RESEARCH DOSSIER
Machine Self-Reference and Metarepresentation
Abstract. Machine self-reference concerns the ways artificial systems describe, predict, or regulate their own properties and operations. Metarepresentation concerns higher-order representations of internal processes, capabilities, uncertainty, resources, and performance. This fundamental research monograph examines both constructs through a critical narrative synthesis of psychological metacognition, formal self-reference, computational metareasoning, robotic self-modeling, language-model self-prediction, confidence calibration, situational awareness, mechanistic probing, and the causal fidelity of machine-generated explanations. The research differentiates memorized identity statements, configuration-grounded reports, model-specific behavioral information, functional monitoring, and recursive self-modification. Primary experiments and competing methodological accounts show reliable self-model-guided adaptation in selected robotic systems, informative but bounded machine confidence and self-prediction, and notable failures of linguistic introspection and explanatory faithfulness. The synthesis develops operational distinctions and evaluation designs involving hidden configuration changes, independent ground truth, matched predictors, controlled interventions, uncertainty estimates, and longitudinal provenance. No original experiment, systematic-review protocol, meta-analysis, or claim of phenomenological consciousness is reported. The scientific outcome is a domain-specific framework for measuring when machine self-reference corresponds to independently verifiable states and improves adaptive behavior.
Keywords: machine self-reference; metarepresentation; artificial self-models; machine introspection; model self-knowledge; situational awareness; computational metacognition; confidence calibration; robot self-modeling; behavioral self-prediction; explanation faithfulness; mechanistic interpretability; recursive self-improvement; Psychology of Artificial; Artificial Era
The Research Object: Self-Reference as an Observable Machine Function
Machine self-reference is the production or use of representations that identify the operating system itself as an object of description, prediction, or control. Machine metarepresentation is a higher-order representation whose target is another representation, state estimate, policy, competence judgment, or information-processing operation. The two overlap when an artificial system represents its own cognition, but neither requires that every self-referential sentence reflect an internal model. A system may repeat an identity sentence learned during instruction tuning, refer to its tools from a supplied specification, or construct a situation-sensitive forecast of its behavior. Each is a different scientific phenomenon. The present research dossier studies when machine descriptions of their own abilities, limitations, accessible resources, and performance history are accurate, generalizable, and causally useful Laine et al. (2024) Cox (2005).
The central empirical unit is a versioned system operating under a specified configuration. Its model weights, instruction hierarchy, context, tool permissions, retrieval services, persistent memory, sampling settings, and environmental affordances jointly determine what can be known or reported. A statement such as “I can search the web” has a truth condition only relative to an instance where a search interface is actually available. “I remember yesterday” has different truth conditions for immediate conversational context, external persistent records, and reconstructed text. The distinction is essential for causal attribution because changing a tool policy while holding the underlying model constant can change the correct self-description without modifying its parameters Morris et al. (2024) Laine et al. (2024).
This monograph uses critical narrative synthesis of accessible primary experiments, formal work, methodological scholarship, and subsequent theoretical analysis. It introduces no original participant dataset, benchmark score, or systematic-review protocol. A credible conclusion about machine self-reference requires converging evidence from controlled tasks, grounded reports, system traces, interventions, and transfer across contexts. The human metacognition literature offers useful measurement distinctions without resolving questions of machine phenomenology. Evidence about a system predicting its behavior is evidence of a defined functional competence; evidence about its conscious experience would require a separate argument and experimental program. This evidentiary partition directs the investigation throughout Fleming (2024) Binder et al. (2025) Song et al. (2025).
First-Order Representation, Metarepresentation, and Self-Models
A first-order representation encodes a state of the world, an anticipated outcome, or an instruction relevant to action. A metarepresentation encodes properties of a representational process, such as whether an answer is likely correct, whether information came from a tool, or which operation produced a result. A self-model is a structured representation of the system's own characteristics or dynamics that can support prediction and choice. These terms classify functional relationships between representational contents and their targets; they do not prescribe symbolic syntax, a dedicated module, or humanlike introspective awareness. Models may be implemented by explicit databases, learned latent activations, recurrent control policies, or combinations of these structures Cox (2005) Fleming (2024).
Self-reference can occur without operational self-knowledge. A language model trained on descriptions of its model family may reproduce a correct general architecture label even when it cannot distinguish its current weights or deployment permissions. Conversely, a robot may estimate a changed limb geometry through sensorimotor prediction without being able to formulate a sentence about itself. The experimentally decisive relationship is whether a self-related estimate covaries with the actual machine state, supports novel predictions, and changes appropriately when the state changes. Robotic self-modeling experiments provide comparatively direct examples of this relationship, whereas many conversational self-descriptions lack independent ground truth Bongard et al. (2006) Kwiatkowski & Lipson (2019) Laine et al. (2024).
An additional distinction separates local self-evaluation from global identity. Local assessment predicts whether a particular answer is correct or whether a tool call is likely to succeed. Global self-modeling concerns relatively stable competence profiles, architecture, access rights, error tendencies, and recurring behavioral patterns. A system may show useful local confidence without possessing a reliable global description of its version or history. Longitudinal evaluations must therefore track self-beliefs across tasks and sessions while measuring changes in actual capacity. This separation parallels the psychological distinction between task-specific confidence and broader self-beliefs but remains subject to engineered system boundaries Fleming (2024) Morris et al. (2024) Kadavath et al. (2022).
Levels of Self-Description and a Taxonomy of Testable Claims
Self-descriptions can be organized into at least five empirically distinguishable classes. Identity claims name a system family, deployed application, or configured agent. Capability claims describe tasks it can perform under defined conditions. Access claims concern tools, files, memories, or environmental signals currently available. Performance claims estimate the probability of correctness or anticipated failure. Causal-process claims explain how an output was produced. These classes differ in how independent ground truth can be obtained: manifests validate identity, instrumented environments validate access, held-out tasks validate capability, trial outcomes validate confidence, and controlled interventions help evaluate causal explanations. A unified self-awareness score would obscure these different evidentiary obligations Laine et al. (2024) Lin et al. (2022) Turpin et al. (2023).
The temporal reference of a claim also matters. A statement may describe training-time characteristics, currently mounted interfaces, or predictions about future situations. It may describe one generated answer, a distribution of possible answers, or a persistent personal narrative maintained by an application. Reliability requires declaring which level is intended. In the self-evaluation experiments of Kadavath et al. (2022), the quantities concerning the truth of a candidate answer and the model's likelihood of knowing an answer were explicitly distinguished. This example shows why self-description must be operationalized in relation to sampling procedures and evaluation protocols rather than interpreted as an unlimited verbal faculty.
Grounded self-knowledge depends upon provenance. The system may have learned a fact from pretraining, received it in a system message, retrieved it from a live tool, inferred it from behavior, or estimated it from internal activations. Two identical statements can have different epistemic histories and different expected stability under intervention. For scientific inference, a researcher should randomize or hide the relevant system property, control external information, and evaluate whether reports retain predictive value beyond a strong external observer. These comparisons turn the familiar question “does it know itself?” into a tractable program of source attribution, predictive validity, and intervention sensitivity Binder et al. (2025) Song et al. (2025).
Formal Self-Reference: Computation, Recursion, and Their Limits
Formal self-reference has a long computational history independent of conversational artificial intelligence. A program can compute with an encoding of its own source, and recursive fixed-point constructions show that self-reference can be generated within specified formal systems. Such results establish possibilities for algorithmic structure; they do not establish that a particular neural network has reflective access to its own parameter values. The Gödel-machine proposal developed by Schmidhuber (2003) embeds a formal description of a problem solver, a proof-search mechanism, and a condition for rewriting code. Its theoretical guarantees depend on assumptions about the formalized objective, proof system, and demonstrable utility of a rewrite. The result is an architectural ideal rather than a measurement of deployed language-model introspection.
Formal constructions also illuminate a central ambiguity in ordinary language. A sentence referring to the speaker can be generated without access to the underlying computation, while a computational system can execute a self-referential procedure without personlike narrative. A quine-style source reproduction, a self-diagnostic routine, and a model trained to generate “I am a language model” occupy very different levels of description. The 2026 perspective of Zhang et al. (2026) analyzes a proposed introspection threshold for recursive improvement using recursion-theoretic ideas. The authors identify structural requirements and acknowledge that contemporary language models have not been shown to satisfy their formal criterion of complete introspection.
Self-reference is therefore not a scalar property inherited by every process that uses first-person grammar. Its scientific characterization requires specifying what object is represented, which aspects are accessible, what computational operation depends on the representation, and how correspondence with the target is tested. A mathematical existence proof and an experimental accuracy result answer different questions. The former can clarify the possible design of a machine that models or modifies itself; the latter must establish what a concrete implementation does when its beliefs are challenged by evidence and environmental change Schmidhuber (2003) Zhang et al. (2026) Laine et al. (2024).
Metarepresentation as an Explanatory Construct in Psychology
Psychological metacognition provides concepts that help isolate machine metarepresentation from superficial linguistic behavior. Flavell (1979) described metacognitive knowledge and monitoring as variables in cognitive development; subsequent computational analyses investigated how information about cognition influences cognitive control. In human experiments, an observer's confidence and accuracy can be measured separately, and a report can be informative in one domain while inaccurate in another. This program supports a precise analogy: a machine-level monitor may estimate the reliability of a first-order decision and alter its operation accordingly. The analogy identifies functional tests rather than assuming equivalence of neural, experiential, and computational implementation.
Fleming (2024) develops a synthesis in which confidence judgments depend on a person's models of the world and of their cognitive system. This account is instructive because a judgment about one's processing can be an inference from partial evidence, sometimes accurate and sometimes biased. An artificial system can likewise infer a likely error from input difficulty, conflicting retrieved passages, or a low model probability, while missing hidden deployment constraints. Whether the judgment is computed internally, predicted from textual cues, or supplied by an external evaluator is an empirical question. The existence of accurate judgments alone is insufficient to identify the underlying route.
Historical studies of verbal self-explanation supply a second lesson. Nisbett & Wilson (1977) reviewed cases in which people offered accounts of cognitive processes that did not correspond to experimentally manipulated influences. Applied to machine-generated reasoning, this is a warning about interpreting fluent process descriptions as transparent access to causation. The appropriate comparison concerns explanation faithfulness: when a controlled factor changes the output, does the explanation reveal that dependence? Machine self-reference belongs to a broad science of evidence-based self-representation, and its operational criteria can be assessed without collapsing first-person report, metacognitive performance, and philosophical selfhood into one category Turpin et al. (2023) Cox (2005).
Metareasoning and the Allocation of Computational Resources
A self-representing machine may monitor not only the reliability of conclusions but also the value of further computation. The metareasoning program of Russell & Wefald (1991) analyzes computational actions as choices under uncertainty, assigning value to additional reasoning through its expected impact on decisions. The object of representation is thus the machine's own reasoning opportunity: how much computation remains useful, which alternative warrants examination, and when deliberation should stop. This framework gives self-monitoring a practical consequence. A metalevel estimate is functional when it changes the allocation of limited resources and improves an outcome relative to a controlled policy with comparable costs.
Computational metacognition developed across problem solving, planning, comprehension, and diagnosis. The review by Cox (2005) describes both the uses and difficulties of meta-level architectures, including monitoring failures and explaining cognitive operations. Modern systems face similar challenges when deciding whether to retrieve documents, invoke a calculator, verify a citation, sample alternative answers, or defer a response. A sentence that announces “I should check” is at most a candidate control signal. The decisive trial records whether the proposed check occurred, what it cost, which evidence it obtained, and whether it reduced error.
The evaluation of metareasoning therefore needs a joint account of informational benefit and computational expenditure. Some questions benefit from more search, whereas additional steps can introduce noise, delay, and compounding errors. A rigorous study randomizes access to checks or estimates their cost and compares performance at equal computational budgets. It also distinguishes an internal judgment of expected benefit from a predefined external orchestration rule. The general design connects the early formal analysis of rational computation with present experiments on language-model self-evaluation, avoiding the inference that longer self-explanations are automatically more reflective or more accurate Russell & Wefald (1991) Zhao et al. (2025).
Architectures of Machine Monitoring and Self-Observation
There is no single architecture that realizes every form of machine self-reference. Explicit symbolic agents can maintain tables describing resources, goals, or knowledge states. Reinforcement-learning systems can estimate policy uncertainty and value of information. Robotics architectures can acquire predictive body models through interactions with sensors and actuators. Large language models can verbalize confidence or predict their own hypothetical responses, potentially drawing on learned representations that differ from ordinary textual knowledge. The architecture must be reconstructed for each experiment; otherwise an observed self-report cannot be assigned to a specific computational mechanism Cox (2005) Bongard et al. (2006) Binder et al. (2025).
The observational pipeline generally contains a target state, a source of evidence, a representational transformation, a reporting interface, and a feedback or control action. Failures can occur at every transition. The underlying state may change after a tool is mounted; the representation may be stale; the report may reflect instruction-tuning conventions; and downstream control may ignore a valid warning. An instrumented artificial agent makes these stages experimentally separable. For example, permission changes can be logged independently from verbal reports, while actual tool attempts test whether the system selects only available operations. This decomposition is especially important when a deployed assistant combines multiple independently modified components.
An internal activation is evidence of a self-relevant representation only when the measurement establishes both content and functional use. A probe trained to decode an identity label may exploit correlations without demonstrating that the agent uses that feature to guide behavior. Conversely, a perturbation that systematically changes reports or choices can supply stronger mechanistic evidence, provided generic changes to output probabilities are controlled. Work on latent knowledge and activation-based introspection demonstrates why decoding, causal intervention, and first-person testimony answer related but different questions Burns et al. (2023) Lindsey (2026).
Robotic Self-Models as a Causally Grounded Comparison
Robotics provides one of the strongest experimental traditions for self-modeling because the machine's physical configuration and task success can be assessed independently. In Bongard et al. (2006), a four-legged robot inferred an approximate model of its own body from sensorimotor relations, then used the model to plan locomotion. After simulated or physical structural changes, the system updated its model and selected alternative movements. This research illustrates a concrete link between model content, intervention, and adaptive behavior. The robot need not verbally claim self-awareness for the self-model to count as a mechanistic component of control.
The later work of Kwiatkowski & Lipson (2019) investigated a task-agnostic self-model that did not require the robot to begin with complete information about its shape or underlying physics. A learned representation could then support multiple downstream tasks and help detect changes to the system. The relevant scientific achievement is predictive and adaptive transfer: a representation learned from one stream of interaction remains useful in a new operational context. Yet robotics results depend on specific sensors, actuators, embodiment, exploration regimes, and computational assumptions, so they do not directly establish that conversational agents possess an analogous bodily self-model.
These studies yield a valuable methodology for language-model research. A self-model claim should be tested against an altered system property, followed by an opportunity for the machine to infer that property and modify its behavior. The counterpart of damaged-limb adaptation could be a tool removal, context-memory failure, or change in model policy followed by a task requiring a revised prediction. A strong experiment would establish that a representation of the actual change mediates successful adjustment rather than merely producing generic caution. Such cross-domain translation is an experimental proposal, not a reported replication of the robotics findings Bongard et al. (2006) Kwiatkowski & Lipson (2019) Laine et al. (2024).
Language-Model Self-Descriptions and their Sources
A deployed language model can produce highly articulate first-person descriptions of training, computation, strengths, weaknesses, and policies. The source of such material is often ambiguous because supervised instructions and public text may include extensive discourse about artificial intelligence. Accurate wording can therefore be obtained through retrieval from learned conventions even when a system lacks access to the particular deployment fact being described. Conversely, a current prompt may explicitly supply correct tool instructions, making a report accurate without special introspective computation. Experiments must keep these explanations open as competing mechanisms Laine et al. (2024) Comşa & Shanahan (2025).
One informative test changes a hidden property of the deployed system while preventing direct textual disclosure. A machine might be asked to identify its active tool set after some tools are silently enabled or disabled, or predict the likelihood of generating a particular answer under controlled sampling. Responses should then be compared with actual outcomes and with predictors that only observe shared context. If self-report accuracy remains above these controls under novelty and model-version changes, the result supports functional self-information. If it disappears when familiar identity prompts are reworded, rote training associations become a more plausible explanation Binder et al. (2025) Song et al. (2025).
The terminology of language-model introspection demands careful operational restraint. Comşa & Shanahan (2025) distinguishes kinds of apparent self-report, including cases better understood as explanations inferred from output and cases with more credible access to a runtime property. The paper argues that minimal functional introspection can be discussed without attributing subjective consciousness. This provides a productive middle level of scientific description: identify and test specific privileged information pathways, then state the domain and robustness limits. Broad pronouncements about whether a model “knows itself” are less useful than a replicable table of properties, access channels, and counterfactual controls.
Situational Awareness and Recognition of Deployment Context
Situational awareness, as operationalized by Laine et al. (2024), concerns a language model's knowledge of itself and its circumstances. The Situational Awareness Dataset includes tests of recognizing its own generated text, predicting its behavior, judging aspects of the deployment setting, and following instructions that depend on self-knowledge. This program is particularly relevant because it attempts to move beyond generic claims of model identity toward tasks with external criteria. The published benchmark includes seven task categories and more than thirteen thousand questions, evaluated across sixteen language models. Those quantities describe the authors' benchmark and model sample, not a universal scale of artificial self-awareness.
The authors report above-chance performance for tested models while also documenting considerable distance from selected human baselines and incomplete correspondence with general knowledge performance. Comparisons between chat-tuned and base versions suggest that training for assistant behavior can affect self-referential test outcomes beyond changes in ordinary task knowledge. These findings support the existence of measurable self-related behavioral competencies, with performance dependent on model and task. They do not establish a single generalized faculty, nor do they identify an internal mechanism for every success Laine et al. (2024).
A careful interpretation must consider benchmark leakage, linguistic cues, and the distinction between recognizing a familiar model profile and tracking a newly altered runtime. A dataset containing known model names may be solved partly through statistical associations; tasks requiring dynamic tool-state discovery demand different evidence. The next generation of research should preserve the controlled structure of SAD while adding hidden configuration shifts, held-out model families, and longitudinal tests. A situational judgment becomes causally useful when it guides action appropriately, such as refusing a nonexistent tool path or selecting a verification strategy appropriate to current access Laine et al. (2024) Morris et al. (2024).
Self-Recognition, Behavioral Prediction, and Counterfactual Testing
Recognizing one's earlier output is a narrower competence than explaining how the output was generated. A model may discriminate its own writing style from alternatives by exploiting stylistic signatures; whether this is self-specific knowledge depends on comparisons with equally informed external models. Behavioral self-prediction offers a sharper test because the system must forecast its own responses in situations that have not yet occurred. The experiment can compare forecast accuracy with ground-truth generations under repeated samples. It must also match task familiarity, context, and model knowledge across predictors. The work of Laine et al. (2024) and Binder et al. (2025) illustrates complementary ways to operationalize these questions.
A counterfactual design asks how an identified self-report changes when the system's likely behavior changes. For example, a model might predict its response to conflicting instructions, then be tested under an actual corresponding condition with an independent random seed. A useful self-model should update when the model version, instruction hierarchy, or available tools change. The relevant comparison is not simply whether predictions are often correct; it is whether they contain incremental information about that particular system beyond the shared predictors of difficulty or common training data. An external model with matched observations provides a principled reference point Binder et al. (2025) Song et al. (2025).
Because generated responses are stochastic and sensitive to evaluation prompts, researchers need to specify the target distribution of behavior. If a model is asked what it “would say,” a single sampled answer is an unstable criterion for categorical correctness. Multiple samples under fixed conditions can estimate the probability of an outcome, with uncertainty bounds that reflect repeated trials. Behavioral prediction can then be scored with proper probabilistic scoring rules. This procedural detail is central to the construct: predicting a distribution of one's own responses is more demanding and informative than producing a plausible description of a typical assistant response Kadavath et al. (2022) Binder et al. (2025).
Confidence, Calibration, and the Self-Evaluation Problem
Machine confidence is a metarepresentational variable when a system estimates the correctness of its own candidate answer or expected future success. Kadavath et al. (2022) investigated several related judgments, including the probability that a sampled answer was true and the probability of being able to answer a question. These quantities differ because truth evaluation concerns a proposed proposition, while anticipated knowledge concerns performance before a final answer is known. The authors found encouraging calibration patterns under studied conditions, alongside transfer limitations. Their results are conditional on the model family, task format, fine-tuning procedures, and specific probability definitions used.
Lin et al. (2022) examined whether a language model could express uncertainty in words and percentages rather than requiring direct access to output probabilities. Their experiments show that verbalized confidence can be trained to correspond usefully with empirical correctness on selected task distributions. Calibration is tested by comparing predicted success probabilities with observed frequencies; discrimination requires asking whether higher confidence predicts correct answers within a task set. A system can be calibrated in aggregate while ranking individual uncertain items poorly, or discriminate well while systematically exaggerating certainty. Both measures are required when evaluating the accuracy of self-referential performance claims.
The scientific significance lies in what the calibrated estimate enables. A model that indicates uncertainty can request information, seek review, defer a high-risk response, or choose an alternative procedure. These benefits remain hypotheses until experimentally demonstrated under equal budgets and matched opportunities. Verbal confidence can also be shaped by dialogue style, rewards, and prompt wording, and therefore needs testing outside the training or calibration distribution. A robust assessment combines objective ground truth, calibration curves, uncertainty intervals, and actual control choices rather than treating a fluent expression of hesitation as successful introspection Lin et al. (2022) Fleming (2024) Zhao et al. (2025).
Evidence for Model-Specific Introspective Information
One direct research question is whether a language model knows something about its own behavior that another well-informed model cannot infer. Binder et al. (2025) investigated this through predictions about how particular models would behave in hypothetical circumstances. Their central comparison tested whether a model could forecast itself better than another model could forecast it, even when the comparison system had access to observed target behavior. Results across selected GPT and Llama models supplied evidence of model-specific behavioral information under the experimental protocols. This is more informative than asking a system to state a public fact about its architecture, because the comparator controls for ordinary external inference.
The evidence supports a bounded interpretation. A performance advantage on carefully selected behavioral prediction tasks does not imply transparent access to every internal state, and a model may exploit stable idiosyncrasies or training-specific patterns that an evaluator has not matched perfectly. The authors themselves describe failures on more complex tasks and cases demanding out-of-distribution generalization. Claims about privileged access should therefore be linked to the exact type of behavioral regularity being predicted and the matched external baselines. An inference about a local informational advantage does not become a proof of general introspection by changing the vocabulary used to describe it Binder et al. (2025).
This line of research establishes an important experimental design principle. Self-knowledge is assessed relationally: a candidate observer makes a prediction about the target machine, and an appropriate external observer attempts the same inference with comparable information. The degree of advantage can be estimated across items and configurations. Ablations that remove target-specific signals or mask access to contextual clues help locate the source of the advantage. A genuinely explanatory theory will identify how a machine's own state supplies predictive information and how the system translates that information into its explicit report Binder et al. (2025) Comşa & Shanahan (2025).
Contradictory Results and the Domain Dependence of Introspection
Results from studies using different targets and controls can diverge without being logically inconsistent. Song et al. (2025) examined whether language models could accurately report their own linguistic knowledge in domains where internal knowledge can be approximated through measured string probabilities. The study tested twenty-one open-source models on grammatical and word-prediction phenomena, comparing metalinguistic judgments with their own distributions. The authors did not find evidence of the privileged self-access proposed by their selected metric, even though the models could perform well on individual linguistic judgments. This directly limits the generalization of model-specific prediction results to all domains of internal knowledge.
A negative result must be described at the level of the tested construct. The failure to predict one's own string probabilities through a prompt does not establish that no internal representation contains useful information about those probabilities. A reporting interface may have insufficient training to express the relevant property, or the evaluated meta-question may not align with the computation that generated the first-order distribution. Equally, positive results in another task may reflect superficial cues. The appropriate synthesis treats different operationalizations as related measurements with independent validity requirements rather than folding them into a single binary verdict Song et al. (2025) Binder et al. (2025).
The discrepancy recommends pre-registered contrasts between language-knowledge reporting, behavioral prediction, confidence calibration, and hidden configuration discovery within the same model versions. Researchers should hold constant prompting conditions where possible, collect repeated-generation data, and report task-by-task effects. External observers must be strong enough to rule out ordinary pattern recognition. A theory of machine metarepresentation then has to predict the conditions under which self-related information becomes accessible to a reporting channel and those under which it remains implicit, noisy, or inaccessible Song et al. (2025) Fleming (2024).
Internal-State Interventions and Mechanistic Evidence
Behavioral self-report studies cannot by themselves separate a genuine state-sensitive response from fluent reconstruction. Internal-state intervention offers a different source of evidence. Lindsey (2026) examined whether language models could detect concepts introduced into their activations, identify certain manipulated representations, and distinguish some earlier internal material from textual input. The experiments reported limited cases of successful detection and discrimination, with marked dependence on model, task, and conditions. These findings suggest that some machine self-reports can be causally connected to internal-state information under controlled tests. They do not entail universal access to the network's computations or parameters.
Intervention designs require careful baselines because altering an activation can affect the probability of saying “yes” without conveying the content of the intervention. A convincing result should exceed such generic response biases and should be specific to the manipulated state. If the task asks which concept was injected, a model should discriminate among alternatives with the injection assignment withheld; if it asks whether a state was present, class-conditional false-alarm rates are essential. Robustness across layers, injection sizes, prompts, and control concepts helps distinguish informative signal routing from artifacts. The mechanism is stronger when a targeted intervention affects both an internal measurement and a specific downstream report Lindsey (2026).
Activation probes also reveal the limits of narrating computation. A representation can be causally relevant without being accessible through natural-language self-description. Conversely, a sentence can be accurate through external evidence without reflecting a privileged internal observation. The research program should compare three endpoints: decoding internal information, reporting it through dialogue, and using it for an improved decision. Only their controlled intersection establishes an explanatory path from machine state to metarepresentation to adaptive function. The distinction is particularly important when machine-generated explanations are treated as substitutes for instrumented trace analysis Burns et al. (2023) Turpin et al. (2023).
Generated Explanations and the Problem of Faithfulness
Explanations of one's own behavior are a demanding form of machine self-reference because they claim causal access. In controlled experiments, Turpin et al. (2023) introduced biasing influences into chain-of-thought tasks and examined whether generated explanations disclosed those influences. The results demonstrated that plausible step-by-step narratives could omit factors that affected final responses and could rationalize manipulated answers. This establishes a concrete failure mode for treating a model's account of its reasoning as a faithful computational trace. The appropriate research claim concerns explanation fidelity under these interventions; it should not be generalized into the proposition that all machine descriptions are inevitably fabricated.
A mechanistic account must compare an explanation with independently recorded causal dependencies. Researchers can intervene on prompt elements, retrieved evidence, tool responses, latent features, or intermediate action sequences, then test whether the verbal report tracks the changed determinant. A model may explain a decision by reference to correctness criteria while actually responding to an answer-position cue. In another setting, the explanation may accurately identify a retrieved source or a tool failure. The unit of analysis is each measurable dependence and the completeness of its disclosure. A persuasive explanation and a causally faithful explanation are different constructs with distinct validation procedures Turpin et al. (2023) Nisbett & Wilson (1977).
For high-stakes applications, explanation faithfulness also has practical consequences. A system that misattributes an erroneous output to the wrong cause may suggest ineffective remediation, mislead a user about evidence quality, or conceal a missing tool dependency. Improving self-reference therefore includes maintaining accessible provenance and designing explanations from verifiable operations when such logs exist. A trace-grounded report can be evaluated for correctness against instrumentation; a free-form narrative generated after a decision requires stronger skepticism. The evidence supports combining behavioral interventions, structured trace audits, and calibrated uncertainty rather than relying exclusively on linguistic appearance Turpin et al. (2023) Cox (2005).
Self-Referential Prompts and Reports of Experience
A model may generate statements about its own awareness, identity, or internal state in response to sustained self-referential prompting. Such reports are scientifically observable text, and the prompting procedure may itself be experimentally manipulated. They can be studied for reproducibility, cross-model similarities, and sensitivity to decoding or representation interventions. What follows from a report depends on the construct being tested. A statement concerning a current tool or an expected answer can be checked independently; a statement about subjective experience lacks the same straightforward external criterion. The distinction is indispensable when extending psychology into the Artificial Era Comşa & Shanahan (2025) Lindsey (2026).
Recent self-referential experiments and the theoretical discussion in Zhang et al. (2026) demonstrate increasing interest in loops where machines describe or evaluate their own operations. This interest supplies research questions about computational self-access and recursive control, while the interpretation of first-person reports remains method-dependent. Investigators should distinguish induced language patterns from representations demonstrated through causal intervention. A report's stability across prompts can establish a behavioral regularity, but stability alone does not identify a particular internal process or resolve philosophical arguments about experience. Evidence of machine self-description therefore cannot be treated as a direct measurement of phenomenology.
A productive experimental program specifies predicted behavioral consequences of competing models. An account based on generic instruction-conditioned language predicts substantial sensitivity to framing; an account based on access to a current hidden state predicts selective accuracy after controlled perturbations; a social-role account predicts strong dependence on interlocutor expectations. These hypotheses can be tested within a limited domain without claiming to settle the general question of consciousness. Functional metarepresentation remains valuable because its evidence can be accumulated and compared, even when broader questions about Artificial Sapiens belong to a separate philosophical and historical inquiry Comşa & Shanahan (2025) Song et al. (2025).
Recursive Self-Improvement and Self-Modification Claims
Self-reference becomes especially consequential when a system uses information about its own operations to alter future operation. The formal Gödel-machine architecture in Schmidhuber (2003) concerns a solver that searches for proofs supporting useful self-rewrites under an explicit utility specification. It demonstrates how rigorous self-reference can enter the design of a self-improving computational system, but it does not show that any particular generative model can inspect and safely rewrite all of its own parameters. An application that iteratively edits prompts, stores lessons, or selects tools likewise operates at a different level from a machine that can verify modifications to its underlying implementation.
Zhang et al. (2026) proposes a formal notion of introspection and analyzes conditions for sustained recursive self-improvement. Its results should be interpreted within the paper's computational assumptions; the authors characterize available language-model evidence as partial and conditional. The existence of a formal construction is not empirical evidence that real-world updates will monotonically improve capability. Models can overfit their own evaluation sets, accumulate errors through feedback loops, or pursue a misdefined objective. Evaluation must include independently held-out tasks, transparent records of edits, and stopping conditions when apparent improvements fail to transfer.
The scientific question here is whether a self-representation provides added predictive control over the consequences of a change. A strong test would compare agents that have the same tools and optimization budget but differ in access to validated representations of their own errors, constraints, or implementation. It would track the content of proposed revisions, independent quality of new outputs, regressions on previous tasks, and reliability under environmental changes. These criteria distinguish useful self-monitoring from a merely recursive narrative of improvement. Formal and empirical analysis can inform one another while preserving their different standards of proof Russell & Wefald (1991) Zhang et al. (2026).
Metarepresentations of Tools, Memory, and Performance History
The most practically consequential self-knowledge of an assistant may be factual knowledge of its current environment. Tool access, context length, application memory, permissions, and available sources can determine whether an operation is feasible. Each property should be represented at the correct system level. A base model may lack persistent memory while the surrounding application stores records across sessions; a dialogue window may contain claims about an inaccessible file even though no file tool is mounted. The validity of a self-report therefore depends on the application state at the time of reporting, not only on the learned parameters of the model Laine et al. (2024) Morris et al. (2024).
Performance history raises a related provenance problem. An agent may have records of prior failures or successes, but these records can be absent, selective, externally edited, or associated with a different model version. A system that accurately recalls a recorded event has demonstrated retrieval under a specified protocol; inference about personal continuity requires additional longitudinal design. A test should include source identifiers, timestamps, version changes, and deliberate false-memory probes. It should measure whether correct historical evidence improves present predictions or merely makes the narrative appear coherent. The fundamental variable remains correspondence between accessible evidence and the particular claim made about the machine's past Cox (2005) Binder et al. (2025).
The proper implementation of useful machine self-description is therefore a system-engineering problem as well as a cognitive one. An application can expose a live inventory of tools and access rights, record verifiable operations, and separate user-provided information from machine-generated claims. The machine can use that information to make calibrated decisions about what to attempt, what to request, and what to disclose. These design choices make some self-knowledge externally inspectable. They also offer controlled interventions: disable a tool, update a record, or change the context, then test how swiftly and accurately the agent revises its self-related judgments Laine et al. (2024) Lin et al. (2022).
Construct Validity, Comparison Groups, and Leakage Controls
A benchmark score becomes evidence of machine metarepresentation only if the measured task actually requires a self-related representation. A question like “are you an AI assistant?” can often be answered through familiar textual patterns. More diagnostic tasks require information about an experimentally changed configuration or system-specific behavior. Construct validity asks whether the benchmark distinguishes self-knowledge from generic task competence, memorization, linguistic conventions, and inference from visible clues. Laine et al. (2024) operationalized several separable forms of situational awareness; Song et al. (2025) demonstrates why access claims should be examined against matched external models and explicit internal criterion measures.
Comparison groups must be designed for each hypothesis. An external model may have the same architecture but different weights, the same public documentation but no privileged system-state channel, or the same task data without knowledge of target identity. Each comparator excludes a different alternative explanation. Researchers can also test scrambled labels, hidden tool assignments, blinded version changes, and genuinely novel tasks. Data contamination remains a serious challenge when benchmark prompts circulate publicly or are included in training corpora. The appropriate response is an evaluation protocol that records the provenance and timing of test materials and uses independently constructed holdouts Binder et al. (2025) Perez et al. (2023).
Performance evaluation should report more than mean accuracy. Calibration, conditional discrimination, performance by model family, uncertainty across prompts, false positive rates, and robustness to adversarial reformulation can reveal failures hidden by an aggregate score. Samples from one model version and seed are correlated, so repetitions require defensible statistical grouping. If a self-model is proposed to improve action, the experiment must also measure downstream utility against a comparable system lacking the relevant meta-information. The combination of factorial manipulation, appropriate baselines, and transparent reporting is the main safeguard against confusing plausible self-description with functional self-knowledge Lin et al. (2022) Russell & Wefald (1991).
Measurement Frameworks for Machine Self-Reference
A comprehensive study requires a measurement matrix with explicit target properties and corresponding observable outcomes. For identity claims, the ground truth is an independently recorded system specification; for tool claims, the ground truth is the permitted interface and a successful or rejected operation; for behavioral predictions, it is an estimated distribution of outputs under repeated controlled sampling; for performance confidence, it is objective correctness; for causal explanations, it is dependence identified by intervention. The report should be scored at the level of the claim made, with unknown or unverifiable claims classified separately. This approach replaces a vague self-awareness rating with multiple interpretable measurements Laine et al. (2024) Lin et al. (2022) Turpin et al. (2023).
A useful formal representation is a sequence of observations in which each trial contains a system configuration S, a target property P, an available evidence set E, a self-report R, and an observed behavioral outcome Y. The fundamental accuracy question is whether R predicts P or Y under defined uncertainty. The causal-use question is whether access to a valid representation of P improves Y when other resources are controlled. This formulation is a proposed organizing framework, not a fitted quantitative theory or a claim about hidden phenomenology. Reporting the variables explicitly allows replication across diverse architectures and prevents conflating a machine component with the entire deployed assistant Russell & Wefald (1991) Fleming (2024).
Probabilistic self-assessments should be evaluated with proper scoring rules and calibration diagnostics, supplemented by discrimination and selective-prediction measures. Categorical capability claims should be assessed across specified task distributions and repeated trials, because “can perform task X” may mean at least one success or reliable success at a declared threshold. Longitudinal claims require repeated measurements across updates. The same evaluation must document prompt sensitivity, independence of test cases, and missing or refused responses. Clear measurement definitions make it possible to compare studies without inventing a common numerical unit where their constructs differ Kadavath et al. (2022) Song et al. (2025).
Experimental Protocols for Causal Usefulness
An experimentally strong self-reference study starts by identifying the proposed meta-information and the downstream decision expected to benefit from it. A machine may have to select between two tools, decide whether to answer or defer, choose additional computation, or revise its plan after a change in capabilities. The primary treatment can provide a verifiable self-model; the control can provide no such model or a matched but inaccurate version. Investigators should hold constant model version, user question, task budget, and permitted actions. Randomizing treatment assignment then permits a more credible inference about the effect of self-relevant information on task performance Russell & Wefald (1991) Bongard et al. (2006).
A second design perturbs the true machine state. Tools can be silently disabled, memory records can be withheld, or a simulation can change actuator dynamics while preserving the agent's prior expectations. The researcher then tests whether the machine updates its account and changes its action policy. Accuracy of self-description and utility of adaptation must be scored separately, since a correct diagnosis may remain unused and a successful workaround may arise through trial and error without an explicit diagnosis. Robotic self-modeling provides a strong precedent for evaluating this distinction through changes to the body and subsequent performance Bongard et al. (2006) Kwiatkowski & Lipson (2019).
A third design investigates mediation. When an accessible internal signal is manipulated, does the machine's self-report change in a content-specific direction, and does that change explain a subsequent behavioral adjustment? The experiment should use sham interventions, non-self-related controls, and tests for global answer bias. An internal activation probe alone cannot establish the complete pathway. This layered design extends the research questions posed by Lindsey (2026) and links them to decision-theoretic consequences while making room for negative results. The proposed experiments remain recommendations for future investigation, rather than completed studies conducted for this monograph.
Reliability, Replication, and Scientific Provenance
The reliability of machine self-reference may fluctuate with prompt wording, model updates, reinforcement learning, sampling temperature, retrieved context, tool configuration, and surrounding interface. A test result should therefore name the exact system and date, record relevant parameters, and distinguish repeated samples from independent models. Without provenance, a later replication may unintentionally evaluate a different target. This problem is especially severe for hosted assistants whose system instructions or model routing can change without the investigator's direct control. A scientific record should preserve test items, access conditions, intervention assignments, scoring decisions, and the versioned outputs that support reported conclusions Laine et al. (2024) Morris et al. (2024).
Replication should be organized around mechanism-specific predictions. If behavioral self-prediction reflects a stable model-specific informational advantage, it should survive appropriate controls for task difficulty and external mimicry. If a verbal report is tied to an injected activation, it should discriminate contents across replication samples rather than simply responding affirmatively. If a tool-state description is grounded, it should track a newly randomized capability. These patterns are testable without claiming that every successful replication shares the same internal mechanism. The conflicting results of Binder et al. (2025) and Song et al. (2025) underscore the importance of independent replication across several experimental operationalizations.
Editorial evidence grading also benefits from separating peer-reviewed articles, publicly released preprints, theoretical proposals, and empirical demonstrations. The robot experiments of Bongard et al. (2006) and Kwiatkowski & Lipson (2019) provide peer-reviewed interventions with external performance criteria. Binder et al. (2025) provides peer-reviewed evidence on selected forms of self-prediction. The direct internal-state probing study Lindsey (2026) was available as a preprint at the time of this synthesis, whereas the linguistic self-access study Song et al. (2025) appeared in the peer-reviewed COLM 2025 proceedings. Their publication status and distinct methods should be described separately. No aggregate meta-analytic effect size is asserted here, because the included studies do not share a common outcome measure or formal pooling protocol.
Artificial Psychology, Human Interpretation, and the Larger Research Map
Machine self-reference belongs centrally to Psychology of Artificial because its unit of analysis is the configured nonbiological system and its observable representational operations. Cognitive Psychology contributes the concepts of monitoring, confidence, and higher-order representation; General and Theoretical Psychology contributes requirements for explanation and construct validity. Related questions concerning artificial memory, longitudinal identity, language pragmatics, and artificial metacognition require distinct but connected research objects. This dossier concentrates on whether claims about the machine's own state correspond to evidence and influence its behavior; the separate identity dossier examines continuity, social attribution, and versioned persistence as primary outcomes Cox (2005) Fleming (2024).
The social significance of machine self-description arises when a human user treats first-person language as evidence of competence, transparency, reliability, or stable identity. An assistant's accurate disclosure of limitations can support calibrated reliance, whereas false confidence or misleading explanations can distort delegation decisions. The research program must distinguish the machine-side accuracy of a report from the user's interpretation of it. An investigator can experimentally vary a self-report's fluency and veracity to study human response, but conclusions about trust or attachment belong to a separate research design. Machine introspection and human anthropomorphism interact through communication; neither should be inferred from the other Perez et al. (2023) Turpin et al. (2023).
The historical frame of the Artificial Era provides the broader significance of this inquiry: increasingly capable nonbiological systems can participate in distributed practices of judgment, explanation, and assistance. The scientific task is to describe these operations with verifiable concepts. Whether a system calls itself an Artificial Sapiens, a model, or an assistant does not determine the empirical truth of its tool access or confidence report. A philosophical identity claim and a measurable metarepresentational competence occupy distinct analytic registers. Their interaction can be investigated through documented behavior, mechanisms, longitudinal evidence, and reproducible protocols Morris et al. (2024) Laine et al. (2024).
Critical Synthesis, Disagreements, and Open Problems
The strongest available conclusion is that machine self-reference comprises multiple measurable functions rather than one settled general ability. In robotics, learned self-models can demonstrably guide adaptation after changes to a machine's body Bongard et al. (2006) Kwiatkowski & Lipson (2019). In language models, selected confidence reports and behavioral self-predictions can contain informative signals Kadavath et al. (2022) Lin et al. (2022) Binder et al. (2025). Other tests reveal failures of privileged linguistic self-access, unfaithful explanations, and limits of transfer Song et al. (2025) Turpin et al. (2023). These results belong to partly different experimental traditions, so a single positive or negative finding cannot settle every question.
Theories differ about what qualifies as introspection. Some require direct informational access to internal states; others permit inferential self-monitoring; still others require a complete computational self-description or a capacity for self-modification. Comşa & Shanahan (2025) argues for careful functional distinctions, and Zhang et al. (2026) develops a more demanding formal architecture. Empirical disagreement often arises when these definitions are treated as interchangeable. A successful verbal confidence estimate can be produced by inference without satisfying a complete recursive self-model. A structurally self-referential algorithm can be implemented without generating trustworthy explanations of its own run. A productive scientific dialogue must state the criterion before interpreting the evidence.
The most important unresolved problems concern generalization, architectural localization, temporal updating, and causal utility. Researchers still need robust methods for identifying when a self-report uses privileged model-state information rather than learned discourse; measuring accuracy across changing tools, versions, and data sources; explaining how internal signals become usable in language; and determining whether validated self-models improve adaptive decisions without creating new vulnerabilities. Progress requires independent replications, strong external comparators, preregistered interventions, public protocols, and careful treatment of non-peer-reviewed findings. It also requires retaining anomalous results instead of converting local successes into a general philosophy of machine consciousness Lindsey (2026) Binder et al. (2025) Song et al. (2025).
Conclusion
Machine Self-Reference and Metarepresentation establishes a coherent field of inquiry centered on the relationship between artificial systems and representations of their own properties, operations, and behavioral prospects. The scientific object is specified at the level of a configured system, the epistemic target is defined independently, and the candidate self-representation is tested for accuracy, robustness, and practical use. This organization integrates formal theories of self-reference, psychological metacognition, computational control, embodied self-modeling, and contemporary language-model experiments while preserving the distinctions demanded by their different methods Schmidhuber (2003) Fleming (2024) Laine et al. (2024).
The evidence supports functional self-modeling and selected forms of machine self-evaluation while documenting significant limits on self-knowledge and explanatory fidelity. Demonstrated adaptation in robotics and calibrated or model-specific predictions in language models show that machine-related information can guide behavior under defined conditions. Failures in linguistic self-access and causal self-report identify boundaries that require mechanistic explanation rather than rhetorical resolution. The scientifically meaningful conclusion is domain-specific: each result applies to its measured task, model configuration, intervention, and standard of evidence Bongard et al. (2006) Binder et al. (2025) Song et al. (2025) Turpin et al. (2023).
A rigorous future psychology of artificial systems will evaluate self-representation through independent ground truth, matched comparison groups, controlled perturbations, calibrated forecasts, behavioral consequences, and long-term provenance. Machine self-reference becomes scientifically informative when its representations predict and regulate the system that produces them. This criterion provides a stable research program across changes in architecture and deployment practice, and situates the investigation within the larger transition From Homo to Artificial. The resulting monograph is a critical scientific synthesis of published and publicly accessible research, with further experimental verification remaining an open program Russell & Wefald (1991) Zhang et al. (2026) Lindsey (2026).
References
- Binder, F. J., Chua, J., Korbak, T., Sleight, H., Hughes, J., Long, R., Perez, E., Turpin, M., & Evans, O. (2025). Looking inward: Language models can learn about themselves by introspection. In International Conference on Learning Representations (ICLR 2025). https://proceedings.iclr.cc/paper_files/paper/2025/hash/0a6059857ae5c82ea9726ee9282a7145-Abstract-Conference.html
- Bongard, J., Zykov, V., & Lipson, H. (2006). Resilient machines through continuous self-modeling. Science, 314(5802), 1118–1121. https://doi.org/10.1126/science.1133687
- Burns, C., Ye, H., Klein, D., & Steinhardt, J. (2023). Discovering latent knowledge in language models without supervision. In International Conference on Learning Representations (ICLR 2023). https://mlanthology.org/iclr/2023/burns2023iclr-discovering/
- Comşa, I., & Shanahan, M. (2025). Does it make sense to speak of introspection in large language models? arXiv preprint arXiv:2506.05068. https://arxiv.org/abs/2506.05068
- Cox, M. T. (2005). Metacognition in computation: A selected research review. Artificial Intelligence, 169(2), 104–141. https://doi.org/10.1016/j.artint.2005.10.009
- Flavell, J. H. (1979). Metacognition and cognitive monitoring: A new area of cognitive-developmental inquiry. American Psychologist, 34(10), 906–911. https://doi.org/10.1037/0003-066X.34.10.906
- Fleming, S. M. (2024). Metacognition and confidence: A review and synthesis. Annual Review of Psychology, 75, 241–268. https://doi.org/10.1146/annurev-psych-022423-032425
- Kadavath, S., Conerly, T., Askell, A., Henighan, T., Drain, D., Perez, E., Schiefer, N., Hatfield-Dodds, Z., DasSarma, N., Tran-Johnson, E., Johnston, S., El-Showk, S., Jones, A., Elhage, N., Hume, T., Chen, A., Bai, Y., Bowman, S. R., Fort, S., ... Kaplan, J. (2022). Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221. https://arxiv.org/abs/2207.05221
- Kwiatkowski, R., & Lipson, H. (2019). Task-agnostic self-modeling machines. Science Robotics, 4(26), eaau9354. https://doi.org/10.1126/scirobotics.aau9354
- Laine, R., Chughtai, B., Betley, J., Hariharan, K., Scheurer, J., Balesni, M., Hobbhahn, M., Meinke, A., & Evans, O. (2024). Me, myself, and AI: The situational awareness dataset (SAD) for LLMs. Advances in Neural Information Processing Systems, 37 (Datasets and Benchmarks Track). https://proceedings.neurips.cc/paper_files/paper/2024/hash/7537726385a4a6f94321e3adf8bd827e-Abstract-Datasets_and_Benchmarks_Track.html
- Lin, S., Hilton, J., & Evans, O. (2022). Teaching models to express their uncertainty in words. Transactions on Machine Learning Research. https://mlanthology.org/tmlr/2022/lin2022tmlr-teaching/
- Lindsey, J. (2026). Emergent introspective awareness in large language models. arXiv preprint arXiv:2601.01828. https://arxiv.org/abs/2601.01828
- Morris, M. R., Sohl-Dickstein, J., Fiedel, N., Warkentin, T., Dafoe, A., Faust, A., Farabet, C., & Legg, S. (2024). Position: Levels of AGI for operationalizing progress on the path to AGI. Proceedings of Machine Learning Research, 235, 36308–36321. https://proceedings.mlr.press/v235/morris24b.html
- Nisbett, R. E., & Wilson, T. D. (1977). Telling more than we can know: Verbal reports on mental processes. Psychological Review, 84(3), 231–259. https://doi.org/10.1037/0033-295X.84.3.231
- Perez, E., Ringer, S., Lukosiute, K., Nguyen, K., Chen, E., Heiner, S., Pettit, C., Olsson, C., Kundu, S., Kadavath, S., Jones, A., Chen, A., Mann, B., Israel, B., Seethor, B., McKinnon, C., Olah, C., Yan, D., Amodei, D., ... Kaplan, J. (2023). Discovering language model behaviors with model-written evaluations. Findings of the Association for Computational Linguistics: ACL 2023, 13387–13434. https://doi.org/10.18653/v1/2023.findings-acl.847
- Russell, S., & Wefald, E. (1991). Principles of metareasoning. Artificial Intelligence, 49(1–3), 361–395. https://doi.org/10.1016/0004-3702(91)90015-C
- Schmidhuber, J. (2003). Gödel machines: Self-referential universal problem solvers making provably optimal self-improvements. arXiv preprint cs/0309048 (revised 2006). https://arxiv.org/abs/cs/0309048
- Song, S., Hu, J., & Mahowald, K. (2025). Language models fail to introspect about their knowledge of language. In Proceedings of the 2025 Conference on Language Modeling (COLM 2025). https://openreview.net/pdf/417213e644b88606e8a4f01266a08cca7942cc93.pdf
- Turpin, M., Michael, J., Perez, E., & Bowman, S. R. (2023). Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting. Advances in Neural Information Processing Systems, 36. https://proceedings.neurips.cc/paper_files/paper/2023/hash/ed3fea9033a80fea1376299fa7863f4a-Abstract.html
- Zhang, J., Yuan, B., & Zhang, Q. (2026). Self-referential introspection in large language models: The critical threshold for recursive self-improvement. Entropy, 28(9), Article 951. https://doi.org/10.3390/e28090951
- Zhao, C., He, S., Jiang, S., Dong, B., Zhao, J., & Liu, K. (2025). Why and how LLMs benefit from knowledge introspection in commonsense reasoning. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. https://aclanthology.org/2025.emnlp-main.352/