Continual learning for frozen models
Assay is a research program on learning after training. New capabilities and facts go in as small modules that switch on only for their own domain and come out bit-for-bit. The base model never changes, so nothing it already knew is overwritten.
Operation library · frozen 4B model
5 / 5 installed“A veks B. C veks A.” Who veks A?
Two placements given. Who is in the middle?
A chain of comparisons. Who is second?
Comparisons stated with “not”. Who ranks higher?
Rule out everyone else. Who holds the role?
Refused. Tracking state across steps would not install at any module size we tried.
01 The idea
Installs
A skill is defined over a whole space of inputs, so it applies wherever the reasoning lands. One module of about 10,000 parameters installs transitive inference: held-out domains go from 73% to 94%, above 10-shot prompting (78%) and above a prompt that spells out the rule (70%).
Needs an interface
A fact only answers the exact question it was stored under. Read in place during reasoning, an installed fact comes back 71% of the time; asked as an isolated lookup, 93%. Giving facts that interface lifted multi-hop accuracy on edited facts from 30% to 44%.
Open question
Tracking state step by step is the one thing no module has carried. It also defeats a single forward pass: permutation tracking stays at chance even after full training, while written step-by-step reasoning solves it. Whether a latent workspace can replace the writing is what we are testing now.
02 The architecture
We never retrain the base model. Everything new lives in parts that can be tested on their own, switched on only where they apply, and removed without a trace.
Each part is marked by its current status: demonstrated, partial, or open.
A frozen pretrained model: Qwen3-4B in our experiments. Every result is checked against the same base-weight hash.
Small installed modules for skills and facts, each routed to its own domain and removable bit-for-bit.
A loop that decides what stays. Modules are built from the model's own verified answers, and a functional-invariance constraint keeps earlier ones alive.
A small adapter that learns what to look up and how to carry it through a chain of reasoning.
Latent reasoning for what cannot be installed: state that has to be tracked step by step.
Rules installed into learned dynamics models, where there is no prompt to state them in.
Choosing among candidate outputs by predicted outcome instead of likelihood.
03 Results
| Result | Measured | Bound |
|---|---|---|
| One installed reasoning operation, on held-out domains | 0.73 → 0.94 10-shot prompting 0.78 | A skill, not new competence: it transfers across content, never to neighbouring operations. |
| Five operations live at once, built from the model's own verified answers | 0.82 – 0.96 unrelated skills unchanged | Synthetic operation families. Step-by-step state tracking was refused. |
| Learning signal: verified answers vs reasoning traces | +0.17 to +0.23 vs +0.00 | Three replications on held-out domains; one 4B model family. |
| Accumulation under a functional-invariance constraint | 23/23 facts alive after 5 generations (vs 3/11 without) | The cost to new learning grows with every generation. |
| Installed facts inside reasoning: read in place vs isolated lookup | recall 0.71 → 0.93 multi-hop 0.30 → 0.44 | Composition gained +0.07 against the +0.20 we set as the bar. |
| A usage adapter on edited multi-hop questions | +0.10 95% CI [0.04, 0.16] | Trained by supervision. Outcome RL added nothing. |
| LLM plus a world model with installed rules, vs the LLM given the rules | 0.98 vs 0.20 | One environment, templated language. |
| Code: pick candidates by predicted outcome vs by likelihood | 0.35 vs 0.15 on problems the model fails | 40 tasks, replicated on 20 more. |
Every bar was written down before the run it judges. Full records, including the retractions, are in the research repository.
04 Negative results
Outcome RL collapsed at the same point in two runs. The policy learned phrasing habits, such as asking which city where the question needed a country, that derail the next step.
Our first continuous-thought model scored just as well with its thoughts replaced by averages, or by another example's thoughts.
Step-by-step state tracking was refused three times, at every module size we tried. One variant fit its training data and transferred nothing.
A full fine-tune lost 62% of its question-answering knowledge while perplexity stayed flat. We measure retention with question batteries instead.
05 Open source
Library · Apache-2.0
A write-path gate for agent memory. It rejects poisoned, contradictory and stale writes before they are stored, with a reason for every decision: 98–100% rejection on LOCOMO and LongMemEval-S, regenerated by one command.
Demo
Memory for local models that folds facts into the weights. The model answers from an empty context, then forgets on command, with the base weights restored.
Record
The research record: every result with its preregistered bar, its controls, its bound, and the failures kept next to the wins.
06 How we work
Every pass or fail threshold is written down before the experiment that it judges, and deviations are recorded with the reason.
Each result is compared against the alternatives that could explain it: scrambled modules, prompts that describe the rule, and the model without the module.
Before and after are measured on the same questions, with confidence intervals, and the base-weight hash is checked on every run.
Assay is built by Arsenios Papachristos in Thessaloniki, Greece. The experiments run on frozen open-weight models, on a single laptop.
Contact
For research collaborations, design partners who want models that learn on the job and can undo it, or a note when the write-up is out.
We reply from arseniospapa@gmail.com. Never sold. See the privacy note.
Good fits