Tag
Masked distillation is a knowledge-distillation framework that trains a student LLM to predict only solution tokens while a reasoning teacher provides feedback, aiming to internalize chain-of-thought computation into model parameters. The method shows task-dependent success, working on GSM8K but requiring small scaffolds for harder tasks like Countdown.
This paper investigates whether a frozen looped transformer can read its own computation quality (pre-answer prediction reaching AUROC 0.797) and whether external interventions can improve outcomes, finding that no tested frozen intervention produces a validated capability gain, a property termed operational proto-introspection.
This paper investigates whether structured synthetic reasoning data can improve arithmetic reasoning in small language models under consumer-hardware constraints. Fine-tuning Qwen3-0.6B and Qwen3-1.7B with LoRA on a synthetic corpus derived from GSM8K improved exact-match accuracy by 12-13 percentage points, with strong transfer to related benchmarks.
LoRA Speedrun is a public wall-clock leaderboard for fine-tuning techniques, measuring how fast LoRA adapters can achieve target accuracy on GSM8K and SQuAD tasks using Qwen2.5-1.5B and SmolLM2-1.7B on a single L40S GPU, with automated verification.
CADENCE introduces a unified framework addressing cold-start collapse, state-agnostic scheduling, and reward sparsity in on-policy knowledge distillation for reasoning, achieving strong results on GSM8K and MATH-500 with compact student models on a single Mac Studio.
This paper introduces Viable Path Entropy (VPE), a finite-budget measure of verified continuation capacity for intelligent systems, decomposing capability into verified reachability and verified-mode diversity. Experiments on GSM8K with Qwen2.5-Instruct models demonstrate that accessible verified continuation capacity, rather than parameter count, determines mirror horizon.
This study introduces section-aware compression for reasoning traces, training models to drop filler narration while preserving compute and verification spans, matching or exceeding original accuracy while using 2-3 times fewer tokens.
SPEAR is a code-augmented agentic prompt optimizer that uses a Python sandbox for structural error analysis, achieving state-of-the-art performance on multiple LLM evaluation suites including industrial judge tasks, BBH, and GSM8K.
A framework called GuardedRepair is proposed for post-hoc replacement of LLM mathematical reasoning, using selective replacement with safety guards to fix errors while minimizing harm to correct traces. On GSM8K it improves accuracy from 95.60% to 96.89% without breaking correct answers.
This paper identifies a 'positional copying' shortcut where small language models answer arithmetic questions by copying the last number before the answer delimiter, bypassing actual reasoning. This effect explains why shuffling CoT steps retains performance; it accounts for 89-92% of teacher-forcing accuracy in 1-3B models on GSM8K.
HRM-Text introduces a Hierarchical Recurrent Model that decouples computation into slow and fast layers, enabling efficient pretraining from scratch on only 40 billion tokens and a $1,500 budget, achieving competitive performance with larger models.
OpenAI trained a system using verifiers to solve grade school math word problems with 90% of child-level accuracy, nearly doubling fine-tuned GPT-3 performance. The approach addresses language models' weakness in multistep reasoning by training verifiers to evaluate candidate solutions and select the best one.