PaperlayerView exact revision on arXiv

Source-linked AI summary

AM-Thinking-v1: Advancing the Frontier of Reasoning at 32B Scale

Yunjie Ji, Xiaoyu Tian, Sitong Zhao, Haotian Wang, Shuaiting Chen, Yiping Peng, Han Zhao, Xiangang Li

arXiv:2505.08311v2cs.CL

TL;DR

AM-Thinking-v1 addresses whether a 32B dense model can achieve strong reasoning without private data or massive MoE architectures. Built from Qwen2.5-32B with public-data curation, supervised fine-tuning, and reinforcement learning, it reaches competitive mathematical, coding, and conversational performance. Its applicability remains limited by missing tool use and multimodal support, preliminary safety alignment, and variable performance across some languages and domains.

  • Problem

    The paper asks whether 32B dense models can unlock strong reasoning without private data or massive MoE architectures that increase deployment and fine-tuning burdens.

    Source: §1 · p. 2

  • Method

    AM-Thinking-v1 uses the public Qwen2.5-32B base, publicly available queries, data processing, supervised fine-tuning, and reinforcement learning.

    Source: §1 · p. 2

  • Results

    AM-Thinking-v1 achieves state-of-the-art reasoning performance among comparable open-source dense models, with 85.3 on AIME2024, 74.4 on AIME2025, and 70.3 on LiveCodeBench.

    Source: §1 · p. 2, §5.2 · pp. 10–11, §6 · p. 12

  • Takeaways & Limitations

    The findings support 32B dense models as a practical balance between deployment efficiency and reasoning performance.

    Source: §6 · p. 12

    The model lacks structured function-calling, tool use, and multimodal inputs; safety alignment is preliminary, and performance may vary across low-resource languages and domain-specific tasks.

    Source: §6 · p. 12

Abstract

from arXiv · show

We present AM-Thinking-v1, a 32B dense language model that advances the frontier of reasoning, embodying the collaborative spirit of open-source innovation. Outperforming DeepSeek-R1 and rivaling leading Mixture-of-Experts (MoE) models like Qwen3-235B-A22B and Seed1.5-Thinking, AM-Thinking-v1 achieves impressive scores of 85.3 on AIME 2024, 74.4 on AIME 2025, and 70.3 on LiveCodeBench, showcasing state-of-the-art mathematical and coding capabilities among open-source models of similar scale. Built entirely from the open-source Qwen2.5-32B base model and publicly available queries, AM-Thinking-v1 leverages a meticulously crafted post-training pipeline - combining supervised fine-tuning and reinforcement learning - to deliver exceptional reasoning capabilities. This work demonstrates that the open-source community can achieve high performance at the 32B scale, a practical sweet spot for deployment and fine-tuning. By striking a balance between top-tier performance and real-world usability, we hope AM-Thinking-v1 inspires further collaborative efforts to harness mid-scale models, pushing reasoning boundaries while keeping accessibility at the core of innovation. We have open-sourced our model on \href{https://huggingface.co/a-m-team/AM-Thinking-v1}{Hugging Face}.

1 Introduction

AM-Thinking-v1 addresses whether a 32B dense model can achieve strong reasoning without private data or massive MoE architectures. Built from Qwen2.5-32B with public queries and SFT-plus-RL post-training, it reaches competitive reasoning performance while retaining moderate-scale deployability.

  • AM-Thinking-v1 targets the gap between efficient 32B dense models and larger MoE systems that impose greater deployment and fine-tuning burdens.Source: §1 · p. 2
  • 85.3 on AIME2024, 74.4 on AIME2025, and 70.3 on LiveCodeBench demonstrate strong mathematical and coding performance for AM-Thinking-v1.Source: §1 · p. 2
  • AM-Thinking-v1 is built upon the publicly available Qwen2.5-32B base model and uses publicly available training queries processed through filtering, deduplication, quality control, and benchmark decontamination.Source: §1 · p. 2
  • Its post-training pipeline combines supervised fine-tuning with reinforcement learning, including a think-then-answer cold start, difficulty-aware selection, and two-stage training.Source: §1 · p. 2
  • The authors conclude that careful post-training can make 32B dense models comparable to leading reasoning models while preserving deployability advantages.Source: §1 · p. 2

2 Data

The data pipeline combines diverse public datasets with deduplication, quality filtering, benchmark decontamination, and verification. Mathematical and synthetic-response processing is especially extensive, while data balancing and multiple responses support broader and more robust training.

2.1 Data Collection

  • Training spans math, code, science, instruction following, and general chat using multiple publicly available open-source datasets.The collection includes verifiable ground truths for mathematical and scientific data and test cases for code data.Source: §2 · p. 2, §2.1 · p. 3

2.2 Query filtering

  • Queries are deduplicated and filtered for URLs, image references, and similarity to evaluation data through exact and semantic matching.Source: §2.2 · p. 3

2.2.1 Mathematical query filtering

  • Mathematical processing removes unclear or unsuitable problems and revises potentially incorrect ground truths using model agreement, math_verify, and o4-mini review.Mathematical proof problems and multi-subquestion queries are filtered, while multiple-choice items are rewritten as fill-in-the-blank questions.Source: §2.2.1 · p. 3

2.3 Synthetic response filtering

  • Synthetic responses are filtered by perplexity, repeated n-grams, dialogue structure, and the presence of complete think-and-answer components.Source: §2.3 · p. 4
  • Response quality and query difficulty are estimated from pass rates for grounded tasks or average reward-model scores for ungrounded tasks.Source: §2.3 · p. 4

3 Reward

AM-Thinking-v1 uses objective verification whenever possible and reward-model scoring otherwise. Math, code, and instruction-following rewards are binary aggregate outcomes, while non-verifiable queries receive averaged helpfulness, correctness, and coherence scores.

3.1 Verifiable Queries

  • Verifiable math, code, and instruction-following queries use rule-based verification or code execution to assess response correctness.Source: §3.1 · p. 4

3.1.1 Math

  • Math rewards are binary: the final boxed answer is extracted, normalized with math_verify, and scored 1 for a match or 0 otherwise.Source: §3.1.1 · p. 4

3.1.2 Code

  • Code queries run in a secure sandbox using method-call or standard-input/output test cases, and receive reward 1 only when all tests pass.The sandbox is distributed across machines with load balancing and queue management for secure, high-concurrency execution.Source: §3.1.2 · pp. 4–5

3.1.3 Instruction Follow

  • Instruction-following queries use the strict IFEval validator, assigning reward 1 only when every associated instruction is successfully followed.Source: §3.1.3 · pp. 5–6

3.2 Non-Verifiable Queries

  • For queries without objective criteria, a reward model scores helpfulness, correctness, and coherence, whose average forms the final reward.Source: §3.2 · p. 6

4 Approach

The approach combines large-scale SFT, difficulty-aware two-stage GRPO reinforcement learning, and infrastructure changes for scalable rollout. Detached workers and streaming load balancing target rollout bottlenecks caused by long and uneven generations.

4.1 Supervised Fine-Tuning

  • 2.84 million SFT samples cover five task categories, with upsampling for sparse data and multiple synthetic responses for difficult queries.Figure 4 reports distributions at instance and token levels, computed over responses rather than queries.Source: §4.1 · p. 6, Figure 4 · p. 6
  • Long-form reasoning SFT uses a larger learning rate and batch size than traditional SFT to achieve stable convergence.The reported learning rate is 8 × 10^-5, with a 32k maximum sequence length and final-response-only loss for multi-turn data.Source: §4.1 · p. 6

4.2 Reinforcement Learning

  • RL retains math and code queries with SFT pass rates strictly between 0 and 1, then transitions to a second stage that removes fully solved items and adds general-chat and instruction-following data.Source: §4.2 · pp. 6–7
  • GRPO is used with no KL constraint, zero advantages for overlong responses, strict on-policy updates, and stage-specific response lengths and learning rates.Source: §4.2 · p. 7

4.3 RL Framework

  • The verl-based framework integrates GRPO with vLLM, FSDP, and Megatron-LM for scalable training across 1000+ GPUs.Source: §4.3 · p. 7

4.3.1 Rollout Speed Optimization

  • Rollout generation consumes more than 70% of a training step, motivating static shuffling and a detached, metric-aware streaming load balancer that reroutes long sequences.The detached design separates rollout logic from the inference engine and is presented as suitable for future agent and tool-use scenarios.Source: §4.3.1 · pp. 7–9

5 Experiments

The experiments evaluate AM-Thinking-v1 across mathematics, coding, and general chat, using standardized protocols and comparisons with strong open-source and proprietary baselines. Results show competitive reasoning at 32B scale, alongside a supervised-fine-tuning pattern shift toward shorter, more task-aligned generations.

5.1.2 Evaluation Methodology

  • Evaluations use standardized generation settings, benchmark-specific response counts, a consistent system prompt, and comparisons with leading proprietary and open-source baselines.AIME uses 64 responses per query, LiveCodeBench uses 16, and Arena-Hard uses one response judged by GPT-4 Turbo.Source: §5.1.2 · p. 9, §5.1.3 · pp. 9–10, §5.1.4 · p. 10

5.2 Results

  • 85.3 on AIME2024 and 74.4 on AIME2025 outperform or closely match larger models such as DeepSeek-R1 and Qwen3-235B-A22B.Source: §5.2 · p. 10, Table 1 · p. 10
  • 70.3 on LiveCodeBench surpasses DeepSeek-R1 (64.3), Qwen3-32B (65.7), and Nemotron-Ultra-253B (68.1).Source: Table 1 · p. 10, §5.2 · p. 11
  • 92.5 on Arena-Hard is competitive with OpenAI-o1 (92.1) and o3-mini (89.0) but trails Qwen3-235B-A22B (95.6).Source: §5.2 · p. 11
  • Figure 6 compares model size with AIME2024 and LiveCodeBench scores, showing AM-Thinking-v1 as strongest among similarly sized dense models and close to much larger MoE models.Source: §5.2 · p. 11, Figure 6 · p. 11

5.3 Supervised Fine-Tuning Pattern Shift

  • Long-form reasoning SFT requires a larger learning rate and batch size than traditional SFT, while generation length decreases and stop ratio increases during training.The observed metric trends are interpreted as increasingly coherent and task-aligned reasoning behavior.Source: §5.3 · pp. 11–12, Figure 7 · p. 11, Figure 8 · p. 12

6 Conclusion and Limitations

AM-Thinking-v1 demonstrates that a 32B dense model can achieve strong reasoning performance through open-source data and carefully designed post-training. The authors position this scale as a practical balance between deployment efficiency and reasoning performance, while noting important capability limitations.

  • AM-Thinking-v1 achieves state-of-the-art reasoning capabilities among open-source models of comparable size and approaches top-tier MoE models.Source: §6 · p. 12
  • Systematic data preprocessing, ground-truth verification, SFT, and RL elicit advanced reasoning capabilities from a moderately sized model.Source: §6 · p. 12
  • 32B-scale dense models remain a practical and competitive choice, balancing deployment efficiency with reasoning performance.Source: §6 · p. 12
  • The model lacks structured function-calling, tool use, and multimodal inputs, limiting applicability in agent-based or cross-modal scenarios.Source: §6 · p. 12
Loading 2505.08311v2…