Climbing the Ladder of Reasoning: What LLMs Can-and Still Can't-Solve after SFT?

• URL: http://arxiv.org/abs/2504.11741v1
• Date: Wed, 16 Apr 2025 03:39:38 GMT
• Title: Climbing the Ladder of Reasoning: What LLMs Can-and Still Can't-Solve after SFT?
• Authors: Yiyou Sun, Georgia Zhou, Hao Wang, Dacheng Li, Nouha Dziri, Dawn Song,
• Abstract summary: We conduct a detailed analysis of model performance on the AIME24 dataset.<n>We categorize questions into four tiers (Easy, Medium, Hard, and Extremely Hard)<n>We find that progression from Easy to Medium tier requires adopting an R1 reasoning style with minimal SFT-1K instances.<n>Exh-level questions present a fundamentally different challenge; they require unconventional problem-solving skills.
• Score: 59.418994222096885
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Recent supervised fine-tuning (SFT) approaches have significantly improved language models' performance on mathematical reasoning tasks, even when models are trained at a small scale. However, the specific capabilities enhanced through such fine-tuning remain poorly understood. In this paper, we conduct a detailed analysis of model performance on the AIME24 dataset to understand how reasoning capabilities evolve. We discover a ladder-like structure in problem difficulty, categorize questions into four tiers (Easy, Medium, Hard, and Extremely Hard (Exh)), and identify the specific requirements for advancing between tiers. We find that progression from Easy to Medium tier requires adopting an R1 reasoning style with minimal SFT (500-1K instances), while Hard-level questions suffer from frequent model's errors at each step of the reasoning chain, with accuracy plateauing at around 65% despite logarithmic scaling. Exh-level questions present a fundamentally different challenge; they require unconventional problem-solving skills that current models uniformly struggle with. Additional findings reveal that carefully curated small-scale datasets offer limited advantage-scaling dataset size proves far more effective. Our analysis provides a clearer roadmap for advancing language model capabilities in mathematical reasoning.

Related papers

• THOUGHTTERMINATOR: Benchmarking, Calibrating, and Mitigating Overthinking in Reasoning Models [65.39456695678713]
  We introduce approximate measures of problem-level difficulty and demonstrate that a clear relationship between problem difficulty and optimal token spend exists. We find that in general, reasoning models are poorly calibrated, particularly on easy problems. We introduce THOUGHTTERMINATOR, a training-free black box decoding technique that significantly improves reasoning model calibration.
  arXiv  Detail & Related papers  (2025-04-17T22:16:30Z)
• Long Is More Important Than Difficult for Training Reasoning Models [21.369780872368143]
  We show that reasoning length, rather than problem difficulty, primarily influences the performance of trained models.<n>We present our model, Long1K-32B, which achieves remarkable performance with only 1,000 training samples.
  arXiv  Detail & Related papers  (2025-03-23T13:33:59Z)
• Large Language Models and Mathematical Reasoning Failures [1.6114012813668932]
  This paper investigates the mathematical reasoning capabilities of large language models (LLMs) using 50 newly constructed high-school-level word problems.<n>We rigorously analyze both final answers and solution steps to identify reasoning failures.<n>We find that while newer models (e.g., o3-mini, deepseek-r1) achieve higher accuracy, all models exhibit errors in spatial reasoning, strategic planning, and arithmetic.
  arXiv  Detail & Related papers  (2025-02-17T09:07:32Z)
• LLMs Can Easily Learn to Reason from Demonstrations Structure, not content, is what matters! [53.84130385074551]
  Large reasoning models (LRMs) tackle complex reasoning problems by following long chain-of-thoughts (Long CoT)<n>We find that a Large Language model (LLM) can effectively learn Long CoT reasoning through data-efficient supervised fine-tuning (SFT) and parameter-efficient low-rank adaptation (LoRA)<n>With just 17k long CoT training samples, the Qwen2.5-32B-Instruct model achieves significant improvements on a wide range of math and coding benchmarks.
  arXiv  Detail & Related papers  (2025-02-11T08:48:48Z)
• Optimizing Language Model's Reasoning Abilities with Weak Supervision [48.60598455782159]
  We present textscPuzzleBen, a weakly supervised benchmark that comprises 25,147 complex questions, answers, and human-generated rationales. A unique aspect of our dataset is the inclusion of 10,000 unannotated questions, enabling us to explore utilizing fewer supersized data to boost LLMs' inference capabilities.
  arXiv  Detail & Related papers  (2024-05-07T07:39:15Z)
• The Unreasonable Effectiveness of Easy Training Data for Hard Tasks [84.30018805150607]
  We present the surprising conclusion that current pretrained language models often generalize relatively well from easy to hard data. We demonstrate this kind of easy-to-hard generalization using simple finetuning methods like in-context learning, linear heads, and QLoRA. We conclude that easy-to-hard generalization in LMs is surprisingly strong for the tasks studied.
  arXiv  Detail & Related papers  (2024-01-12T18:36:29Z)
• MuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning [63.80739044622555]
  We introduce MuSR, a dataset for evaluating language models on soft reasoning tasks specified in a natural language narrative. This dataset has two crucial features. First, it is created through a novel neurosymbolic synthetic-to-natural generation algorithm. Second, our dataset instances are free text narratives corresponding to real-world domains of reasoning.
  arXiv  Detail & Related papers  (2023-10-24T17:59:20Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.