Show HN: Linguistic RL – A 7B model discovers Occam's Razor through reflection (github.com)

Researchers released "Linguistic RL," an experiment showing a 7B LLM learn to improve reasoning purely by reflection — no weight updates, no extra training data — by journaling failures and iteratively distilling its own textual strategies. On a scheduling constraint-satisfaction task (can N overlapping meetings fit into M rooms?) the model's accuracy rose from a 51.3% baseline to 78.0% after reflective distillation. Early runs featured confident, complex hallucinations (interval trees, DP, graph theory) that underperformed; successive journaled critiques created selection pressure that pruned elaborate hypotheses in favor of simple, robust strategies (e.g., counting/interval checking). The repo includes runnable code, three artifacts (thought logs, batch reflections, strategy evolution), and claims reproducibility on consumer hardware (8+ GB RAM, ~35–50 minutes CPU). This matters because it demonstrates meta-cognition and emergent Occam's Razor in a current LLM without changing weights: strategies are readable, transferable text rather than opaque parameter edits. Practically, LRL promises improved interpretability, cheaper experimentation (CPU-only), and a pathway to safer, self-correcting models that learn humility from empirical feedback. Open questions remain — generalization across domains/models, stability, and comparative efficacy versus fine-tuning or RLHF — but the experiment highlights a low-cost, human-readable alternative for iteratively improving model behavior.