The research notebook / Updated 4 November 2026
Take the possibility
seriously.
Test the explanation.
Creativity research, computational discovery and a few useful equations help explain the design of innoMOS. They also help explain why a convincing idea still needs work.
Three different thingsResearch can support a component. Software checks can show that a workflow runs. Comparative evidence is needed to establish that the complete innoMOS method improves outcomes. These are separate claims.
01 / What are we trying to stretch?
A model can propose.
Reality gets a vote.
Beyond the first plausible answer
A familiar representation can steer the whole search. innoMOS deliberately varies the starting point: a problem to reframe, an encounter between documented pieces, an anomaly to investigate, or a function looking for a new purpose.
Randomness chooses inputs without asking the model to select its favourite combination. Abstraction and analogy change what can be transferred. A same-input comparison asks whether the resulting relation actually does anything different.
What the method does not change
It does not change the model's weights, supply missing scientific facts or guarantee novelty. A well-written explanation can still omit a critical condition. A strange combination can still lead to a standard solution.
The useful ambition is broader exploration coupled to more discriminating tests. It is not a claim that ordinary language models cannot invent, or that a prompt removes every limit.
02 / Five primary sources
What the studies say.
What we take from them.
01 / Science Advances · 2024Individual creativity and collective diversity
Doshi & Hauser — Generative AI enhances individual creativity but reduces the collective diversity of novel content.
In a short-story experiment, access to AI ideas improved evaluated individual stories while increasing similarity across AI-assisted stories.
- Design implication
- Keep different routes into the problem, rather than equating one polished answer with a varied search.
- Boundary
- A short-story result is not a universal result about invention, current models or innoMOS.
Publisher / DOI ↗
02 / ICLR · 2025Ideas can be judged novel before they are built
Si, Yang & Hashimoto — Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers.
The study found higher judged novelty for LLM-generated proposals, alongside feasibility and generation-diversity limitations.
- Design implication
- Take model-generated possibilities seriously. Preserve the conditions that would make them work.
- Boundary
- Reviewers' novelty judgements are not implemented discoveries or exhaustive prior-art checks.
Authors' paper ↗
03 / Execution follow-up · 2025 preprintA proposal's appeal can change after execution
Si, Hashimoto & Yang — The Ideation-Execution Gap: Execution Outcomes of LLM-Generated versus Human Research Ideas.
Experts implemented assigned research ideas. Earlier advantages for LLM proposals weakened after execution.
- Design implication
- Separate the exciting proposal from the result of a discriminating implementation or experiment.
- Boundary
- These tasks and implementations do not settle every domain or establish an innoMOS benefit.
Authors' follow-up ↗
04 / Nature · Published online 2023Mathematical discovery needs an evaluator
Romera-Paredes and colleagues — Mathematical discoveries from program search with large language models.
FunSearch coupled model-generated programs with systematic evaluation and search, producing cap-set constructions and bin-packing heuristics.
- Design implication
- Use an external check that can distinguish a better proposal from a persuasive one.
- Boundary
- This is a particular evaluated search system, not a general theorem-solving guarantee. innoMOS is not FunSearch.
Paper abstract / DOI ↗ Authors' implementation ↗
05 / Quality-diversity search · 2015Look for a landscape of useful possibilities
Mouret & Clune — Illuminating search spaces by mapping elites.
MAP-Elites preserves high-performing solutions across predefined behavioural dimensions, rather than returning only one optimum.
- Design implication
- Diversity and intermediate stepping stones can be useful search objectives.
- Boundary
- innoMOS does not implement MAP-Elites by borrowing this idea; a curated operation catalogue is a different mechanism.
Authors' paper ↗
03 / Equations that have a job
Count the space.
Then challenge the function.
C(n, k) = n! / [k! (n − k)!]There are 2,024 unordered three-piece groups in a pool of 24 distinct pieces. A pool of 100 has 75,287,520 five-piece groups. These are counts of input sets, not original ideas.
E[Uₘ] = N [1 − (1 − 1/N)ᵐ]With N possible groups and m independent uniform draws, this is the expected number of distinct groups observed. For 15 draws from 220 groups it is about 14.53. Repeats are possible even with a sound random generator.
The first result follows from counting selections without replacement or order. For the second, each group has probability 1 − (1 − 1/N)m of appearing; adding those probabilities gives the expectation. Within each draw the pieces are distinct. Groups may repeat between draws. The homepage controls are mathematical illustrations, not a claim that a 100-item research population has been curated.
Operate the calculation ↗f_proposal(x) ≠ f_known(x)A test criterion, not a law of creativity: on the same valid input x, does the proposed relation enable a different relevant action or inference than the nearest known functional counterpart?
Then remove or replace the claimed coupling. If the difference survives unchanged, the proposed explanation may be wrong. A difference from a weak comparison does not establish a difference from the best known solution. None of these equations proves originality.
04 / Our evidence, separately
Research-informed.
Still being tested.
What exists in the product
The inspected Discovery implementation offers problem, encounter, anomaly and exaptation entry profiles. It uses curated, source-linked operation hypotheses, a three-outcome Explore → Realize flow, and retention checks. Random and Ortho routes address different parts of exploration.
The host model supplies the creative proposals. Structural acceptance is not creative-quality acceptance.
What our pilots establish
Our bounded development pilots include actual paired runs and functional checks. They do not yet establish general superiority over a strong direct prompt, verified originality, first-mover status or human benefit. A three-pair core.4/core.5 pilot ended INCONCLUSIVE.
innoMOS 0.4.0 remains a preserved reference. A newer version's additional workflow features do not, by themselves, make it a better invention method.
Project basis: the 20-part discovery research synthesis (17 October 2026); inspected Discovery source; NEXT3 paired-pilot report (26 October 2026). The research loops are literature work, not 20 method-effect experiments. Source inspection here used paper abstracts, publisher excerpts and author project material; it was not a full-text systematic review.
Back to the possibility ↗