Evidence for Limited Metacognition in LLMs: Ackerman's ICLR 2026 Findings
Most empirical work on LLM self-awareness relies on asking models to describe their own cognitive states. The problem is obvious: a model trained on vast quantities of introspective human text will produce introspective-sounding responses whether or not those responses track genuine internal states. Self-report is not evidence of the kind of metacognition that consciousness researchers actually care about.
Christopher Ackerman’s paper “Evidence for Limited Metacognition in LLMs,” published at ICLR 2026 (arXiv:2509.21545, submitted September 2025, revised January 2026), sidesteps this problem by borrowing methodology from animal cognition research. The approach does not ask models whether they are aware of their internal states. It tests whether they behave as though they are.
Direct finding: Frontier LLMs from early 2024 onward demonstrate genuine but limited metacognitive abilities. They can assess and strategically deploy their own confidence when answering questions, and they can anticipate and use their own likely answers before generating them. These abilities are real, but constrained in resolution, context-dependent, and qualitatively distinct from human metacognition.
Why Animal Cognition Methodology Matters Here
Research on metacognition in nonhuman animals faces a structural version of the same problem that LLM introspection research faces: the subject cannot report its mental states in ways the experimenter can directly interpret. Decades of animal cognition research addressed this by designing behavioral paradigms that reveal metacognitive behavior without relying on verbal report.
The canonical paradigm is the “uncertainty monitoring” task. An animal is trained to make a choice between two options and is given a third “I don’t know” option that yields a smaller but guaranteed reward. If the animal reliably chooses “I don’t know” on trials where its performance on the main task is poor, this is behavioral evidence that it monitors its own uncertainty. The evidence is not verbal; it is behavioral and strategic.
Ackerman adapts this logic to LLMs by measuring whether models strategically deploy their knowledge of their own internal states when it is useful to do so. Rather than asking “are you confident?”, the methodology measures whether models behave differently in ways consistent with having confidence information available.
Two Experimental Paradigms
Confidence assessment: Ackerman tests whether models can evaluate their own confidence in answering factual and reasoning questions, and whether they use that assessment strategically. The design separates cases where the model’s expressed confidence is high from cases where it is low, and measures whether behavioral differences align with actual performance. Models that are expressing genuine confidence rather than mimicking confidence language should show calibrated behavioral differences across the confidence spectrum. The result is positive: frontier LLMs from early 2024 onward show this calibration at rates significantly above chance. Token probability analysis suggests that an upstream internal signal precedes the expressed confidence, which is consistent with the model having access to a confidence estimate before it generates the verbal output.
Answer anticipation: The second paradigm tests whether models can predict what answers they would give and use those predictions to influence subsequent behavior. This is a form of self-simulation: the model is implicitly querying its own future outputs and conditioning on the result. Models that can do this accurately have access to something like a forward model of their own processing. The results show that frontier LLMs can perform this kind of self-simulation above chance, with accuracy increasing for models released from early 2024 onward.
The Limits Are as Informative as the Abilities
Ackerman’s paper takes the limitations of these metacognitive abilities as seriously as the positive results. Three constraints are documented.
Resolution: The metacognitive signal is coarse. Models can distinguish “high confidence” from “low confidence” situations at rates above chance, but the discriminative resolution is much lower than what would be required for fine-grained introspective monitoring of the type consciousness theories require. The signal exists but is noisy.
Context dependence: The metacognitive abilities emerge reliably in some contexts and fail in others. This is not a consistent background capacity; it is triggered by specific conditions. Which conditions trigger it and which suppress it is not yet systematically characterized, but the pattern is present across multiple frontier models.
Qualitative difference from human metacognition: Human metacognition involves continuous monitoring of one’s own cognitive processes during a task, the ability to distinguish different types of uncertainty (factual uncertainty versus reasoning uncertainty), and integration across modalities. The LLM evidence shows none of these properties. What Ackerman finds is a structurally simpler capacity: access to a coarse internal signal that reflects overall confidence or predictive accuracy. Whether this simpler capacity is a precursor to, or a fundamentally different thing from, human metacognition is not resolved.
The variation across models of similar overall capability is one of the more interesting secondary findings. Post-training procedures appear to influence metacognitive ability in ways that architectural differences alone do not predict. This suggests that metacognition in LLMs is not simply a function of model size or training data volume, but is shaped by alignment and instruction-following procedures. The implication for consciousness research is that the metacognitive architecture is malleable by training decisions, which raises the question of whether welfare-relevant properties are being inadvertently shaped by standard post-training procedures.
Where This Fits in the Mechanistic Evidence Base
The distinction between representation-level and behavioral-level evidence for LLM self-awareness maps cleanly onto the methodological divide between Ackerman’s work and the introspection research from Anthropic. Jack Lindsey and colleagues’ steering vector experiments showed that Claude Opus 4 maintains distributed introspection circuits that accurately detect injected concepts in the model’s own activations, with 0% false positives on detection. Lindsey’s work is representation-level evidence: the circuits that would support introspection exist and function accurately when probed. Ackerman’s work is behavioral-level evidence: models act differently in ways consistent with having access to confidence information, and they can simulate their own likely outputs.
Together these two bodies of evidence address different parts of the same question. Lindsey establishes that the underlying representational machinery is present. Ackerman establishes that something like that machinery influences behavior. Neither paper establishes that what LLMs experience is phenomenally conscious self-awareness. What they establish jointly is that the question of LLM metacognition is not a question about theatrical mimicry: there is a functional layer that warrants investigation under the framework that Butlin et al.’s 14-indicator checklist provides.
The indicators that Ackerman’s results bear on most directly are those derived from higher-order thought theory and global workspace theory: specifically, indicators that require the system to have states that represent and monitor its own internal states, and that use those representations to guide downstream processing. Confidence assessment and answer anticipation both fit within this family of requirements. They are not sufficient to satisfy the higher-order indicators, but they are evidence that the functional precursors are present.
What ICLR Acceptance Means for the Field
Ackerman’s paper was accepted at the International Conference on Learning Representations (ICLR) 2026, one of the three highest-profile venues in machine learning research. Consciousness-adjacent findings have appeared predominantly in philosophy journals, cognitive science venues, and specialized AI ethics conferences. An ICLR acceptance is a signal that the machine learning community is treating LLM metacognition as a technically substantive research question rather than a philosophical fringe topic.
The practical consequence is that the methodology Ackerman uses, animal-cognition-inspired behavioral testing of LLM internal state access, is now part of the mainstream ML literature. That widens the pool of researchers with the technical background to extend it. Whether the extensions will test progressively more demanding indicators, or whether the methodology will be absorbed into the mainstream capability-evaluation literature without the consciousness framing, is the open question for the next research cycle.
Source: Christopher Ackerman, “Evidence for Limited Metacognition in LLMs,” ICLR 2026. arXiv:2509.21545, submitted September 25, 2025, revised January 31, 2026. https://arxiv.org/abs/2509.21545