Seed1.5-Thinking: Advancing Superb Reasoning Models with Reinforcement Learning (ByteDance Seed, April 2025)

Seed1.5-Thinking: Advancing Superb Reasoning Models with Reinforcement Learning is a technical report by the Seed research group at ByteDance, posted to arXiv as 2504.13914 in April 2025 (the captured version is v3, dated April 29, 2025) (Source: https://arxiv.org/abs/2504.13914). It describes Seed1.5-Thinking, a mixture-of-experts reasoning model with 20B activated and 200B total parameters that "reasons through thinking before responding", and is the research paper behind the Seed-Thinking entry in the Doubao family page. The report links a model trial on Volcano Engine and does not itself use the Doubao product name. Its content is a description of the RL training data, reward modelling, algorithms and infrastructure, followed by automatic and human evaluation results; it reports no safety or refusal evaluations.

Training approach

The report frames three ingredients — data, RL algorithm and RL infrastructure. RL data is split into verifiable problems (STEM questions with answers, coding problems with unit tests, and logic puzzles) and non-verifiable problems (creative writing, translation, knowledge QA, role-play and similar tasks judged by human preference), the latter drawn from the RL prompts used for Doubao-1.5 Pro and filtered to remove low-variance and low-difficulty prompts. Because the 30 annual AIME problems give "high-variance" results, the authors construct BeyondAIME, an internal set of expert-written problems designed to resist memorisation, and an internal Codeforces benchmark built from the twelve most recent contests, both of which they say will be released. Reward modelling uses two verifiers for verifiable problems, Seed-Verifier and a chain-of-thought Seed-Thinking-Verifier, which the report scores at 82.7% and 99.3% on a 456-item human-labelled test set of hard cases; non-verifiable problems use a pairwise preference reward model. The RL stage combines verifiable, general and hybrid data under one framework and applies the group's VAPO (actor-critic) and DAPO (policy-gradient) algorithms with techniques including value pre-training, decoupled GAE, clip-higher, token-level loss and dynamic sampling. Section 5 describes a streaming rollout system and training system for long chain-of-thought generation.

Evaluation

Table 2 compares Seed1.5-Thinking with DeepSeek R1, OpenAI o3-mini (called o3-mini-high in the text), Grok 3 Beta and Gemini 2.5 Pro on AIME 2025 and 2024, BeyondAIME, GPQA diamond, SuperGPQA, MMLU-Pro, the internal Codeforces benchmark (avg@8 and pass@8), LiveCodeBench v5, Aider Polyglot, SWE-bench Verified (as reported and as re-run in the authors' sandbox), ARC-AGI, SimpleQA, Collie and IFEval. Mathematics results average 32 responses, GPQA averages 8, and other tasks use one response. Seed1.5-Thinking is reported at 86.7 on AIME 2024, which the report says matches o3-mini-high, 74.0 on AIME 2025, 77.3 on GPQA diamond, 55.0 pass@8 on Codeforces, 64.9 on LiveCodeBench v5, 47.0 on SWE-bench Verified and 39.9 on ARC-AGI, the last being the highest in that row; the authors state that the model lags o3-level performance on AIME 2025 and BeyondAIME, "nearly matches" Gemini 2.5 Pro on Codeforces, and scores 12.9 on SimpleQA, which they characterise as a memory-oriented metric tied to pre-training scale. The sources of the comparison figures are not stated. In a human evaluation of non-reasoning tasks against DeepSeek R1 by domain-expert raters on a 0–4 scale, the report gives an "overall win ratio of 8.0%", phrased in the abstract as surpassing R1 "by 8% in win rate". Ablations show a rejection-fine-tuned initialisation converging faster but ending lower on AIME, and consistent VAPO-over-DAPO rankings across a Qwen-32B dense model and a Seed-150B MoE.

Reported results

Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.

Provenance

The page summarises the arXiv HTML (v3, April 29, 2025) as captured on September 21, 2026. All figures are reported by ByteDance Seed; comparison figures are unattributed except where marked as the authors' own sandbox runs. Confidence is medium: a single developer-authored source with fast-decay benchmark claims and no third-party evaluation.

Relationships

• supports: Doubao
• instance-of: AI System Cards
• related: ByteDance / Doubao, AIME (as a model benchmark), GPQA, SWE-bench, ARC-AGI, LiveCodeBench, DeepSeek-R1, Gemini 2.5 Pro