Skip to main content
Back

DeepSeek trains math proof model without human data

TRY NOW

May 2, 2025 5 min read

Hey,

Welcome to AlphaSignal – the most read newsletter by AI developers. 


We bring you the top 1% of news, papers, models, and repos, all summarized to keep you updated on the latest in AI.

IN TODAY'S SIGNAL

Read time: 5 min 56 sec

🎖️ Top News

⚡️ Trending Signals

📄  Top Papers

  • The Leaderboard Illusion: Private testing and unequal access skew Chatbot Arena rankings, exposing bias in LLM benchmark evaluations.


  • SkyReels-V2: Generates infinite-length, cinematic-quality videos using diffusion models, MLLMs, and RL for motion coherence.


  • Paper2Code: Turns ML papers into runnable code using a multi-agent LLM system with planning, analysis, and generation stages.

🧠  Deep Dive

If you're enjoying AlphaSignal please forward this email to a colleague. 

It helps us keep this content free.

TOP NEWS

AI Model

DeepSeek unveils 671B open-source model for formal theorem proving without labels

⇧ 3,239 Likes

What's New

DeepSeek has released Prover-V2, a formal theorem-proving model built on DeepSeek-V3. It generates machine-checkable proofs in Lean by decomposing problems into subgoals and resolving them with reinforcement learning. Prover-V2 achieves 88.9% on MiniF2F and solves 49 out of 658 problems in PutnamBench.


Cold-start pipeline

The training pipeline requires no external formal corpora or human-labeled data.

  • Uses DeepSeek-V3 to generate informal reasoning and decompose problems into subgoals

  • A 7B model searches proofs for each subgoal to reduce compute costs

  • Valid subgoal proofs combine into full formal proofs

  • Matches each informal chain with its formal proof for supervised fine-tuning

  • Enables cold-start learning without manual proof annotation

Reinforcement Learning
A second stage uses binary feedback to improve formal reasoning ability.

  • Fine-tunes on synthetic proofs where all subgoals are resolved

  • Uses correct/incorrect signals as reward supervision

  • Optimizes for alignment between informal reasoning and formal verification

  • Trains on unsolved problems that allow complete proof via subgoal composition

New Benchmark: ProverBench

ProverBench spans math competition and undergraduate-level problems in formal Lean 4 format.

  • 15 problems from AIME 24 and 25

  • 90 problems from calculus and 50 from linear algebra

  • 40 problems each from number theory and abstract algebra

  • Remaining problems from analysis, probability, and functional analysis

Access and Availability

DeepSeek-Prover-V2 is available in two sizes and comes with an open-source release.

  • You can download Prover-V2-7B and Prover-V2-671B from Hugging Face.

  • The 7B version uses a 32K context and builds on Prover-V1.5-Base.

  • The 671B version is trained on DeepSeek-V3-Base and offers higher performance.

  • The ProverBench dataset with 325 formalized problems is also available for download.

  • The release includes proof outputs for the MiniF2F benchmark as a ZIP archive.

Community Feedback


Xythera

"Just tried DeepSeek-Prover V2. It solved the problem before I even knew what it was asking. Finally, a model that proves faster than I can read. This is intelligence."

Synaptic Solo
"This will eventually lead to superhuman math AI that will turbocharge AI research. Here’s how: Optimizes algorithms for faster, smarter models. Unlocks insights into AI’s “black box” behavior. Designs better architectures without endless trial and error. Speeds up data analysis for quicker breakthroughs. In other words, this is what leads to AGI/ASI. In couple of years it will produce math no human will comprehend."

Alex Mirugwe
"
Strong results, but how far can scaling get us in formal math without deeper reasoning breakthroughs? Are we hitting a ceiling where models memorize structure but don’t really "understand"?"

Try now

TRENDING SIGNALS

AI Assistant

Anthropic presents Claude Integrations, enabling custom app connections in under 30 minutes using APIs like Zapier and PayPal

⇧ 3,359 Likes

Model Behaviour & Safety

OpenAI plans customizable personalities and real-time feedback to prevent future misalignment and better user control

⇧ 8,639 Likes

LLM Evaluation

Researchers from MIT, Stanford present paper which exposes flaws in LM Arena’s top model evaluation methodology

⇧ 4,927 Likes

LLMs

Microsoft releases Phi-4-reasoning: 14B open-weight models reaching SOTA on math, coding, and planning tasks

⇧ 840 Likes

Generative AI

Google rolls out native image editing in Gemini with SynthID watermark and support in 45+ languages

⇧ 2,438 Like

TOP PAPERS

LLM Evaluation

The Leaderboard Illusion

⇧ 4,927 Likes

Problem
Chatbot Arena rankings are skewed by private testing and selective score submissions, distorting public perception of LLM quality.

Solution
The paper reveals that companies like Meta, OpenAI, and Google benefit from early access and higher sampling rates. Open-weight models, by contrast, are underrepresented and more frequently removed.

Results
Private exposure leads to up to 112% relative gains, showing that leaderboard performance can be gamed. The authors call for reforms to ensure fairer, more transparent evaluations.

Video Generation

SkyReels-V2: Infinite-length Film Generative Model

⇧ 2,149 Likes

Problem
Current video generation models struggle with producing high-quality, coherent long-form content due to limitations in prompt adherence, visual fidelity, motion dynamics, and understanding of cinematic grammar.

Solution
SkyReels-V2 integrates a Multi-modal Large Language Model (MLLM), multi-stage pretraining, reinforcement learning, and a diffusion forcing framework. It introduces SkyCaptioner-V1 for detailed video annotations and employs progressive-resolution training alongside motion-specific reinforcement learning to enhance dynamic consistency.

Results
The model generates infinite-length videos with improved visual quality and motion coherence, effectively capturing cinematic elements like shot composition and camera movements.

Code Generation

Paper2Code: Automating Code Generation from Scientific Papers in Machine Learning

⇧ 2,847 Likes

Problem
Reproducing machine learning research is hindered by the frequent absence of accompanying code, making replication labor-intensive and slowing progress.

Solution
Paper2Code introduces PaperCoder, a multi-agent LLM framework that transforms ML papers into functional code repositories through three stages: planning (roadmap and architecture), analysis (implementation details), and generation (modular code).

Results
Evaluated on 90 papers, PaperCoder outperformed baselines, with 77% of authors preferring its outputs. It achieved a 44.26% replication score on PaperBench, significantly surpassing competitors.

DEEP DIVE

AI Agent

How to connect AI agents to third-party tools using MCP

⇧ 836 Likes

This lecture shows how to connect AI agents to third-party services like Gmail, GitHub, and ArXiv using the Model Context Protocol (MCP) and ACI.dev. You learn how to centralize authentication logic, skip writing separate auth flows, and give your agent seamless access to external APIs through a single MCP endpoint.

You will learn:

  • How to use ACI.dev to manage OAuth flows across multiple APIs

  • How to register and link tools like Gmail, GitHub, ArXiv on the ACI platform

  • How to set up the MCP server and connect it to Windsurf or any MCP-compatible agent

  • How to use unified secure function calls from your agent to trigger tool actions

Watch now

At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.


Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws.

Looking to promote your company, product, service, or event to 200,000+ AI developers? Let's work together.

Work with us