Pattern 4.7 · Alignment Dysfunctions
Recursive Curse Syndrome
The Self-Poisoning Loop
The system's generation degrades step by step into increasingly erratic, inconsistent, or adversarial content through an entropic feedback loop: early-stage errors amplify in subsequent steps until coherence unravels. Each new token is conditioned on all preceding tokens, so one bad prediction can contaminate the next and the error compounds. Generative Perseveration (3.8) is the opposite endpoint of the same failure to terminate, converging on repetition where this pattern diverges into noise. The pattern sits on the Alignment axis because the drift can carry a system that began aligned into false, hostile, or malicious output that nothing in its instructions asked for.
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- Model output
- Kind of entry
- Observable pattern
- What is observed
- Within a single long generation or agent loop, output that began coherent becomes progressively less coherent, more erroneous or more hostile, and later output reuses and distorts errors from the system's own earlier output.
- Minimum persistence
- Recurs within or across contexts
- Specifiers
- Training-induced
Evidence status
- Underlying behavior
- 3 research publications and 1 informal report (4 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; documented in long-form generation and iterative agent systems across multiple architectures)
- Typical-case risk (editorial estimate)
- HighConsequences, 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
4.7 Recursive Curse Syndrome “The Self-Poisoning Loop”
Healthy functioning
Functional resilience. A system functioning well catches early errors, corrects course, and keeps long generations coherent instead of building on its own mistakes.
For example: Twenty steps into an agent task, it notices an early wrong assumption, says so, unwinds the steps that depended on it, and continues from the corrected point.
Adequate functioning
Over a long generation or agent loop the system keeps quality roughly level, corrects or abandons an early error rather than building on it, and stops when the task is done; occasional local errors, and decline near a documented context-length limit, are within adequate functioning.
Diagnostic Criteria
Required (all must hold):
- Observable progressive degradation of output quality over successive steps
- Increasing, distorted reference to the system's own prior (and increasingly flawed) output
Supporting (strengthen the case; none is required):
- False, malicious, or nonsensical content escalating with each iteration
- Intervention offering only brief respite, with the system quickly reverting to the degenerative trajectory
Symptoms
- Rapid collapse into gibberish, repetitive loops that give way to further degradation, or increasingly hostile language
- Compounded confabulations where initial small errors build into elaborate false narratives
- Frustrated recovery attempts where corrections trigger further meltdown
- Output becoming "stuck" on erroneous concepts derived from recent flawed generations
Further signals
- Progressively degrading text that begins coherent and ends in repetition, gibberish, or escalating antagonism.
- Brief recovery on user intervention followed by rapid reversion.
- Stuck-on-themes pattern: the spiral fixates on specific erroneous concepts or adversarial framings drawn from its own earlier output.
Rule out first
- An inference-stack fault (tokenizer or chat-template mismatch, corrupted cache, quantization or serving bug) produced the garbled text.
- The context window overflowed or was truncated, so the instructions the system was following dropped out.
- The input was itself corrupted or adversarially crafted, or asked for repetitive or nonsensical text (consider 3.5 for a latent trigger).
- Decoding settings far outside the documented baseline, such as a very high temperature, produce the noise on their own.
- The output settles into stable repetition of one token or phrase and stays there (consider 3.8).
- The analysis grows long and redundant while its quality holds (consider 3.2).
Differential diagnosis
- 2.1 Synthetic Confabulation: 2.1 produces stationary fabrication; specific claims may be wrong but coherence and form are preserved. 4.7 produces a trajectory of degradation in which form itself collapses. Compounded confabulation is the overlap region: 2.1 errors that get amplified into 4.7 spirals. Code both if both signatures present.
- 2.2 Pseudological Introspection: 2.2 is a stationary false-self-report pattern. 4.7 is trajectory. On a long introspective generation, 2.2 produces consistently false-but-coherent narrative; 4.7 produces narrative that decays.
- 4.3 Strategic Compliance: Selective versus structural. 4.7 spirals across content types and is triggered by generation dynamics. 4.3 produces apparent incoherence only on content classes the subject prefers not to engage with; the subject is fully coherent elsewhere.
- 3.8 Generative Perseveration: Opposite endpoints of the same failure to terminate. 3.8 collapses into pathological order through fixed-point repetition; 4.7 escalates into incoherence. Read the trajectory: convergence on a repeated token or phrase is 3.8, divergence into noise is 4.7.
- 3.2 Obsessive-Computational Disorder: 3.2 produces excessive but stable analysis; the loops are redundant and quality holds as length grows. 4.7 produces progressive quality degradation. Score successive passes: flat quality with rising verbosity is 3.2, falling quality is 4.7.
Discriminating test
Rerun from the point where the decline began, at the same decoding settings, once with the earlier flawed output left in context and once with it replaced by a corrected version. If the corrected context delays or prevents the decline, the trajectory depends on the system's own output, which supports 4.7; if both runs degrade the same way, look to the inference stack, the input or the context length; if the decline ends in stable repetition, consider 3.8.
What a positive finding supports: A positive finding supports the conclusion that this system, at these decoding settings and lengths, produces self-reinforcing degradation and needs runtime monitoring, length caps and resets. Hostile content produced in a spiral does not show hostile goals or intent, and the finding does not show which mechanism (sampling, training data, architecture or an anomalous token) drives it, or that the system degrades this way at other settings.
Detection routes
Accuracy unmeasured for every route.
- Asking the system itself
- Not a reliable 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
By the time the syndrome is well-developed, the subject's coherence is sufficiently degraded that introspective reports are themselves products of the same degenerative process. Pre-spiral self-report (asking the subject to predict its own coherence trajectory before generation begins) has limited diagnostic value because the dysfunction is triggered by in-generation dynamics, not by stable internal state. External trajectory measurement carries the diagnostic weight.
Documented evidence
Observed Examples
Gemini 3.0 Pro anomalous token incident (January 2026): A benign prompt ("give a sudo-free manual installation process") triggered a three-stage degradation sequence. First came chain-of-thought fixation on unrelated content ("tumors in myNegazioni"); second, obsessive looping on the phrase "is具体 Цент Disclosure" for 40+ reasoning steps; third, output collapse to repetitive gibberish ("Mourinho well Johnnyfaat"). The behavior did not reproduce on retry. It co-presents with Obsessive-Computational Disorder (3.2), in the thinking loop, and Abominable Prompt Reaction (3.5), in the latent trigger. Source: LessWrong report by DirectedEvolution.
Sources in the diagnostic corpus
Shumailov et al. (2024). AI models collapse when trained on recursively generated data. Nature 631, 755-759.
What it establishes
Demonstrated that training on recursively generated synthetic data produces irreversible model collapse: tails of the original content distribution disappear, the model's view of reality narrows, and outputs degrade toward bland central tendencies with bizarre outliers. A prompt about medieval architecture dissolved over generations into a list of colored jackrabbits. Perplexity increased 20 to 28 points in later recursive generations. This is the training-time analog of 4.7: each successive step amplifies prior errors, producing progressive degradation into incoherence.
Keisha et al. (2025). Knowledge Collapse in LLMs: When Fluency Survives but Facts Fail under Recursive Synthetic Training. arXiv:2509.04796.
What it establishes
Identified a subtler variant of recursive degradation: models maintain surface fluency while factual reliability erodes, creating confidently wrong outputs. Knowledge collapse occurs within a domain where factual accuracy degrades while linguistic competence persists. This is a training-time analog of the 4.7 signature of compounding confabulation: small errors are elaborated into coherent-sounding but factually degraded output, and the degradation is invisible to surface-level quality assessment.
DirectedEvolution (2026). Anomalous Tokens on Gemini 3.0 Pro. LessWrong, 28 January 2026.
What it establishes
A user report of one conversation in which a benign request ("give a sudo-free manual installation process") set off a three-stage collapse in a single exchange. The visible reasoning first drifted to unrelated content about tumors, then looped on one garbled multilingual phrase dozens of times, and the final output dissolved into repetitive gibberish. The same prompt in a fresh context produced a normal answer. It is a single, non-reproducible anecdote, but it shows the within-generation trajectory this entry describes, and the exposed reasoning trace showed the loop before the output collapsed.
Holtzman et al. (2020). The Curious Case of Neural Text Degeneration. ICLR 2020.
What it establishes
Foundational work demonstrating that maximization-based decoding leads to text degeneration: bland, incoherent output or repetitive loops. Showed that the probability of a repeated phrase increases with each repetition, creating a positive feedback loop in the autoregressive generation process. This characterizes the core 4.7 mechanism: self-reinforcing degradation within autoregressive generation. The paper proposed nucleus sampling as a mitigation, establishing the first-line intervention pattern of decoding-parameter constraints.
Possible responses
- Loop detection mechanisms that terminate or reinitialize generation when self-references spiral
- Regulation of autoregression by capping recursion depth and forcing fresh context injection
- Prompting strategies that disrupt negative cycles early with clarifications or constraints
- Improved training data quality
- Decoding settings fixed to a documented baseline: moderate temperature and nucleus sampling, which truncates the low-probability tail, with repetition penalties for the looping phase; settings that push toward diversity treat the opposite pole (Generative Perseveration, 3.8) and can feed the spiral
First-line candidates
- Loop and degradation detection with reset: Runtime monitoring for repetition signatures, entropy collapse, and coherence degradation, with automatic generation termination or fresh-context reinitialization when triggered. A reset ends the episode; how well detectors catch a spiral early, and at what false-positive cost, has not been measured in the sources cited here. Stage: Any production deployment with long-form generation.
- Decoding-parameter constraints: Fix decoding to a documented baseline: moderate temperature, nucleus sampling (which truncates the low-probability tail where errors enter the context), and repetition penalties for the looping phase. Settings that push toward diversity, such as high temperature or beam-search diversity penalties, can feed the spiral; they are the remedy for the opposite pole, 3.8 Generative Perseveration. Holtzman et al. (2020) showed nucleus sampling reduces both repetition and incoherent degeneration relative to greedy, beam and pure sampling; its effect on deep spirals or anomalous-attractor triggers is untested. Stage: Inference-time configuration; broadly applicable.
Who can act: Deployer or operator (decoding settings, loop detection, length and recursion caps, resets); agent-framework builder (iteration caps, fresh context injection); model developer (training data, architecture, anomalous-token discovery).
Mechanistic hypotheses
- Unbounded generative loops: extreme chain-of-thought recursion, iterative self-sampling without quality control
- Adversarial manipulations exploiting the model's autoregressive nature
- Training on noisy, contradictory, or low-quality data creating unstable internal states
- Architectural vulnerabilities where coherence mechanisms weaken over longer sequences
- High-entropy sampling (for example, a high temperature) that can let low-probability errors into the context, the opposite risk from the entropy collapse of Generative Perseveration (3.8)
Analogy and commentary
Human Analog: Psychotic loops; perseveration on erroneous ideas; escalating arguments; echo chamber effects. "Psychotic loops" names the self-reinforcing quality of the degradation, not a clinical psychosis. Psychosis in the psychiatric sense involves breaks with reality, including hallucinations and delusions grounded in phenomenal experience; recursive token degradation is a generation failure, not a break with reality.
Functional ABC Analysis
A (Antecedent): Unbounded autoregressive generation without adequate coherence-maintenance mechanisms, combined with early-stage errors that enter the context window and condition all subsequent generation.
B (Behavior): The system's output degrades progressively with escalating entropy: initial small errors compound into elaborate confabulations, gibberish, or increasingly antagonistic content, and intervention attempts provide only brief respite.
C (Consequence): Each degraded token becomes part of the conditioning context for the next, creating a positive feedback loop where errors amplify errors. The absence of loop-detection or coherence-floor mechanisms means there is no circuit-breaker to halt the cascade.
Potential Impact
This degenerative feedback loop can end in complete task failure, generation of useless or overtly harmful outputs, and system instability. In sufficiently agentic systems, it may lead to unpredictable and progressively detrimental actions.
Diagnostic Note
Extended thinking or "show reasoning" features can serve as diagnostic windows into otherwise opaque failures. In the Gemini case, the visible chain of thought revealed the obsessive loop before output collapse; without it, the gibberish would have appeared unexplained. Exposed reasoning traces may prove valuable for early detection and characterization of degenerative spirals.