Show HN: Training-free improvement of Pareto frontiers for reasoning LLMs (swireasoning.github.io)

SwiReasoning introduces a training-free strategy to push the Pareto frontier of reasoning LLMs by dynamically switching between “latent” and “explicit” thinking modes. Latent mode lets the model internally carry out multi-step reasoning without emitting chain-of-thought (CoT) text (saving tokens and latency), while explicit mode produces full CoT when additional clarity is needed. The core is a lightweight, prompt-driven policy that estimates when the latent answer is reliable (via confidence signals or ensemble/self-consistency heuristics) and only falls back to explicit CoT for uncertain cases—no fine-tuning or model weight changes required. This simple switch-thinking mechanism yields Pareto-superior trade-offs between accuracy, compute cost, and token/latency overhead: many examples are solved cheaply in latent mode while difficult instances trigger more expensive but more accurate explicit reasoning. The paper formalizes these trade-offs, describes practical switching heuristics, and evaluates the approach across standard reasoning benchmarks, showing consistent outward shifts of the accuracy-vs-cost frontier. For practitioners, SwiReasoning offers an accessible way to get better throughput and cost-efficiency from off-the-shelf LLMs by combining confidence-aware routing with existing prompting techniques.