Zero-data multimodal self-training stops chewing old tasks: a training-log alarm swaps in 400 fresh ones, lifting the 11-benchmark average from 57.77 to 61.05
Synopsis
Zero-data multimodal self-training has focused on what tasks to generate next and less on when a task pool has stopped supplying learning signal. A-RSI hands the timing to the training log: when the count of prompts with within-group reward variation falls at least 15% relative to two windows earlier, or logged accuracy rises at least ten points, the whole pool is replaced with 400 unused program-generated tasks whose answers are computed rather than voted. Qwen3-VL-8B-Instruct rises from 57.77 to 61.05 on the equal-weight average of eleven benchmarks using 800 fresh tasks, while an every-window control reaches 60.64 with 3,200.
Interpretation
Zero-data self-training now has a pool-swap switch that reads no test scores: when the reward-variation signal in the model's own training log weakens, the current pool is treated as spent. Earlier zero-data methods concentrated on what experience to generate next, such as harder problems or tasks near the model's estimated competence, and paid less attention to when fresh experience should arrive. On Qwen's old pool, the number of prompts with nonzero group advantages fell from 239 to 195 over three windows, a retention of 81.59% below the 85% threshold, triggering replacement; the accuracy condition never fired alone.
Fresh tasks carry answers computed by a program from the same underlying configuration, not labels produced by self-consistency or majority voting. When training targets come from self-consistency or majority voting, errors in the model's own predictions can become labels for its next update. From actually consumed replacement pools, 200 tasks were drawn from each of five generator groups for 1,000 total; independent solvers agreed with all 1,000 reference answers, and source labels and image hashes also agreed.
On the equal-weight average of eleven benchmarks, log-triggered refresh with random fresh tasks reaches 61.05 after eight continuation windows. Under the same protocol, an every-window arm with adaptive difficulty scores 60.64, a periodic arm 59.89, and a detector-triggered arm using frontier selection 59.59. A-RSI introduces 800 fresh tasks for a gain of 4.09 points per thousand tasks, while the every-window arm introduces 3,200 for 0.90; at the matched sixth continuation window, A-RSI scores 60.40 against 58.35 for continued training on the old pool.
The same fixed-threshold controller makes different decisions on MiMo-VL-7B-SFT: it replaces on the first two pools and keeps the third. A fixed three-window schedule would replace the third pool for both models. MiMo's second replacement pool retains 98.68% of its signal count while accuracy rises 6.63 points, and the replay keeps it; Qwen's second replacement pool retains 83.98% and the replay requests replacement.
Perspective
The demonstrated setting is continued training of Qwen3-VL-8B-Instruct and MiMo-VL-7B-SFT on eleven benchmarks at an 8,192-token output limit, with tasks drawn from a fixed library of program templates covering geometry, charts, tables and counting, and omitting function plots, numerical estimation and spatial tasks. The authors propose replication across seeds, a fixed-schedule random control for detector timing, a 16k evaluation of this run, and a refresh policy for low-entropy models.
The authors state that each arm was trained once, that small score differences lack significance testing, and that the first two random refreshes coincided with a three-window schedule, so the existing trajectory does not establish that history-dependent timing beats a fixed schedule; No-Refresh halted at window seven on a health check, leaving the pre-registered endpoint unobserved; MiMo's margins over every-window and periodic refresh are small and do not establish a timing advantage.