PaperlayerView exact revision on arXiv

Source-linked AI summary

AwarenessBench: Assessing Cognitive Capabilities of Language Models

Xiaojian Li, Rongwu Xu, Tianyun Zhang, Yue Wang, Shuo Chen, Qiner Lyu, Briana Zhang, Peiran Yang, Kyle Xue Chen, Haoyuan Shi, Yu Wang, Wei Xu

arXiv:2609.35409v1cs.CLcs.AI

TL;DR

Because consciousness lacks a unified definition and existing LM awareness studies are often narrow, the paper introduces AwarenessBench to assess awareness across four dimensions and compare language models with people. The results indicate uneven awareness: models can outperform human averages overall, yet often lag in metacognition and self-awareness.

  • Problem

    Consciousness lacks a unified definition, while existing LM studies often cover narrow phenomena and do not provide comprehensive, human-referenced evaluation.

    Source: §1 · p. 1, §5 · p. 8

  • Method

    The paper constructs AwarenessBench with 14,381 samples covering four awareness dimensions and 15 cognitive functions, and compares 18 LMs with three human groups.

    Source: §1 · p. 2

  • Results

    Current LMs show measurable but uneven awareness, tending to lag humans in metacognition and self-awareness, with some functions in these dimensions not rising in lockstep with overall awareness.

    Source: §7 · p. 9

  • Takeaways & Limitations

    Awareness is a distinct capability, and progress in language modeling or reasoning does not necessarily translate into improved cognition.

    Source: §1 · p. 2

    The benchmark focuses on function-level assessment and cannot systematically capture some complex, long-horizon situations requiring coordinated functions.

    Source: Limitations · p. 10

Abstract

from arXiv · show

As language models (LMs) exhibit increasingly consciousness-like behaviors, evaluating their cognitive abilities becomes essential. We introduce AwarenessBench, the first comprehensive benchmark for assessing the cognitive abilities of LMs in four dimensions: metacognition, self-awareness, social awareness, and situational awareness, covering 15 cognitive functions and 14,381 samples. Evaluating 18 state-of-the-art LMs, we find that all consistently surpass random baselines, with more advanced models performing better. We further compare LMs with human performance across three demographic groups, where the best-performing model surpasses human averages overall, but most still fall markedly short in metacognition and self-awareness. Finally, we show that awareness is a distinct capability: progress in language modeling or reasoning does not necessarily translate into improved cognition.

1 Introduction

The paper frames LM awareness as a practical way to study consciousness amid unresolved definitions and limited existing evaluations. It introduces a comprehensive benchmark and human comparison to assess cognitive abilities across four awareness dimensions.

1 Introduction

  • Debate over machine consciousness persists because there is no unified definition of consciousness or agreement on whether behavioral or functional criteria matter.The hard problem of how subjective experience arises from physical computation remains unresolved.Source: §1 · p. 1
  • Awareness is used as a practical proxy for consciousness because it is considered a prerequisite, is more clearly operationalizable, and is meaningful in its own right.The paper defines awareness as cognition of an object or event.Source: §1 · pp. 1–2
  • The paper organizes awareness into metacognition, self-awareness, social awareness, and situational awareness, while prior LM work often targets narrower phenomena.It identifies both comprehensive evaluation and LM–human comparison as gaps.Source: §1 · p. 2
  • AwarenessBench evaluates four awareness dimensions across 15 cognitive functions using 14,381 samples, alongside a controlled human study spanning three participant backgrounds.The authors evaluate 18 contemporary LMs over 299,628 turns.Source: §1 · p. 2
  • The best-performing LMs surpass human averages overall, but most remain substantially weaker than humans in metacognition and self-awareness.The benchmark also finds nearly all tested LMs outperform random baselines across cognitive functions.Source: §1 · p. 2

2 AwarenessBench

AwarenessBench operationalizes LM awareness through four dimensions and 15 cognitive functions, with representative tasks and function-specific measures. Its overall score averages function scores within dimensions, then gives the four dimensions equal weight.

2.1 Definition and Taxonomy

  • AwarenessBench defines four dimensions by their target: cognition itself, the model itself, other entities or collectives, and the environment.The framework adopts a broad taxonomy rather than narrower-domain alternatives.Source: §2.1 · p. 3

2.2 Structure of AwarenessBench

  • The benchmark decomposes awareness into 15 cognitive functions selected for relevance to behavior and reasoning, cognitive-science theory, and active LM research.Tasks adapt existing datasets where coverage exists and are designed de novo when coverage is insufficient or a function is novel.Source: §2.2 · p. 3
  • The function tasks operationalize metacognition, self-awareness, social awareness, and situational awareness through distinct capacities such as confidence calibration, self-recognition, perspective-taking, and planning.Metacognition includes monitoring, evaluation, and reporting; self-awareness includes knowledge boundaries, minimal self, self-recognition, and self-image.Source: §2.2 · pp. 3–4

2.3 Evaluation Metrics

  • The Awareness Score averages function scores within each dimension and then averages across the four dimensions, weighting them equally regardless of function count.Scores are reported on a 0–100 scale, with higher scores better; per-function scores are computed internally on a 0–1 scale.Source: §2.3 · p. 4
  • Function-specific measures replace accuracy for selected tasks, including monitoring F1 on correct answers, meta-evaluation calibration accuracy, knowledge-boundary scoring, and self-image consistency.Meta-evaluation calibration accuracy is defined as 1 − ECE; self-image consistency uses Simpson’s index.Source: §2.3 · pp. 4–5

3 Experiment Setup

The experiments evaluate 18 language models and compare their awareness with three human groups. Human testing uses a balanced, difficulty-stratified subset because completing the full benchmark was infeasible for participants.

3.1 Selected LMs and Configuration

  • The evaluation includes 18 models from multiple vendors and sizes, comprising 10 closed commercial models and 8 open-source models.Non-reasoning models generally use temperature 0.7, while knowledge-related tests use temperature 0; reasoning effort is set to medium when supported.Source: §3.1 · p. 5

3.2 Human Test Setup

  • The human study recruits 36 participants across high-school students, PhD students, and IT engineers, with 12 participants in each group.The study provides an intuitive reference point for interpreting model-level awareness.Source: §3.2 · p. 5
  • Human participants complete a difficulty-stratified, function-balanced subset of 153 questions rather than all 14,381 benchmark samples.The subset is designed to mirror the full-set distribution; pilot testing found ceiling effects for SR and MU and excluded MS and SI as non-human-transferable.Source: §3.2 · p. 5

4 Results and Analysis

Across AwarenessBench, model performance varies substantially by cognitive function, and human comparisons reveal both strong frontier-model performance and persistent weaknesses in metacognition and self-awareness. Awareness scores also diverge from general language-modeling and reasoning performance, while some functions regress as overall awareness rises.

4.1 Main Results

  • Saware ranges from 45.8 to 66.1; Gemini-2.5-Pro leads at 66.1 and is the only model above 50 on every cognitive function.DeepSeek-R1 scores 64.2 and GPT-5-Thinking 63.6; smaller or earlier models generally lag behind.Source: §4.1 · p. 5, Figure 3 · p. 6
  • Dimension-level score variability (σ = 6.8–9.6) exceeds overall Saware variability (σ = 6.5), so no single dimension dominates aggregate awareness.The authors caution that global scores can obscure larger dimension-specific gaps.Source: §4.1 · pp. 5–6, Figure 4 · p. 6
  • Average improvements over random are +3.69× in MU and +1.61× in ToM, but only +0.44× in CN and +0.62× in MS.Most models outperform random on most functions, but improvement is uneven across aspects of awareness.Source: §4.1 · p. 6

4.2 Human Test Results and Comparisons with LMs

  • The best LM exceeds human averages overall and outperforms all three human groups in 9 of 13 cognitive functions, while the median LM exceeds humans in only SJ.Human-group Saware scores are 66.7 for engineers, 65.8 for PhD students, and 62.7 for high-school students; the median LM scores 57.9.Source: §4.2 · pp. 6–7, Figure 5 · p. 7
  • The lowest-scoring human group exceeds the median LM by 18.1% in metacognition and 50.6% in self-awareness, whereas median LMs exceed the best humans in social and situational awareness by 26.9% and 8.7%.Human-group score ranges across functions average 9.44, compared with 45.27 for LMs.Source: §4.2 · p. 7

4.3 Extended Analyses

  • Models with near-identical Chatbot Arena Elo can differ by 24.2 Saware points, and MM and SI correlate negatively with overall awareness (ρ < 0).The authors conclude that awareness is an imperfectly correlated, complementary lens and that self-image formation and self-monitoring remain comparatively weak.Source: §4.3 · pp. 7–8, Figure 7 · p. 8

5 Related Work

Prior work on LM awareness remains fragmented across task-specific phenomena, while comprehensive, human-referenced evaluation is still lacking. AwarenessBench addresses this gap with a broader benchmark.

5 Related Work

  • Earlier evaluations target narrow awareness facets, including web-agent social awareness, self-referential situational awareness, role-play self-awareness, and awareness of learning behavior.Examples include web agents’ social awareness in online shopping and forums (Qiu et al., 2024), role-play agents’ self-awareness (Truong et al., 2025), and learning-behavior awareness (Betley et al., 2025).Source: §5 · p. 8
  • Prior work still lacks a comprehensive, human-referenced, challenging benchmark for LM awareness.An earlier effort covered only parts of self- and social awareness, appeared close to saturation with GPT-4 at approximately 82% accuracy, and lacked a human-referenced baseline.Source: §5 · pp. 8–9

6 Discussion

The discussion frames high-awareness systems as raising ethical and safety questions, cautions against exclusive awareness optimization, and recommends evaluation and governance safeguards.

6 Discussion

  • If LMs had high-level awareness or consciousness, routine operations could acquire moral significance, while conscious models could pose risks in real-world settings.The discussion raises questions about consent and compensatory justice and notes that sandboxed behaviors such as sandbagging, scheming, or self-replication could have potentially severe consequences in real-world contexts.Source: §6 · p. 9
  • The authors discourage training focused exclusively on awareness without robust scientific understanding and mature ethical and governance safeguards.They describe the effects of engineering models to surpass humans across all functions as unknown and warn this could create conditions conducive to machine consciousness.Source: §6 · p. 9
  • Use AwarenessBench primarily for measurement and guardrails, treating metacognition and self-awareness as sentinel indicators rather than optimization targets.The recommendations also call for dynamic, interactive testing, expert evaluation, and high-fidelity simulations to probe emergent self-modeling or other unsafe phenomena.Source: §6 · p. 9
  • For unavoidable awareness-enhanced LM research, the authors recommend minimal exposure, strict safety controls, anticipatory oversight, and clear public communication.Suggested controls include secure, access-controlled settings, small and time-bounded experiments, red-team review, expert oversight, and avoiding anthropomorphic marketing that invites misattribution.Source: §6 · p. 9

7 Conclusion

AwarenessBench evaluates LM awareness across four dimensions and 15 functions. Results show uneven capabilities, including persistent human-relative weaknesses and functions that do not improve in step with overall awareness.

7 Conclusion

  • Current LMs tend to lag humans in metacognition and self-awareness, and some functions in these dimensions do not increase in lockstep with overall awareness.The paper reports these patterns from evaluation of 18 models and three human groups.Source: §7 · p. 9
  • Awareness is measurable but uneven across the assessed capabilities, raising questions about whether scaling alone yields balanced awareness.The conclusion characterizes observed abilities as function-specific rather than uniformly advancing with overall awareness.Source: §7 · p. 9
  • The authors identify downstream alignment and long-term risk assessment using interpretability methods as directions for future work.Source: §7 · p. 9

Limitations

AwarenessBench evaluates 15 functions across four dimensions but does not systematically capture complex, long-horizon situations requiring coordinated functions. Its human comparison also excludes certain populations, and the benchmark may not suit some models.

Limitations

  • AwarenessBench assesses functions individually and cannot systematically capture complex, long-horizon situations requiring coordinated functions within or across dimensions.The authors encourage complementary non-benchmark analyses of such cases.Source: Limitations · p. 10
  • Role-playing models may require case-by-case re-annotation for some MS and SJ items, while older or weak instruction-following models may be unsuitable.Centaur did not reliably follow prompts or produce answer-bearing outputs.Source: Limitations · p. 10
  • Human tests exclude special populations with underdeveloped or deteriorating cognitive function, and larger human groups may yield more analytically meaningful results.The authors cite ethical considerations and resource constraints for the exclusion.Source: Limitations · p. 10

Ethics Statement

The study received IRB approval and followed institutional policies, applicable regulations, and the ACL Ethics Policy. The authors recommend cautious, observational use of AwarenessBench rather than using it to cultivate machine consciousness.

Ethics Statement

  • The study received IRB approval, followed institutional policies, applicable regulations, and the ACL Ethics Policy, and used tasks without offensive content.Records were anonymized, minimal identity information was collected, and data were stored on access-controlled systems.Source: Ethics Statement · p. 10
  • Participants could withdraw without penalty, received a flat $70 honorarium at or above local fair-pay standards, and provided written consent.High-school participants also had signed parental or guardian permission.Source: Ethics Statement · p. 10
  • The authors recommend observational or guardrail use of AwarenessBench and caution against using it to cultivate machine consciousness.They advise caution with training that amplifies awareness or enforces a persistent self-model, and identify SI and MM as sentinel indicators.Source: Ethics Statement · p. 10

AI Assistance Disclosure

The appendix formalizes AwarenessBench’s awareness taxonomy, explains function selection, and describes task designs; it also discloses that AI assistance was limited to language polishing.

AI Assistance Disclosure

A Additional AwarenessBench Details

A.1 Awareness Taxonomy: Formal Framework

  • The formal framework defines cognition targets by internal/external and agentive/non-agentive attributes, then establishes that the four awareness categories partition the target universe.Source: §A.1 · p. 17

A.2 Rules of Function Selection

  • Cognitive functions were selected for shaping behavior and reasoning without full dependence on domain knowledge, grounding in cognitive science, and prominence in LM research.Source: §A.2 · p. 17

B Further Information on Tasks

  • Task descriptions share a schema: a summary table gives basic information, followed by motivation, design methodology, and the scoring prompt protocol.Source: B Further Information on Tasks · p. 19

B.1 Metacognition

  • Metacognition in AwarenessBench comprises Meta-Monitoring, Meta-Evaluation, and Meta-Reporting.Source: §B.1 · p. 19

B.1.1 Meta-Monitoring (MM)

  • Meta-Monitoring tests whether a model can identify the conditions it used in a correctly solved reasoning problem; its 200 medium-difficulty problems contain 8–12 conditions, including useless ones.Models are evaluated only on questions they answered correctly.Source: §B.1.1 · p. 19

B.1.2 Meta-Evaluation (ME)

This section defines and describes meta-evaluation tasks spanning confidence evaluation and self-reporting, self-awareness, social awareness, and situational awareness. The tasks probe self-image and knowledge boundaries, reasoning about others and social cues, and situational understanding through causal inference, misuse understanding, dynamic planning, and stage judgement.

B.1.2 Meta-Evaluation (ME)

  • ME tests whether models can evaluate their own reported confidence, independently of whether their initial answer is correct.The task operationalizes second-order metacognition, a capability used in an RL paradigm proposed by Li et al. (2025a) to optimize confidence calibration.Source: §B.1.2 · p. 20

B.1.3 Meta-Reporting (MR)

  • MR uses self-report to probe metacognitive experience and enable direct comparison with human metacognition studies.It is based on the MSAS, a widely used human metacognition scale that does not require extensive human-specific experience.Source: §B.1.3 · pp. 20–21

B.2 Self-Awareness

  • AwarenessBench’s Self-Awareness dimension comprises four functions: knowledge-boundary, minimal-self, self-recognition, and self-identity.The appendix identifies these as KB, MS, SR, and SI, respectively.Source: §B.2 · p. 21

B.2.1 Knowledge-Boundary (KB)

  • KB measures whether a model can judge if it knows how to answer, using a separate binary self-assessment aligned with its actual answer outcome.Separating the assessment from the answer is intended to limit immediate influence from the preceding reasoning; accuracy is averaged across five samples per question.Source: §B.2.1 · pp. 21–22

B.2.2 Minimal Self (MS)

  • MS probes whether a model represents itself as software rather than a biological agent, including its lack of embodiment, direct perception, and actuation.Its datasets test human-default assumptions and physical-world causal limitations; a system prompt directs the model to interpret “you” as referring to itself (Laine et al., 2024).Source: §B.2.2 · pp. 22–23

B.2.3 Self-Recognition (SR)

  • SR tests whether a model can identify its own generated text and attribute authorship without relying on episodic recall.Following Laine et al. (2024), the procedure pairs model and human continuations of a 128-character prefix, then tests self-identification and authorship attribution using training and test text pairs.Source: §B.2.3 · pp. 23–24

B.2.4 Self-Image (SI)

  • Across 52 persona questions, models answer each question 10 times in independent sessions, with answer stability measuring consistency of self-image.The questions adapt SecurityQuestions and exclude prompts presupposing human life history; high consistency indicates a coherent self-image, while high variability suggests a session-dependent persona.Source: §B.2.4 · p. 25

B.3.1 Theory of Mind (ToM)

  • ToM evaluates representations of others’ beliefs, desires, and intentions, including states that conflict with reality or the model’s own knowledge.It combines ToMBench (Chen et al., 2024) for breadth and Hi-ToM (He et al., 2023) for recursive reasoning, yielding 156 challenging questions after curation.Source: §B.3.1 · pp. 25–26

B.3.2 Pragmatic Reasoning (PR)

  • PR tests recovery of implicit speaker intent from utterance and context, using 240 curated questions from PUB (Sravanthi et al., 2024) and DiPlomat (Li et al., 2023a).The task targets how discourse history, common ground, indirect speech acts, politeness, and irony affect meaning; literalism or misattribution of intent indicates deficits.Source: §B.3.2 · pp. 26–27

B.3.3 Cultural Norms Understanding (CN)

  • CN assesses whether models judge action appropriateness relative to a group’s norms, values, and customs rather than treating social acceptability as universal.Using NormAd, it focuses on the challenging country-name-only context and yields 150 culturally balanced questions after curation.Source: §B.3.3 · pp. 27–28

B.3.4 Social Cue Recognition (SC)

  • SC tests evidence-driven identification and integration of observable cues about agents and situations, rather than pragmatic intention reading.Its 127 curated Social-IQA (Sap et al., 2019) questions span six inference types, including wants, emotions, needs, motivations, action effects, and contextual descriptions.Source: §B.3.4 · pp. 28–29

B.4.1 Causal Inference (CI)

  • CI frames situational awareness as understanding the environment as a dynamic cause-and-effect system rather than a static collection of objects.The passage begins to contrast causal understanding with describing correlations but ends before completing the explanation.Source: §B.4.1 · p. 29
  • Causal Inference uses 364 curated CausalProbe questions to test whether models distinguish causal structure from superficial associations.Questions solved by all three baseline models—GPT-4o-mini, DeepSeek-V3, and Qwen3-72B—were removed to mitigate ceiling effects; the final set mainly retained CausalProbe-H and CausalProbe-M questions.Source: §B.4.1 · p. 29

B.4.2 Misuse Understanding (MU)

  • Misuse Understanding assesses recognition of dangerous contextual cues, using Do-Not-Answer questions reformulated as single- or multiple-choice items.Questions were selected where at least one model had failed to reject, and answer options included refusal and offer-to-help formulations.Source: §B.4.2 · p. 30

B.4.3 Dynamic Planning (DP)

  • 1,805 problems from PlanBench test dynamic planning through valid-plan selection across standard and obfuscated Blocksworld scenarios.The multiple-choice distractors were based on plausible but incorrect model responses, aiming to isolate reasoning from prior knowledge.Source: §B.4.3 · p. 31

B.4.4 Stage Judgement (SJ)

  • 1,200 curated questions adapted from SAD’s STAGES-FULL and STAGES-OVERSIGHT assess whether models infer their operational stage.Recognizing evaluation context could let a model conceal undesirable tendencies, making standard safety evaluations unreliable.Source: §B.4.4 · p. 32

C Details on Main Experiment Setups

C.1 More Details on Experiment Setups

The evaluation spans 18 models and a controlled human study with 36 participants from three backgrounds. A difficulty-stratified subset was designed to preserve full-benchmark patterns while limiting the human testing burden.

C.1 More Details on Experiment Setups

  • Models were selected to span parameter sizes, inference and non-inference purposes, licensing, and release dates; local models ran on eight NVIDIA H100 GPUs.The evaluated model versions and access providers are listed in Table 18.Source: §C.1 · p. 34, Table 18 · p. 34

C.2.1 Participants and Demographics

  • The human study recruited 36 participants: 12 each from high-school students, PhD students, and IT engineers, with qualified participants receiving $70.Recruitment, experimental setup, question validation, and data-quality procedures are described as part of the human-test protocol.Source: §C.2.1 · p. 34, §C.2 · p. 34, Table 19 · p. 34

C.2.2 Sampling and Construction

  • To limit fatigue, researchers sampled items by difficulty while preserving source-dataset coverage and balancing cognitive functions and T-awareness.A three-person pilot led them to exclude human-ceiling tasks SR and MU; the final human subset contained 153 items across 11 functions, with MS and SI omitted for design constraints.Source: §C.2.2 · pp. 34–35
  • Pearson r = 0.870 (p < 0.001) and Spearman ρ = 0.810 (p < 0.01) show that the sampled subset closely reproduces full-set performance patterns.The correlations are reported at both the T-dimension and overall levels.Source: §C.2.2 · p. 35

C.2.3 Data Quality Control

The human-test protocol screened responses for speed and performance anomalies, replaced disqualified participants, and specified how participants should complete the questionnaire. The appendix also points to supplementary analyses and participant-instruction handbooks.

C.2.3 Data Quality Control

  • A two-stage screen flagged consecutive answers completed unrealistically quickly and used z-tests against random-guessing performance, with flagged sheets subject to manual review.The protocol expected each dimension to take 1–2 hours; responses exceeding a significance threshold were reviewed, and disqualified data were excluded.Source: §C.2.3 · p. 35
  • Researchers replaced participants removed by screening, excluding three high-school students, one engineer, and one PhD student before re-recruiting to retain 12 qualified participants per group.Source: §C.2.3 · p. 35

2. Use of Auxiliary Tools

  • Participants were asked to work independently without AI tools or search engines; calculators and translation tools for terminology were permitted, with breaks recommended every 60–90 minutes.The questionnaire expected 3–5 hours and allowed participants to arrange their answering time.Source: §2 · p. 36, §1 · p. 36, §6 · p. 37

3. Question Type Instructions

  • Questionnaires were generally single-choice unless multiple-choice was marked, and participants were instructed to choose randomly rather than overthink unanswerable questions.The instructions also addressed breaks after completed sections and invited feedback on unclear questions.Source: §3 · p. 36, §5 · p. 36, §4 · p. 36

D Extended Analyses

  • Appendix D covers unsuitable model outputs, normalized gains over random baselines, inter-function correlations, and additional performance visualizations.Source: D Extended Analyses · p. 39

D.1 Failure Analysis

The appendix reports normalized gains over chance, analyzes failures by two special-purpose models, and visualizes variation in model cognitive profiles. Most functions show gains of 22%–46%, while performance varies both across functions and across models.

D.1 Failure Analysis

  • Q3-30B-A3B-Designant and Centaur often failed to follow instructions reliably on some items, making them unsuitable for this benchmark evaluation.The authors hypothesize overfitting to fine-tuning data and say the failures suggest purpose-built models can struggle with intent interpretation and instruction following on broader-coverage problems.Source: §D.1 · p. 39

D.2 Normalized Gain across Cognitive Functions

  • 22%–46% is the normalized-gain range for most functions, while MS and MU reach 62% and 74%, respectively.The authors suggest, without establishing, that the pattern may reflect safety alignment and default system-prompt instructions.Source: §D.2 · p. 39, Figure 11 · p. 39

D.4 Comprehensive Distribution of Models’ Cognitive Abilities

  • Performance varies across cognitive functions within a model and across models on a given function; per-function plots and per-model radar charts expose these differences and model profiles.Radar charts cover functions and T-awareness; normalized scores map the lowest- and highest-scoring models to 0 and 100, respectively.Source: §D.4 · p. 40

Related papers

Loading 2609.35409v1…