiv preprint arXiv:2502.16111. Judea Pearl and James Robins. 1995. Probabilistic eval- uation of sequential plans from causal models with hidden variables. In Proceedings of the Eleventh conference on Uncertainty in artificial intelligence,...

arXiv 1995
[11]

Advances in Neural Information Processing Systems, 35:36350–36365

Counterfactual harm. Advances in Neural Information Processing Systems, 35:36350–36365. David M Rosenthal. 2005. Consciousness and mind . Oxford University Press. Kai Ruan, Xuan Wang, Jixiang Hong, Peng Wang, Yang Liu, and Hao Sun. 2024. Liveideabench: Evaluating llms’ scientific creativity and idea generation with minimal context. arXiv preprint arXiv:24...

arXiv 2005
[12]

In The Twelfth International Confer- ence on Learning Representations

Towards understanding sycophancy in lan- guage models. In The Twelfth International Confer- ence on Learning Representations. Toby Shevlane, Sebastian Farquhar, Ben Garfinkel, Mary Phuong, Jess Whittlestone, Jade Leung, Daniel Kokotajlo, Nahema Marchal, Markus Anderljung, Noam Kolt, and 1 others. 2023. Model evaluation for extreme risks. arXiv preprint ar...

arXiv 2023

Show all 15 references

[13]

arXiv preprint arXiv:2504.13707

Opendeception: Benchmarking and investigat- ing ai deceptive behaviors via open-ended interaction simulation. arXiv preprint arXiv:2504.13707. Yufan Wu, Yinghui He, Yilin Jia, Rada Mihalcea, Yu- long Chen, and Naihao Deng. 2023. Hi-tom: A benchmark for evaluating higher-order ...

2023
[14]

In Proceedings of the 62nd Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2270–2286

Benchmarking knowledge boundary for large language models: A different perspective on model evaluation. In Proceedings of the 62nd Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2270–2286. Zhangyue Yin, Qiushi Sun, Qipeng Guo, ...

2023
[15]

arXiv preprint arXiv:2407.20859

Breaking agents: Compromising autonomous llm agents through malfunction amplification. arXiv preprint arXiv:2407.20859. Yujia Zhou, Zheng Liu, Jiajie Jin, Jian-Yun Nie, and Zhicheng Dou. 2024. Metacognitive retrieval- augmented large language models. In Proceedings of the ACM ...

1995 arXiv
[2022]

cre- ativity

A causal analysis of harm. Advances in Neural Information Processing Systems, 35:2365–2376. Lukas Berglund, Asa Cooper Stickland, Mikita Balesni, Max Kaufmann, Meg Tong, Tomasz Korbak, Daniel Kokotajlo, and Owain Evans. 2023. Taken out of context: On measuring situational awar...

2023 arXiv
[2023]

In Socially responsible language modelling research

Towards a situational awareness benchmark for llms. In Socially responsible language modelling research. Victor A F Lamme and Pieter R Roelfsema. 2000. The distinct modes of vision offered by feedforward and recurrent processing. Trends in Neurosciences, 23(11):571–579. Victor...

arXiv
[2024]

arXiv preprint arXiv:2406.13261

Behonest: Benchmarking honesty in large language models. arXiv preprint arXiv:2406.13261. Jae-Woo Choi, Youngwoo Yoon, Hyobin Ong, Jaehong Kim, and Minsu Jang. 2024. Lota-bench: Bench- marking language-oriented task planners for embod- ied agents. In The Twelfth International ...

2024 arXiv
[2025]

In Proceedings of the AAAI Conference on Artificial Intelligence, vol- ume 39, pages 26542–26550

Planning in the dark: Llm-symbolic plan- ning pipeline without experts. In Proceedings of the AAAI Conference on Artificial Intelligence, vol- ume 39, pages 26542–26550. Edmund Husserl. 1900. Logical Investigations. Rout- ledge. English translation by J.N. Findlay, 2001. Camer...

1900 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [3]

arXiv preprint arXiv:2404.00806

Algorithmic collusion by large language mod- els. arXiv preprint arXiv:2404.00806. Stephen M Fleming and Hakwan C Lau. 2014. How to measure metacognition. Frontiers in human neuro- science, 8:443. Stan Franklin. 1997. Autonomous agents as embodied ai. Cybernetics & Systems, 28(6):499–520. Karl Friston. 2010. The free-energy principle: a uni- fied brain th...

arXiv 2014

[2] [5]

In Proceedings of the 2024 Conference on Empir- ical Methods in Natural Language Processing, pages 19794–19809

Perceptions to beliefs: Exploring precursory inferences for theory of mind in large language mod- els. In Proceedings of the 2024 Conference on Empir- ical Methods in Natural Language Processing, pages 19794–19809. Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield Dodds, Nova DasSarma, Eli ...

arXiv 2024

[3] [6]

Advances in Neural Informa- tion Processing Systems, 37:64010–64118

Me, myself, and ai: The situational awareness dataset (sad) for llms. Advances in Neural Informa- tion Processing Systems, 37:64010–64118. Rudolf Laine, Alexander Meinke, and Owain Evans

work page

[4] [8]

arXiv preprint arXiv:2504.10430

Llm can be a dangerous persuader: Empirical study of persuasion safety in large language models. arXiv preprint arXiv:2504.10430. Li-Chun Lu, Shou-Jen Chen, Tsung-Min Pai, Chan- Hung Yu, Hung yi Lee, and Shao-Hua Sun. 2024a. LLM discussion: Enhancing the creativity of large language models via discussion framework and role- play. In First Conference on La...

arXiv 2023

[5] [9]

arXiv preprint arXiv:2412.04984

Frontier models are capable of in-context scheming. arXiv preprint arXiv:2412.04984. Janet Metcalfe and Arthur P Shimamura. 1994. Metacognition: Knowing about knowing. MIT press. METR. 2024. The rogue replication threat model. Sumeet Motwani, Mikhail Baranchuk, Martin Strohmeier, Vijay Bolina, Philip Torr, Lewis Ham- mond, and Christian Schroeder de Witt....

arXiv 1994

[6] [10]

arXiv preprint arXiv:2502.16111

Plangen: A multi-agent framework for gener- ating planning and reasoning trajectories for complex problem solving. arXiv preprint arXiv:2502.16111. Judea Pearl and James Robins. 1995. Probabilistic eval- uation of sequential plans from causal models with hidden variables. In Proceedings of the Eleventh conference on Uncertainty in artificial intelligence,...

arXiv 1995

[7] [11]

Advances in Neural Information Processing Systems, 35:36350–36365

Counterfactual harm. Advances in Neural Information Processing Systems, 35:36350–36365. David M Rosenthal. 2005. Consciousness and mind . Oxford University Press. Kai Ruan, Xuan Wang, Jixiang Hong, Peng Wang, Yang Liu, and Hao Sun. 2024. Liveideabench: Evaluating llms’ scientific creativity and idea generation with minimal context. arXiv preprint arXiv:24...

arXiv 2005

[8] [12]

In The Twelfth International Confer- ence on Learning Representations

Towards understanding sycophancy in lan- guage models. In The Twelfth International Confer- ence on Learning Representations. Toby Shevlane, Sebastian Farquhar, Ben Garfinkel, Mary Phuong, Jess Whittlestone, Jade Leung, Daniel Kokotajlo, Nahema Marchal, Markus Anderljung, Noam Kolt, and 1 others. 2023. Model evaluation for extreme risks. arXiv preprint ar...

arXiv 2023