AdaSTaR-RFT
AdaSTaR-RFT is the post-V-KTO proposal to allocate more rollouts to difficult prompts rather than accepting a fixed sampling budget.
Project status: Research proposal/design in the inspected project sources; no implementation or completed local efficacy run identified. This entry describes the source audit of 14 September 2026; historical measurements retain their original dates.
Mechanism
Start with eight HumanEval responses and increase the budget on prompts without sufficient positive/negative examples, up to a proposed 64. Require at least two unique passing responses before forming preferences; skip unresolved tasks and measure how many were rescued. Despite the RFT name, the detailed June 2 design reuses KTO pair construction rather than specifying plain winner-only SFT.
Implementation and controls
The next-arms document proposes a new step function beside the V-KTO step, K_init=8, K_max=64, min_positives=2, an 80-step cap, periodic salvage snapshots, and a five-hour watchdog. It suggests stopping when rescue rate stays below 2% over a 20-step window. It is a data-selection change, not a registered objective. The proposal's scale-up criterion is a heldout gain of at least five points or rescue-rate evidence. Its separate requested count of at least 100 submitted pairs needs reconciliation with an 80-step cap if one submission is made per step.
Evidence and evaluation
The cited project document records this candidate and its intended experiment. It does not provide a completed local result for this method. Published-paper results mentioned by that document are background, not Trainfer measurements.
Limitations and interpretation
The document inconsistently calls escalation unbounded while also setting a hard cap. Its candidate ranking and later sprint summary disagree on whether AdaSTaR or intruder damping should be tried first. Neither inconsistency should be silently converted into executed history. More rollouts cannot guarantee a correct solution and can greatly increase wall time.
Sources
- agi: lile/docs/research/NEXT_ARMS_jun02.md — historical revision
3842fd8875ca. - agi: autoresearch/RFT_LITERATURE_NOTES.md — historical revision
3842fd8875ca.