This study asks whether an agent can learn more usefully from an episode by choosing what to train on as the episode develops. It begins with online parameter adaptation, such as LoRA updates, and builds on existing test-time training research. It is an independent investigation arising from Construct-2's S1 question.
Does choosing between self-generated text, environmental observations, summaries, and no update improve later decisions when the most useful source changes during an episode?
The comparison that matters is against the strongest fixed source or fixed mixture selected on development data, with comparable resources and information access. A selector might identify useful evidence as circumstances change. It might instead gain only by making fewer updates, or by choosing the source that is generally best for that model. Those are competing explanations to examine.
Begin with a focused investigation: understand the closest existing methods, inspect whether their implementations can support the question, and develop a small, informative empirical comparison where feasible. Start with existing implementations where useful. The investigator owns the exact protocol and may revise the question in light of prior work or early observations.
The intended first outcome is a short local research note explaining what is already answered, what was investigated, what the evidence supports, and whether a larger investigation is worthwhile. Finding that the question is already answered, that dynamic selection offers no advantage, or that the available setup cannot distinguish the explanations is useful progress. A positive result is not required.
No numerical time, compute, or spending cap has been specified. This starting brief does not imply an unlimited campaign: keep the initial work focused and bring substantial expansion or new resource requirements back to the human. Publication polish should fit the evidence and effort.
- aTTT, 2607.03441v1, §§3–5 compares Self, Env, and Summary text for episode-specific adapter updates. It leaves online selection or combination of those sources open in that paper. Inspect its exact update, reset, and evaluation implementation first.
- Self-Guided TTT, 2607.09415v1 already selects training spans with a known question. Establish what an episode-level selector would add beyond that setting.
- VANE, 2608.09448v2 provides related adaptation-validation work in robotics; its modality and information access differ from textual agent experience.
- Beyond Perplexity, 2607.00368v1 distinguishes improved prediction loss from useful deployment memory.
- Decision-theoretic TTT, 2606.15569v1 is a follow-up reading candidate; the parent project has reviewed only its abstract. Check its relevance before making novelty claims.
The parent's B0 explanation and evidence map provides mechanism background and links to Construct and Formation findings. In particular, producing useful information and using it successfully can fail separately. The map is a starting assessment, not an exhaustive novelty review.
Hold the base model and adaptation mechanism fixed initially. Compare dynamic selection with fixed sources, a fixed mixture, a matched random selector, and no update. Include an explicit-context comparison using the same available information. Common recorded episode prefixes could isolate the immediate effect of selection before live trajectories introduce different subsequent experiences. These are design suggestions, not a frozen protocol.
Measure later task behavior alongside training loss, update counts, and total cost. Charge selection and summary generation as well as adaptation. Any source verification information available to the selector must also be available to its controls. Distinguish novel but incorrect text from repeated but useful corrections; novelty alone does not establish training value. A gain that disappears against a well-chosen fixed policy or after cost matching supports a simpler explanation.
- Open-weight models through
docker modelare preferred. - OpenAI models through
codexand SpaceXAI models throughagentare available resource routes; their usable roles depend on the interfaces provided. - Use
uvfor Python and Docker/Compose for model or supporting services where appropriate.
Inspect local model availability, hardware, and training support before sizing experiments. Inference access alone does not provide the gradients or mutable parameters required by the neural treatment.
On 10 September 2026, the human also offered a second Apple M3 laptop with 48 GB memory for larger models if needed. Notify the human when a move is warranted. Access and training support on that machine remain unverified.
The first paper/code audit and feasibility proposal is available, with saved supporting evidence. The current laptop has 8 GB memory and an unreachable Docker model runner. No aTTT implementation was located in the focused search; the S-TTT repository was empty. Pinned MLX source provides a candidate training route. No model experiments or empirical adaptation findings are recorded yet. Keep methods, code, data, saved outputs, and reproduction instructions in this project. Add links here as evidence and a manuscript become available. LaTeX source, a bibliography, and a compiled PDF are the intended publication format, scaled to the investigation; an early useful result need not wait for polish.
The ancillary-study guidance describes the relationship: this project investigates and publishes locally; Construct-2 reads those publications and updates its broader theories.