The Core Claim
A vision-language model shown an obscure plant photo and asked "what is this?" will not say "I don't know." It fabricates a confident, plausible-sounding name. Even after correction, it reverts to the original fabrication on the next turn. This is a symptom of two long-standing AI problems usually tackled separately: catastrophic forgetting (new learning erases old) and hallucination (confident invention of unknown facts).
Louis Mouchon's 2026 research note argues these are not two problems but two faces of one gap: the model lacks a "is this familiar?" signal. Supply it, and both behaviors fall into place. Mouchon calls this signal surprise.
Inspiration From Neuroscience
The brain has run a similar signal for hundreds of millions of years. The hippocampus rapidly encodes new episodes; the neocortex slowly integrates them into long-term knowledge. Sleep performs offline replay between the two — a mechanism known as Complementary Learning Systems (McClelland & O'Reilly, 1995). The switch that drives the system is prediction error: a surprising stimulus gets written, a predictable one does not. The same signal also tells the organism "this is unfamiliar," tempering downstream responses. One signal, two jobs: a plasticity gate and a metacognition substrate. Mouchon transplants the structure directly into a neural architecture.
System 1: Surprise-Gated Memory
The architecture is deliberately minimal:
- A frozen backbone (DINOv2 or I-JEPA) embeds images into vectors and never updates.
- A small JEPA predictor (JEPA anchor) attempts to reconstruct full embeddings from masked views; reconstruction error becomes the surprise score.
- A hippocampus modeled as a non-parametric store. New images are written only when surprise exceeds a threshold.
- A neocortex modeled as a linear classifier, updated only during the sleep phase using replayed samples.
- Surprise < 0.35 (known): the model is instructed to answer confidently, drawing on retrieved facts.
- 0.35 ≤ surprise ≤ 0.65 (partial familiarity): the model hedges, providing candidates while flagging uncertainty.
- Surprise > 0.65 (novel): the model enters "student mode" — it describes only what it sees, explicitly states the concept is unfamiliar, asks the user, and then learns the concept from a single sentence.
- Familiar → no new memory needed (plasticity off), confident answer allowed (metacognition permits).
- Unfamiliar → write new memory (plasticity on), suppress guessing (metacognition blocks).
- Paper: Surprise as a Signal for Plasticity and Metacognition, Louis Mouchon, 2026-06-28 (Note: the source post cites this URL; the publication date is taken from the source)
- Author: Louis Mouchon (independent researcher)
- Code: not yet released (proof-of-concept)
Freezing the backbone eliminates representation-level forgetting by construction. All adaptation happens in the lightweight modules above.
Key Numbers
On a 1000-class continual ImageNet stream:
| Configuration | DINOv2 old-class retention | I-JEPA old-class retention | |---|---|---| | No replay (naive) | 65.8% | 25.9% | | Sleep replay | 83.5% | 77.2% | | Upper bound (i.i.d. training) | 84.4% | 78.5% |
Sleep replay recovers 17.7 percentage points for DINOv2 and 51.3 percentage points for I-JEPA. The harder the backbone's features are to linearly separate, the more replay matters.
A Counterintuitive Finding
Replaying only the most recent tasks seems like a sensible memory-saving strategy. On DINOv2, old-class retention drops to 41.2%; on I-JEPA it collapses to zero — worse than no replay at all. Partial replay continuously reinforces new memories at the expense of old ones, which is precisely the catastrophic forgetting mechanism. All-or-nothing replay is the only viable choice, mirroring biology: sleep replays distant and recent memories together.
System 2: Surprise-Driven Metacognition
The same surprise signal is reused for a different purpose: enabling the model to recognize what it does not know.
The backbone switches to SigLIP (shared text-image space) plus a BGE-M3 text retriever. Surprise scores are calibrated to [0, 1]: known concepts score around 0.14, novel concepts around 1.0. Two thresholds partition the space:
Why the Detector Must Be External
The paper's most striking experiment pits three novelty detectors against each other:
| Detector | AUROC | |---|---| | External surprise detector (proposed) | 0.966 ± 0.024 | | Model self-reported confidence (0–100 scale) | 0.618 ± 0.060 | | Model token-level confidence ("yes, I know" probability) | 0.292 ± 0.043 |
Token-level self-confidence is below random chance. The model is equally confident about fabricated and real concepts. This is not a bug but a training inevitability: a model optimized to answer becomes structurally incapable of judging its own knowledge. Therefore the surprise signal must come from outside — a frozen detector that does not participate in generation.
After Sleep: 99.2% vs 0%
The system is taught 50 facts (one shot each), a sleep phase runs (hippocampal consolidation followed by hippocampal clearing), the entire dialogue history is wiped, and the 50 facts are re-queried. Result: 99.2% correct recall (mean over 5 seeds, 95% CI ± 0.9). The baseline model without episodic memory: 0%. Facts survive independently of dialogue context — they have been integrated into the neocortex.
Authority of Memory
The most counterintuitive design choice: when memory conflicts with pretrained priors, memory wins. The user states "Mont Blanc is 4806 m" (the model's pretraining says 4810 m). The base model sticks with its prior in roughly two-thirds of cases. The full system injects retrieved facts with an instruction that they override prior knowledge. Result: 3/3 answers use the corrected value, and the model explicitly acknowledges the new teaching. Memory is not advice — it is command.
Why One Signal Serves Two Functions
The paper's title is the thesis: *Surprise as a Signal for Plasticity and Metacognition*. Both behaviors rest on the same judgment — "is this input familiar to me?"
Evolution spent hundreds of millions of years discovering this economy. AI researchers spent decades patching the two symptoms separately. Mouchon's contribution is to reunite them behind a single scalar residual from a JEPA predictor.
Engineering Takeaways
The paper is a proof-of-concept with small benchmarks and honest limitations, but several insights are likely to persist:
1. Frozen backbone + lightweight adaptation layers is the right posture. Large models stay fixed; adaptation occurs in small, inspectable modules. This excludes representation-level forgetting by structure and yields interpretability — every memory has a known location and write time.
2. External detectors are far more reliable than self-reported confidence. The 0.966 vs 0.618 vs 0.292 gap is alarming. Any hallucination-mitigation pipeline that depends on a model evaluating its own confidence may be built on an unreliable foundation.
3. Non-parametric stores do not forget. Parametric memory (e.g., test-time training) forgets old concepts whenever writing stops. Prototypes written once are immune to this failure mode by design.
4. Partial replay is worse than no replay. Counterintuitive but obvious once stated: either replay everything or skip replay entirely — sliding-window memory savings backfire.
5. Sleep is not optional. Teaching 50 facts without a consolidation phase and then clearing the dialogue window yields 0% recall. Sleep converts fragile short-term traces into stable long-term representations. Skip it, and the system is just a chatbot with a context window.
Why This Matters Beyond the Benchmarks
The dominant AI research trajectory of the past decade has been scaling — more parameters, more data, more compute. That path assumes all problems dissolve with scale. This work identifies something scale cannot deliver: a model needs a signal for what it does not know, and that signal cannot emerge from scale. GPT-5.5 and Gemma 4 12B alike show token-level self-confidence below chance.
Scale lets a model know more. It does not let a model know what it does not know. The latter requires an architectural commitment: place the surprise signal outside the generator.