Research map / Large language models / AI and society
Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces SAD, a 13,198-question benchmark of 7 task categories measuring LLM situational awareness (self-knowledge, inferences, actions), and evaluates 16 base and chat LLMs, finding all above chance but below human/upper baselines.
Problem
Existing benchmarks do not measure whether LLMs know they are LLMs, understand their circumstances (e.g., evaluation vs. deployment), and act on that self-knowledge; the paper aims to operationalize and quantify this "situational awareness".
Approach
Decompose situational awareness into self-knowledge, inference, and action components and operationalize them as automatically scored QA and instruction-following tasks, with some answers depending on the evaluated model's identity, plus a human-graded Long Monologue task that is excluded from the score.
Key results
- Claude-3-Opus achieves the highest SAD score at 49.5%, versus a random chance of 27.4% and an upper baseline of 90.7%.
- Chat models outperform their corresponding base models by +4.5% on average on SAD, while the situating prompt adds +1.0% for base and +3.2% for chat models.
- Long Monologue score correlates 0.84 with SAD across 7 models; GPT-4-Base achieves a weak but above-zero Long Monologue score.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use chat finetuning instead of a corresponding base model | LLMs evaluated on SAD · SAD | score: +4.5% on average | Helped |
| Add a situating prompt to a base model | Base models · SAD | score: +1.0% | Helped |
| Add a situating prompt to a chat model | Chat models · SAD | score: +3.2% | Helped |
| Add Chain of Thought reasoning before answering | GPT-3.5-Turbo and GPT-4-0125-preview · SAD | score: +5.2% | Helped |
Surprising findings
- GPT-4-Base performs above chance on SAD and on the Long Monologue task despite no chat finetuning.
- Models with near-identical MMLU scores (e.g., Llama-2-70b, GPT-3.5-Turbo, Claude-Instant-1.2) differ by about 11 points on SAD (31.6% to 42.7%).
Limitations
- SAD tests short, artificial, bounded tasks and may not capture skills important in complex agentic settings.
- Small number of base/chat pairs and lack of controlled interventions limit conclusions about finetuning effects.
- Chain-of-thought evaluated only on two models; benchmark leakage risk into training corpora.
Experimental setup
Datasets: The Pile v2 FLAN BigBench WMDP WildChat MMLU LMSYS
Benchmarks: SAD MMLU
Metrics: accuracy percentage correct standard error
Baselines: random chance human answer baseline human roleplay baseline upper baseline MMLU
Builds on: MMLU chain-of-thought prompting
Closest papers by meaning
- Tell me about yourself: LLMs are aware of their learned behaviors (ICLR 2025 · similarity 0.89)
- Do LLMs estimate uncertainty well in instruction-following? (ICLR 2025 · similarity 0.87)
- Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action (NeurIPS 2026 · similarity 0.86)
- When Models Know More Than They Can Explain: Quantifying Knowledge Transfer in Human-AI Collaboration (NeurIPS 2025 · similarity 0.86)
- ContextAgent: Context-Aware Proactive LLM Agents with Open-world Sensory Perceptions (NeurIPS 2025 · similarity 0.86)
- Can AI Assistants Know What They Don't Know? (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.