Found first: a primary source the press has not covered yet.
Researchers at Alibaba Cloud Computing tested whether successful post-training updates could be reused across domain shifts and found that 13 of 24 candidate-context pairs failed to improve the target task when naively applied. The paper, "Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training", proposes a method called Boundary-Calibrated Intervention Transfer (BCIT) that conditionally authorizes past updates based on whether the conditions for their success still hold.
What the source says
Tingyun Li and colleagues ran experiments on a 4B-parameter model across three domain shifts: finance reasoning (FinQA to TAT-QA), text-to-SQL (Spider to BIRD), and function calling (xLAM to BFCL). Of 24 candidate-context pairs, 11 improved the target task when updates were naively reused and 13 did not. Compared with a flat-additive baseline, BCIT cut the harmful authorization rate from 62.5% to 25.0% and raised beneficial candidate coverage from 80.0% to 90.0%, though the authorization quality difference did not reach conventional significance (McNemar p = .25). In a cross-task evaluation over 36-GPU-hour episodes, BCIT achieved a mean gain of 2.63 percentage points over the flat-additive baseline (95% CI [2.10, 3.16]; sign-flip p = .031). A short-run validation protocol using 20% of the compute budget recovered direction agreement on 20 of 24 candidates (83.3%; Spearman ρ = 0.72).
Why it matters
Autonomous post-training pipelines that adapt models across tasks without human oversight at each step commonly assume that what worked before will work again in a new context. This paper quantifies how often that assumption is wrong: 13 of 24 times, across domains that represent standard production adaptation scenarios. BCIT adds an applicability check before each reuse decision, and the 2.63 pp gain held consistently across seeds and domains. The 20%-budget validation result matters separately: most of the signal about whether a past update will transfer can be recovered at a fraction of full-training cost, which makes the check cheap enough to run routinely in an automated pipeline.