Roberto I. Ono Filho
Independent researcher, São Paulo. Machine novelty: what stimulating a language model buys, measured.
In one paragraph
Is the answer to an open problem already inside a language model, waiting for the right stimulus? I have been trying to find out at office scale, one laptop and rented GPUs by the hour, under one discipline: pre-registered batteries read once, blind judges that score only generated text, exact verifiers with classical baselines, sham controls, every candidate logged. So far every inference-time stimulus bought variation or integration and none bought progression; moving the weights moved the mass toward the known good and paid with the tails; and in verified search every proposer stalls at the same plateau, a record away. Negative results are published with the same care as positive ones.
Papers
- 04
Pulsed, Not Pressed: What State-Space Injection Buys in Open-Ended Generation, and What It Costs
Intervening in the residual stream during generation. Pulsing a concept direction for a few tokens every few hundred raises judged surprise where pressing it continuously collapses coherence; the direction's content carries the effect (shams fail); against the written subject, the pulse matches text interruption at its peak, trails it everywhere else, and costs coherence.
- 03
Operator Packages, Proposer Strength, and Construction-Family Plateaus in Office-Scale Verified Search
A FunSearch-style loop instrumented at office scale, on nine construction problems from a public repository. Three operator packages compose additively and close more of the seed-to-record gap; a frontier proposer reaches in tens of samples what the local model does not in hundreds; and every proposer stalls at the same plateau, a record away; a later analysis reads the flagship wall as the grid's resolution limit rather than the family. Named in words, the record's family is adopted and loses; handed as code, it is optimized.
- 02
Moving the Mean Toward the Known Good, Not Beyond It: What Inference-Time Interventions and Weight Consolidation Buy in Open-Ended Generation
Every inference-time intervention bought variation or integration and none bought progression; a verifier placed inside the flow was imitated rather than obeyed. Consolidating the model on its own verified findings moves what it writes toward the known good, replicated in three lineages, and never beyond it: the best candidate stays the classical heuristic, and the tails where better-than-classic candidates live get thinner.
- 01
Interrupting the Loop: Periodic Subject Changes Raise Judged Surprise and Connection in Base Language Models
A base model writing with no task drifts into repetition. Of the operations usually credited for novelty, one survives measurement: a new subject injected every few hundred tokens into a habituated stream raises judged surprise and connection, replicated across models and on held-out premises. The anti-probable sampler and the improbable prompt do not.
Method
- Pre-registration, single read. Hypotheses, arms, outcomes and the analysis script are fixed and committed before any data; each battery is read once.
- Judges that cannot be gamed by the intervention. Language-model judges from another model family score windows of generated-only text and never see the injected material; human blind rounds check the instrument.
- Exact verifiers and classical baselines. In verified search, every candidate is scored by an exact verifier against published records and classical constructions; claims are stated as “below the published value”, never “better than”.
- Controls that can undo a result. Shams that copy the surface of an intervention without its content, and no-source controls that reveal what the model already knew.
- Everything released. Code, every logged candidate and judgment, and the dated lab notebook, on GitHub.
Code and data
github.com/RobertoOno/interrupting-the-loop: the harness (samplers, the reverie loop, state-space operators, the verified-search loop and its domains), the analysis scripts, the papers’ LaTeX, and the lab notebook. Run data as releases: run-data-v1 (papers 2 and 3) and run-data-v2 (paper 4).
Contact
- roberto@robertoono.com
- ORCID
- 0009-0006-8650-629X
- GitHub
- RobertoOno