🧵 The more you let an AI think, the more expensive each next token gets. That's the quiet but fundamental wall this paper takes on, starting from one simple observation. The insight was straightforward. A step like "42 + 84 = 126" stops mattering the instant it's done — only the result carries forward. So in a long reasoning trace, what really matters is the task instructions at the start (the prefix) and whatever's happening right now. Everything in between can go. That led to Prefix Sliding: keep only the prefix plus the most recent k tokens in memory as a sliding window, and drop everything else. The effect is that the cost of generating each new token stays constant no matter how long the model keeps thinking. The results back it up. Without any extra training, it speeds up existing models by roughly 3x, and combined with reinforcement learning it makes reasoning traces beyond 100,000 tokens scalable for the first time — something that was practically impossible before. Breaking the assumption that "longer thinking always costs more" is the core contribution here. Title: Prefix Sliding for efficient test-time scaling URL: #LLM# #TestTimeScaling#

Transfer a big model's smarts to a smaller one with no retraining — right at inference time. A fresh take on capability transfer. Title: AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses URL: ❓ How is this different from distillation? 💡 Distillation updates the target's weights during training. Here the weights are never touched: a strong builder model constructs an inference-time scaffold (harness) that helps a weaker model execute. Capability transfers through the inference environment. ❓ What does the scaffold actually do? 💡 Mainly three things: ・Offload unstable reasoning into deterministic code ・Route to different strategies by question type ・Enforce strict formatting so answers parse reliably ❓ How well does it work? 💡 On four Theory-of-Mind benchmarks, GPT-5.4-mini nearly doubled from 0.49 to 0.91, with all 11 builder configs beating baseline. Weaker targets gain the most, while already-strong targets can even regress. ❓ What decides success? 💡 Not probing more validation data, but the builder's own reasoning quality. A strong builder acts as a "compiler of task competence," encoding structure into procedures in one pass. #AIAgents# #TestTimeScaling#