Meta Labs Paper Finds Base LLMs Beat RL Post-Trained Models at Scale
Post-training raises single-shot reliability while shrinking the ceiling of what test-time compute can reach, a trade-off most teams never measure. Builders choosing between base and RL-tuned models for agentic pipelines can use the paper's few-rollout estimate to decide when extra sampling beats extra training. PTGS offers a concrete fix for labs running their own RL pipelines.
A paper from Meta Superintelligence Labs shows that pre-trained base LLMs with a light inference harness can surpass RL post-trained counterparts on agentic tasks when given a large test-time sampling budget.
Benchmarks include BFCL v4 multi-turn, ACEBench and WebShop: post-trained models win at pass@1, but at large K base models often solve tasks the post-trained versions never solve.
The authors name this lost test-time scalability the Sharpening Tax, observing it across 42 base and post-trained model pairs, with the effect growing with model size.
They propose PTGS, which sets per-prompt sampling temperature from estimated task difficulty during RL, reducing the tax and raising pass@1.
Great paper for all you pre-training and post-training nerds. The surprising finding is that pre-trained LLMs, equipped with a light inference harness, can serve as capable agents and can even surpass post-trained counterparts with a sufficient test-time budget. DAIR.AI @dair_ai Banger paper from Meta Superintelligence Labs. They find something super interesting and unexpected. (bookmark it) Base models with a light harness often solve more agentic tasks than their RL post-trained versions when both get enough samples. Post-trained models win on pass @1 . At large K, base models frequently solve tasks the post-trained ones never solve on BFCL v4 multi-turn, ACEBench, and WebShop. This is because post-training pushes each task toward always solved or never solved. Consistency goes up, and coverage goes down. The authors call the lost test-time scalability the Sharpening Tax. Across 42 base and post-trained pairs, it shows up in most settings, grows with model size, and can be estimated from a few rollouts. Their fix, PTGS, sets the sampling temperature per prompt from its estimated difficulty during RL. It pays a smaller tax and also raises pass @1 . Paper: academy.dair.ai/papers/sharpen… 🔗 View Quoted Tweet📊 11 ⚡
This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Sharpening Tax, RL post-training, Meta Superintelligence Labs, BFCL v4.
Source attribution: curated from elvis (@omarsar0).