Interrupting the Loop: Periodic Subject Changes Raise Judged Surprise and Connection in Base Language Models

Abstract: Where does the novelty a base LLM produces with no task come from, and what can an LLM judge of a long stream actually see? We dismantle a cognitively inspired generation loop over 24 conditions on three base models. Most of its effect lives in one operation: a new subject injected every few hundred tokens (an interruption) into a stream whose literal repetition is damped (habituation). We judge windows of generated text only, with the premise as the unit (n=10) and a judge measured for repeatability, against a second judge family and against human readers. Under that protocol the interruption raises judged surprise by 1.2 to 1.4 points and connection by 0.8 over habituation alone. A connective that asks for continuity hurts; a bare paragraph break adds nothing detectable on fresh text; a reset context does at least as well as a kept one; and a pre-registered replication on new premises confirms the primary contrast. Three things the window judge could not see changed the first version of this study, and we think they are of general use. The judge scores the experimenter's injected sentence as the model's own. A fixed rotation of injected sentences makes the model replay its earlier segments from beyond the judge's horizon, and the judge scores the replay as surprise and connection (65-80% of post-interruption windows at periods 150-300). And the local gains do not compose: no arm produces an integrated document. The salience monitor, the in-loop judge, memory across interruptions and a judge-gated Review run with a gate that opens add nothing. On a problem with a verifier (online bin packing), the interruption multiplies valid, distinct candidate heuristics three- to fourfold without raising the quality of the best. We report an evaluation protocol for long generation and a controlled characterization of a simple intervention, not a mechanism of creativity.

• The paper finds that combining repetition penalties with neutral subject changes every 150 tokens raises judged surprise from 0.45 for bare generation to 3.02, while matching or exceeding a larger creativity-inspired scaffold.
• The study shows that fresh-text gains persist across three base models, multiple judges, and human raters, but much of the apparent connection improvement comes from self-replay and disappears at the document level.
• Periodic interruption produces three to four times more distinct valid bin-packing heuristics without improving the best solution, suggesting it is a variation operator that requires selection to create value.

Overview and scope

This paper asks where the novelty a base LLM produces under task-free, forced open-ended continuation comes from, and which of the operations usually credited for it survive measurement. The program began with a cognitively inspired generation scaffold — the "DREAM" architecture, engineered around the default-mode/executive/salience triad of creative cognition, incubation, and re-encounter — and then dismantled it across 24 conditions on three base models (Qwen3-30B-A3B-Base as main generator, Qwen3-8B-Base, OLMo-2-13B), all run locally at 8-bit quantization. The outcome measured is deliberately narrow: judged narrative surprise, connection, and coherence on short windows of generated text, scored by an LLM judge from a different model family than the generator, with the premise (not the window) as the unit of inference. The paper is explicit that this is one ingredient of creativity — the appropriately unexpected turn — not creativity itself.

Two motivating nulls frame the contribution. An entropy-banded anti-probable decoder roughly doubles 4-gram novelty against the public OLMo-2 training corpus (21.5% to 45.5%) and cuts verbatim eight-word training blocks fourfold, yet shows no detectable effect at the level of ideas inside a generation loop or in verified search on online bin packing. Improbable prompts likewise show no benefit over typical ones. Both results pushed the investigation toward the loop itself.

The ladder: habituation and interruption

Under forced continuation with end-of-text masked, bare generation collapses within a few hundred tokens into degeneration modes (literal loops, exam keys, website footers), scoring 0.45 surprise / 0.30 connection / 2.52 coherence. Adding only habituation — a windowed repetition penalty of the same family as prior self-reinforcement remedies — removes literal loops and lifts surprise to 1.58 ($p=0.004$, Cliff's $\delta=+0.80$). Adding a neutral subject change injected every 150 tokens takes the stream to 3.02 / 3.68 / 6.12, matching or beating the full scaffold (2.70 / 1.85 / 6.02) on every dimension. The salience monitor, the in-loop judge (which never fired in 80 reviewed events), and the forgetting reseed add nothing detectable; the scaffold actually gives back connection, plausibly because it interrupts too rarely (0–3 reseeds per cell versus 30 for the clock).

The factorial confirms near-additivity: interruption without habituation gets most of the way alone (+0.87 to +0.91 surprise over habituation, $p \le 0.04$), and a stronger penalty or allowing EOS changes nothing material. The ladder replicates on all three generators from two families and is robust to quantization (bf16 gives nearly identical numbers).

What an LLM judge cannot see

A substantial part of the paper is a study of its own instrument, and three artefacts changed the first version's conclusions:

• Injected text read as the model's own. The original protocol cut windows containing the experimenter's injected sentences; the judge scored them as model output.
• Self-copy beyond the judge's horizon. A fixed rotation of four subject-change sentences teaches the model the cycle: it replays its earlier segments, and 65–80% of post-interruption windows at periods 150–300 are copied, with sources often outside the judge's 600-token context. Copied windows score high on connection by construction and on surprise when their source is invisible. On fresh windows only, the interruption's effect shrinks but survives: +1.20 to +1.35 surprise over habituation ($p \le 0.010$), while the large all-windows connection gain (+2.40) collapses to +0.75 — largely a replay artefact.
• Local gains do not compose. At the document level (whole 4,500-token streams, injections removed), no arm exceeds 2.5 on any dimension, and the interrupted stream over preserved context reads below uninterrupted habituation on document surprise (−1.30, $p=0.002$): a sequence of restarts, not a developing text.

The resulting protocol — generated-only windows, fresh-only estimates, document-level judgment, premise as unit, $k=5$ median judgments, a second judge family, human raters — is presented as a general corrective for windowed-judge evaluation of long generation.

What carries the effect

Battery 3 decomposes the interruption into boundary, content, and context. A bare paragraph break every 300 tokens does nothing detectable; a continuity connective ("And so, as before,") is if anything worse than nothing; injecting the premise itself or the stream's own past is as bad as not interrupting (−1.86 and −1.54 surprise against the neutral change, all $p \le 0.006$). The new subject is what works, and it works whether context is preserved or reset — reset scores higher (3.72 vs 2.90 surprise at period 300), partly because resetting prevents replay. A pre-registered replication on ten new premises confirms the primary contrast (H1: +1.48, $p=0.003$; fresh-only +0.82, $p=0.04$) and refutes two exploratory expectations in the less flattering direction: the reset context beats preserved context on connection, and habituation is not what makes the interruption work (H4 unsupported).

Timing results are similarly deflationary. Salience events as an interrupt trigger are indistinguishable from a clock at matched frequency; no tested period beats another on fresh text; judged surprise decays back toward the habituation level within several hundred tokens after an injection. A judge-gated Review run with a gate that opens (13% of reads) adds nothing over the clock, locally or at document level.

Verified search and the residual stream

On online bin packing, the interruption multiplies valid, distinct candidate heuristics roughly three- to fourfold (5.1 distinct valid per stream versus 1.8 for plain, $p=0.012$) without raising the quality of the best candidate — no arm reliably beats first-fit/best-fit on held-out instances. This profiles the operator as a variation source, not a value source; whether variation becomes value inside a FunSearch-style selection loop is explicitly left as the next experiment.

A descriptive residual-stream analysis at 13 sampled layers finds that bare generation moves least at every layer, that judged surprise co-varies with surface-layer departure (layer-0 novelty $\delta=+0.80$0 pooled) over an intact deep state, and that the best-scoring interruption barely moves deep representations — whereas the forgetting reseed moves them substantially toward the premise's state while being judge-rated lowest. Deep movement and judged quality dissociate throughout. The authors correctly flag this section as correlational, logit-lens-based, and non-causal.

Instrument calibration

Opus 5 shows median intra-window spread of about ±0.7 across five calls; Sonnet 5 is consistent only by scoring zero almost everywhere. A free test–retest arose from two byte-identical arms judged independently (91% exact agreement on surprise medians). Kimi K2.6 reproduces every condition ordering ($\delta=+0.80$1–0.85 across dimensions). Three blind human raters reproduce the condition ordering on surprise (interruption arms highest, bare lowest) and agree moderately with Opus ($\delta=+0.80$2 for the consensus), though they penalize the fluency cost of subject changes more than the judge does, and they did not rate connection in round 1. The authors treat the judge as the study's principal limitation and note that repeatability is not validity.

Limitations

The design is exploratory except for the confirmatory battery, the judge-gated arm, and the verifier probe; batteries were designed sequentially after seeing earlier results, and the corrected analysis protocol was fixed only after the first version was externally reviewed. Ten paired cells per comparison bound statistical power; half-point effects are undetectable at present resolution. Generators are 8–30B parameters, mostly quantized, with one post-trained model run only in raw continuation — nothing here speaks to dialogue use. Stronger decoding baselines (look-back, contrastive decoding) were untested. The "connection" dimension means connection within a 600-token horizon and cannot separate long-range integration from a return to the shared beginning, which is precisely where the reset-versus-preserved ambiguity lives. The network section establishes no mechanism.

Conclusion

Three findings carry the paper. First, in the regime studied, the novelty of task-free base-model generation comes from loop structure, and mostly from one operation: damp literal repetition and inject a new leading-away subject every few hundred tokens; the effect is real on fresh generated text, confirmed pre-registered, replicated across families, judges, genres, and human readers. Second, the gains do not compose into documents, and on a verified problem the operator multiplies candidates without improving the best — a variation operator awaiting selection. Third, windowed LLM judging of long streams is exposed to specific, correctable failures — injected text, out-of-horizon replay, and the window/whole gap — whose fixes constitute a reusable evaluation protocol. The open question the study hands forward is concrete: whether the variation this operator supplies becomes value inside a selection loop with a verifier.