A single agent only reweights SFT data: splitting execution, four analyses and synthesis lifts IFEval to 63.33 by round six
Synopsis
Language model agents are increasingly handed the post-training stack, yet a single agent converges on supervised fine-tuning and spends its effort reweighting data categories. AutoSFT splits the loop into an executor, four analysis subagents that read each round from training data, failures, training signals and history, and a synthesis subagent that writes new examples for undersupplied, verifier-checkable capabilities. Every round retrains from the same base checkpoint, so each component is attributable: analysis reaches 63.33 on IFEval at round six, above the single agent's best of 62.35 across 50 rounds.
Interpretation
AutoSFT widens SFT data research from one kind of change to a search over six data-action families, with execution, analysis and synthesis handled by separate subagents. Earlier automated systems gave the agent one control, either selecting rows from a fixed pool or calling a generator; all 33 rounds of the reproduced single-agent run stayed inside category-proportion adjustment. On Qwen3-1.7B-Base with IFEval and IFBench, the single agent submitted 33 rounds and evaluated 32, peaking at round 9 on both benchmarks with no improvement in the following 23 rounds.
An executor subagent moves training, checkpoint conversion and evaluation out of the deciding agent's session, returning only file paths and a short status report. The single-agent session reached 94.7% of a 258,400-token window and was compacted ten times, discarding and rebuilding research history each time. With the same base checkpoint and context window, evaluated from snapshots at 11.5 hours, compactions fall from ten to five and shell commands from 58% to 29% of recorded events.
Four analysis subagents each propose one data change from a different evidence source: training data, failure cases, training signals and earlier rounds. The single agent stayed in the mixture family throughout, and its reasoning never mentioned repairing a row, replacing a response or writing a new example. Under a 3M-token cap with avg@10 evaluation, the analysis configuration reaches 63.33 at round 6 and 67.10 at round 30, against a base best of 60.60 in the first 30 rounds and 62.35 over 50.
A synthesis subagent generates examples for capability targets the pool undersupplies, and a row enters the pool only after passing that target's programmatic verifier. Neither of the first two components introduces data the pool did not already contain, so capabilities with almost no staged rows stay out of reach. Under a 100M-token cap, the synthesis run reaches a best IFEval of 81.58 and IFBench of 51.95 against 78.15 and 49.02; repeat_prompt sits near 40% for eight rounds and rises to 75.6% in the round its batch enters.
Perspective
The results come from one base model at one scale, Qwen3-1.7B-Base, with one search per configuration in instruction following and in code. Instruction following is scored on IFEval and IFBench with avg@10, code on MBPP+ with avg@3, and every round retrains from the base checkpoint under a fixed token budget. The code transfer study uses short runs, so it shows the loop can transfer rather than how broadly.
Each comparison runs a single search, and repeated evaluation bounds the noise of one checkpoint but not the variation between independent searches. Synthesis gains concentrate on a few learnable capabilities the pool undersupplies and do not move constraints the model cannot yet follow, with odd_even_syllables scoring zero despite batches in three rounds. The authors note the effects may not hold at other scales.