Research map / Large language models / Agents, code and math
Reflection-Bench: Evaluating Epistemic Agency in Large Language Models
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 Reflection-Bench, a parameterized, cognitive-psychology-inspired benchmark of seven tasks that evaluates LLM epistemic agency across prediction, decision-making, perception, memory, counterfactual thinking, belief updating, and meta-reflection.
Problem
Existing LLM evaluations examine isolated abilities or application-specific agent performance, not the base model's intrinsic epistemic agency in agent-environment interaction, and text-heavy benchmarks risk leakage.
Approach
Decomposes epistemic agency into seven dimensions and adapts validated cognitive psychology tests (weather prediction, WCST, oddball, n-back, double-choice Iowa gambling, probabilistic reversal learning, meta-bandit) with randomizable parameters to reduce contamination and allow difficulty scaling.
Key results
- Claude-3.5-Sonnet with CoT scored highest at 68.78 on entry-level settings, with o1-preview at 66.57 and DeepSeek-Reasoner at 64.56, yielding a three-tier hierarchy.
- No model exceeded chance-level thresholds across all five applicable tasks, and all models failed the meta-bandit task.
- On hard parameters, scores decreased for evaluated models; Centaur showed no improvement over its base Llama-3.1-70B-Instruct (overall 46.34 vs 55.17 on easy).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use free output instead of direct generation | LLMs evaluated on Reflection-Bench · Reflection-Bench tasks | overall performance | Helped |
| Use zero-shot CoT instead of direct generation | LLMs evaluated on Reflection-Bench · Reflection-Bench tasks | overall performance | Helped |
| Use zero-shot CoT prompting | LLMs evaluated on Reflection-Bench · Wisconsin Card Sorting Test and N-back | score | Helped |
| Use zero-shot CoT prompting | LLMs evaluated on Reflection-Bench · Double Choice Iowa Gambling Test | score | Hurt |
| Use zero-shot CoT prompting instead of free output | GPT-4o · Reflection-Bench Easy (average score) | score: 56.66 -> 68.06 | Helped |
| Use zero-shot CoT prompting instead of free output | GPT-4o-mini · Reflection-Bench Easy (average score) | score: 58.58 -> 52.64 | Hurt |
| Increase task difficulty from Easy to Hard | o1-preview (direct generation) · Reflection-Bench overall | score: 74.86 -> 49.66 | Hurt |
| Fine-tune on human cognitive test data | Llama-3.1-70B-Instruct · Reflection-Bench Easy (overall, direct generation) | score: 55.17 -> 46.34 | No clear effect |
| Recognize the predictable 2-step reversal pattern | LLMs evaluated on Reflection-Bench · Meta-bandit task | reward acquisition | Hurt |
| Use a fixed win-stay-lose-switch strategy rather than adapting to the global pattern | LLMs evaluated on Reflection-Bench · Meta-bandit task | reward acquisition | Hurt |
| Apply CoT prompting across tasks and models | LLMs evaluated on Reflection-Bench · Reflection-Bench tasks | score | Mixed |
Surprising findings
- Centaur, fine-tuned on human cognitive test data, did not outperform its base model Llama-3.1-70B-Instruct on Reflection-Bench.
- No model recognized the 2-step reversal pattern in the meta-bandit task, defaulting to win-stay-lose-switch strategies.
Limitations
- Ecological validity of adapted cognitive tests is uncertain.
- Focuses on text-only linguistic interaction, not multimodal or embodied settings.
- Evaluates base models only, not multi-LLM or agent architectures.
Experimental setup
Benchmarks: Reflection-Bench
Metrics: score (1-MAE/Max_MAE)*100 matching accuracy accuracy cumulative net earnings chance-level threshold (95th percentile of random simulation
Baselines: Claude-3.5-Haiku GPT-4o-mini Qwen-2.5-7B-Instruct Centaur Llama-3.1-70B-Instruct
Builds on: Iowa gambling task Wisconsin card sorting test probabilistic reversal learning chain-of-thought
Closest papers by meaning
- Position: Theory of Mind Benchmarks are Broken for Large Language Models (ICML 2025 · similarity 0.89)
- Eliciting Reasoning in Language Models with Cognitive Tools (NeurIPS 2025 · similarity 0.89)
- OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling (NeurIPS 2026 · similarity 0.89)
- AutoToM: Scaling Model-based Mental Inference via Automated Agent Modeling (NeurIPS 2025 · similarity 0.88)
- Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action (NeurIPS 2026 · similarity 0.88)
- LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical Study (AAAI 2025 · similarity 0.88)
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.