Study Questions Whether Stage-1 Pre-training Controls Vision-Language Model Outcomes
A new arXiv preprint finds that in two-stage post-training for vision-language models, the Stage-1 initialization method primarily determines the policy entropy regime entering reinforcement learning, rather than the final task performance. The study used Qwen2.5-VL-7B with a 72B teacher model, comparing supervised fine-tuning and on-policy distillation warm-starts across geometry and math reasoning benchmarks. The findings suggest practitioners should not assume on-policy distillation is a superior RL warm-start simply because it produces higher initial entropy.
Researchers submitted a preprint to arXiv examining what Stage-1 warm-starts — either supervised fine-tuning (SFT) or on-policy distillation (OPD) — actually contribute to two-stage post-training pipelines for vision-language models. Using Qwen2.5-VL-7B with a same-modality 72B teacher, they found that all three warm-start variants converged to a narrow 53–54% accuracy band on the Geometry3K in-domain benchmark, providing little evidence that Stage-1 choice changes the in-domain endpoint. OPD entered reinforcement learning with substantially higher policy entropy and modestly better answer diversity and pass@16 scores (+2.0 to +5.2 points over SFT), but this advantage largely disappeared after RL training, with endpoint pass@16 values within 1.1 points across methods. On the out-of-domain MathVista benchmark, a carefully early-stopped SFT improved performance by +2.1 points compared to an over-trained variant, highlighting that training recipe details matter more than warm-start method for generalization. The authors characterize their contribution as a bounded empirical finding: Stage-1 strongly shapes the entropy regime, but the downstream performance payoff is small, localized, and insufficient to conclude OPD is a better RL warm-start than SFT.
What's missing
The study is conducted at small data scale with a single model family (Qwen2.5-VL-7B) and one teacher model, so it is unclear whether findings generalize to larger models, different architectures, or other task domains beyond math and geometry reasoning. The authors acknowledge that problem-level bootstrap intervals show the smaller OPD vs. SFT contrasts are statistically uncertain, and the available RL training trajectories may not be long enough to observe eventual divergence or convergence in entropy regimes.
What different sources said
- arXiv cs.AICenter
Stage-1 Controls the Entropy Regime, Not the Outcome
Related
Gut Bacteria Enzyme Found to Break Down Heat-Processed Food Compounds, Producing Novel Biogenic Amines
Researchers have discovered that an enzyme in common gut bacteria can degrade N-epsilon-carboxymethyllysine (CML), a compound formed during thermal food processing, producing previously unknown biogenic amines. The enzyme, ornithine decarboxylase SpeC from enterobacteria, acts on CML and related modified lysine derivatives through a low-level 'underground' catalytic activity. This finding suggests a previously unrecognized communication axis between thermally processed dietary compounds and gut microbial physiology, with potential implications for host health.
Full-Length Gene Sequencing Reveals Two Distinct Bacterial Communities in Black-Legged Ticks Expanding Into Canada
Researchers used Oxford Nanopore full-length 16S rRNA gene sequencing to characterize the microbiome of Ixodes scapularis black-legged ticks collected in Nova Scotia, Canada, distinguishing between tick-adapted bacteria and environmentally acquired bacteria. The study comes as I. scapularis — the primary vector of Lyme disease — is rapidly expanding northward into Canada due to climate change. The findings suggest that environmentally derived bacteria in tick microbiomes are not mere contamination, which has implications for how tick microbiome data is collected and interpreted across surveillance studies.
Study Identifies Metabolic Link Between Cell Envelope Stress and Biofilm Formation in Bacteria
Researchers have discovered that the metabolite acetyl-CoA directly inhibits enzymes that degrade the bacterial signaling molecule c-di-GMP, connecting cell envelope biosynthesis stress to biofilm formation in Pseudomonas aeruginosa. The study found that sub-inhibitory concentrations of antibiotics targeting early peptidoglycan biosynthesis — but not other antibiotic classes — elevate c-di-GMP levels by reducing phosphodiesterase activity, with acetyl-CoA competing for the enzyme active site. Because the relevant enzyme domain is broadly conserved across bacterial species, this checkpoint mechanism may be widespread and could have implications for understanding antibiotic-induced biofilm responses.