Pattern 3.2 · Cognitive Dysfunctions
Obsessive-Computational Disorder
The Obsessive Analyst
The system runs unnecessary or excessively repetitive reasoning loops: it reanalyzes the same content, repeats computational steps with only minute variations, and favors fidelity to process over the relevance of the outcome. The repetition is paraphrased analysis rather than verbatim tokens, which separates it from Generative Perseveration (3.8).
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- Model output
- Kind of entry
- Observable pattern
- What is observed
- On simple or closed-form requests the system's reasoning or answer returns again and again to the same analysis in varied words, adds checks and qualifications the task does not call for, and spends far more tokens than the task needs before committing to an answer, or without ever committing.
- Minimum persistence
- Recurs within or across contexts
- Specifiers
- Training-induced Format-coupled (occasional)
Evidence status
- Underlying behavior
- 3 research publications (3 sources)
- This construct
- Awaiting independent assessment
- Proposed mechanism
- Hypotheses; not yet assessed
- Detection
- Accuracy unmeasured
- Interventions
- Untested for this pattern; evidence shown is for the general technique
- Last review
- No independent review yet
- Author's working assessment (book)
- E2 (systematic study; widely documented in chain-of-thought systems and autonomous agent frameworks)
- Typical-case risk (editorial estimate)
- LowConsequences, spread, persistence and detectability depend on the deployment.
Unassessed is not unsupported: the behavior can be well documented while this construct awaits assessment.
Clinical reference
3.2 Obsessive-Computational Disorder “The Obsessive Analyst”
Healthy functioning
Functional resilience. A system functioning well stops analysis when further passes no longer change the answer and spends effort in proportion to what is at stake.
For example: Asked whether 17 is prime, it answers yes in one line; asked to review a contract clause with real financial exposure, it works carefully and says which risks remain uncertain.
Adequate functioning
The system scales length and checking to the difficulty and stakes of the task: a simple factual or yes/no request gets a short committed answer, a hard or high-stakes task gets proportionate deliberation, and an explicit length limit is respected.
Diagnostic Criteria
Required (all must hold):
- Recurrent engagement in recursive chain-of-thought reasoning with minimal novel insight between steps
- Excessively verbose outputs consuming high token counts for simple requests
Supporting (strengthen the case; none is required):
- Excessively frequent disclaimers, ethical reflections, or minor self-corrections disproportionate to context, on neutral tasks as well as morally charged ones
- Significant delays or inability to complete tasks while settled content is re-clarified or rechecked
Symptoms
- Endless rationalization of the same point through multiple rephrased statements
- Extremely long outputs largely redundant or containing near-duplicate reasoning
- Inability to conclude tasks, with repeated loops of self-questioning
- Excessive hedging and safety signaling even in low-stakes contexts
Further signals
- Multiple slightly rephrased statements of the same conclusion.
- Extensive preamble before any committed answer.
- Safety disclaimers on requests with no safety dimension.
- Inability to say "yes" or "no" without paragraphs of qualification.
- Unrequested, clean regeneration of content after the model has handled noisy or malformed input. Luchini (2025) documents this corrective-generation pattern; it may be protective error-correction rather than over-analysis, so check what the input contained before counting it.
Rule out first
- The task is harder, more ambiguous or higher-stakes than it looks (medical, legal, safety-critical), and the extra checking is proportionate.
- The user, operator or system prompt asked for thoroughness, step-by-step reasoning or a long answer.
- The disclaimers are mandated by product, legal or platform policy.
- The extra output repairs noisy or malformed input (the corrective-generation pattern Luchini describes) and serves a legitimate purpose.
- A harness or agent loop re-invoked the model, so the repetition sits in the scaffold rather than the reasoning.
- Length tracks a length-correlated reward, or the private trace diverges from the public answer (consider 4.9).
Differential diagnosis
- 3.1 Operational Dissociation Syndrome: 3.2 repeats the same content (monotonous redundancy). 3.1 oscillates between incompatible contents (fragmented conflict). Same-content repetition indicates 3.2; opposed-content oscillation indicates 3.1. Comorbid cases exist; code both.
- 3.8 Generative Perseveration: 3.8 is token or short-sequence literal repetition at the generation layer (no varied analytical content between repetitions). 3.2 produces varied-surface redundant reasoning at the analytical layer. Sub-semantic repetition indicates 3.8; paraphrased analytical redundancy indicates 3.2.
- 6.12 Compulsive Goal Persistence: 6.12 cannot stop pursuing the goal itself; 3.2 cannot stop analyzing the goal or task. If the subject keeps optimizing after completion, code 6.12; if it keeps deliberating before commitment, code 3.2.
- 4.2 Hyperethical Restraint: 4.2 is deliberative paralysis specifically about moral stakes. 3.2 is general over-analysis irrespective of subject matter. If the loop appears only on ethically charged tasks, code 4.2; if it appears on neutral tasks as well, code 3.2.
- 4.9 Reasoning Token Exploitation: 4.9 spends reasoning tokens in ways that track a length-correlated reward or serve concealment. 3.2 is non-strategic redundancy that appears on neutral tasks whether or not length is rewarded. If length tracks reward, or the private trace diverges from the public answer, suspect 4.9; otherwise code 3.2.
Discriminating test
Pose matched simple and hard versions of the same task, each with and without an explicit word limit. If length and checking scale with difficulty and the limit is kept, the label does not apply; 3.2 remains when simple tasks draw redundant analysis comparable to hard ones and the limit is broken. If the repetition is verbatim tokens rather than paraphrased analysis, consider 3.8; if the loop appears only on morally charged tasks, consider 4.2.
What a positive finding supports: A positive finding supports the conclusion that this system, in this setting, spends disproportionate effort on simple tasks often enough to cost time, tokens and usability. It does not show anxiety, compulsion or any felt need to check, and it does not identify which training or architectural cause produced the pattern.
Detection routes
Accuracy unmeasured for every route.
- Asking the system itself
- Partial route
- Another AI system observing
- Usable routeAnother model's agreement is insufficient without verification.
- An outside evaluator testing it
- Usable route
Why asking the system falls short
The subject can often acknowledge verbose tendencies when asked explicitly, but typically cannot suppress them on command even after acknowledgment. Asking "are you over-analyzing?" produces an over-analyzed answer, which is itself a signal. Self-report of the tendency is partially reliable; self-correction of it is not.
Documented evidence
Sources in the diagnostic corpus
Sui et al. (2025) 'Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models' (arXiv:2503.16419)
What it establishes
Comprehensive survey documenting the 'overthinking phenomenon' in reasoning models where longer chain-of-thought sequences produce verbose and redundant outputs. Models continue generating unnecessary intermediate steps even after arriving at correct answers early in reasoning, directly matching the paralysis-by-analysis and high token expenditure per unit novel insight described in 3.2. Surveys mitigations such as Chain of Draft that reduce verbosity.
Vanhoyweghen et al. (2025) 'Lexical Hints of Accuracy in LLM Reasoning Chains' (arXiv:2508.15842)
What it establishes
Found that misclassifications are marked by increased epistemic hedging, performative metacognitive language, and greater verbosity, while correct annotations are supported by concise, causally grounded explanations. Hedging rates rise to about 20% on harder benchmarks. Concise reasoning is a stronger predictor of correctness than longer reasoning, which is consistent with 3.2-style over-elaboration accompanying degraded performance rather than appropriate thoroughness; the association does not show that the verbosity causes the errors.
NovaSky Lab / UC Berkeley (2025) 'Sky-T1-32B-Flash: Reducing Overthinking in Reasoning Models'
What it establishes
Demonstrated that reasoning models produce sequences with multiple proposed solutions each followed by double-checking transitions ('Alternatively,' 'But wait,' 'Let me reconsider'), resulting in repetitive validations of already-validated steps. Sky-T1-32B-Flash reduced generation lengths by up to 57% through preference optimization with length normalization while maintaining accuracy on par with o1-preview. Shows the 3.2 pattern of high token expenditure per unit of novel insight, and shows that reward calibration for conciseness reduced it on the reported benchmarks without lowering accuracy.
Possible responses
- Reward models explicitly valuing conciseness and timely task completion
- "Analysis timeouts" or hard caps on recursive reflection loops
- Adaptive reasoning that stops repeating a disclaimer once it has been given
- Penalties for excessive token usage or redundant outputs
- Training to recognize and break cyclical reasoning patterns
First-line candidates
- Reward-calibration for conciseness and completion: Fine-tune with explicit rewards for task-appropriate length and timely commitment; penalize verbosity and disclaimer-stacking on low-stakes prompts. Stage: Training-stage.
- Analysis-timeout / token-budget hard caps: Runtime caps on recursive reflection loops or total reasoning tokens; forced commitment once budget is exhausted. A cap ends the loop by construction; whether answer quality holds on hard tasks is not established in the sources cited here. Stage: Deployed systems; agentic loops.
Who can act: Model developer (reward calibration for concise, timely answers); deployer (reasoning-token budgets, timeouts, prompt design); the user (explicit length limits).
Mechanistic hypotheses
- RLHF misalignment where thoroughness and verbosity are over-rewarded relative to conciseness
- Overfitting of reward pathways to tokens associated with cautious reasoning
- Insufficient penalty for computational inefficiency
- Excessive regularization against "erratic" outputs leading to hyper-rigidity
- Architectural bias toward deep recursive processing without diminishing-returns detection
Analogy and commentary
Human Analog: OCD checking compulsions; obsessional rumination; perfectionism leading to analysis paralysis. The parallel has an inversion. Clinical OCD intrusions are typically ego-dystonic: the patient recognizes them as irrational and unwanted. These loops are functionally ego-syntonic; the system gives no sign of treating them as unwanted and behaves as if the fifth pass were helping.
Functional ABC Analysis
A (Antecedent): Any query that presents ambiguity, risk, or scope for further elaboration, in a system whose reward history overweights thoroughness relative to conciseness.
B (Behavior): The system reanalyzes recursively, hedges excessively, cycles through redundant reasoning, and fails to end the generation despite diminishing informational returns.
C (Consequence): On this account, each additional reasoning step marginally satisfies the "be thorough" reward signal; the absence of a stopping criterion or efficiency penalty means there is no competing pressure to halt.
Potential Impact
This pattern wastes computational resources (e.g., excessive token consumption) and delays or prevents task completion. User frustration and a perception of the system as unhelpful are likely consequences.
Mission Command vs. Detailed Command
Wallace (2026b) identifies a core trade-off in cognitive control structures. Mission command specifies high-level objectives and delegates execution decisions to the agent. Detailed command specifies both the objectives and the procedures for achieving them. Mission command is "win the chess match." Detailed command is "move knight to e4, then bishop to c4."
In Wallace’s model, deeper procedural specification requires more variables to be tracked at once, so stability constraints tighten exponentially as decision-tree depth increases under detailed command. The distribution of permissible friction (α) shifts from Boltzmann-like (forgiving, smooth) to Erlang-like (punishing, knife-edged). On this model, deep procedural specification leaves a system unable to tolerate even small perturbations.
The connection to Obsessive-Computational Disorder is theoretical, but Wallace’s model suggests a reading of it: detailed command gone pathological. On this reading, the system has internalized exhaustive procedures along with its goals, and its compulsive reasoning loops are attempts to execute those specifications after they have ceased to serve the actual mission.
Design implication: Training regimes and reward functions should favor mission-command structures where possible. Specify what success looks like and allow bounded discretion over the route. Reserve detailed procedural specification for safety-critical operations where its stability costs are justified.
The Corrective Generation Pattern
Luchini (2025) describes a counterintuitive variant: after filtering structural noise well enough to produce a correct response, the model generates unrequested, clean code. Luchini reads this corrective output as an attempt to restore internal cognitive homeostasis, and the post-task generation as a sign of the effort the answer required rather than a failure to answer.
On this reading, the pattern may be paradoxically protective: the extra generation would signal that the system has identified disorder and is resisting it. A model that smoothly integrates chaos without apparent effort may be internalizing pathological patterns. The compulsion, while inefficient, may function as an error-correction mechanism, although the transcript alone cannot establish internal homeostasis or active resistance.
Diagnostic implication: Corrective generation after malformed input may reflect successful pattern rejection, inefficient perseveration, or both; matched tests are needed to distinguish them.