Skip to main content
Back

Deepmind's Breakthrough: A Model That Wins Gold in Math Olympiads

CODE

January 21, 2024 4 min read

Hey 

Welcome back to AlphaSignal, where we bring you the latest developments in the world of AI.


In the past few days 1,500+ AI papers have been released, and among them, our model has picked the six.

Lets now dive straight into the papers!

Lior

IN TODAY'S SUMMARY

  • AlphaGeometry

  • Self-Rewarding Language Models

  • Tuning Language Models by Proxy

  • Other notable papers

Reading time: 4 min 05 sec

TOP PUBLICATION

AlphaGeometry: Solving Olympiad Geometry without Human Demonstrations

Trieu H. Trinh, Yuhuai Wu, Quoc V. Le, He He, Thang Luong 

Code

What's New

Google DeepMind's new model can solve geometry problems at the level of a top high school student in the International Math Olympiad. It combines a Large Language Model (LLM) with a symbolic model to bridge the gap between pattern recognition and rule-based problem-solving.


It employs a neuro-symbolic approach and is trained on a large synthetic dataset.

Problem

Geometry problems pose a unique challenge for AI due to their complex reasoning requirements and the difficulty in translating human proofs into a machine-understandable format. This has led to a scarcity of training data for AI models in this domain.

Solution

The team developed AlphaGeometry, which synthesizes millions of theorems and proofs, creating a vast pool of synthetic training data. The model combines a neural language model with a symbolic deduction engine, enabling it to navigate through infinite branching points in complex problems and generate human-readable proofs.

Results

In a benchmark test comprising 30 recent Olympiad-level problems, AlphaGeometry successfully solved 25, significantly outperforming previous methods. It demonstrated the ability to solve all geometry problems from the 2000 and 2015 IMO, equivalent to earning a bronze medal, establishing a new standard in AI's capability in mathematical reasoning.

Read more

Latitude.sh's brand new Container Solution is now open for early access

Effortlessly train, fine-tune and run inference on your Machine Learning models.


Reduce GPU compute complexity and costs with fast cold-boots, fully dedicated GPUs and on-demand billing.


Launchpad runs on Latitude.sh's enterprise-grade bare metal infrastructure allocating fully dedicated GPUs to your containers, ensuring total scalability and amazing performance with hourly billing.


Register to the early access and enjoy on-demand customized resources and full scaling elasticity based on your data traffic.

Join early access

TOP PUBLICATION

Self-Rewarding Language Models

Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sainbayar Sukhbaatar, Jing Xu, Jason Weston

What's New

Self-Rewarding Language Models (Self-Rewarding LMs) show remarkable performance by using an LLM-as-a-Judge mechanism during Iterative DPO training, outpacing models like Claude 2 and GPT-4 0613 in AlpacaEval 2.0.

Problem

The goal is to surpass human-level performance in language models by overcoming the limitations of fixed, human-crafted reward models and enabling continuous self-improvement in instruction following and reward modeling abilities.

Solution

Employ LLM-as-a-Judge to enable Self-Rewarding LMs to assign and improve rewards autonomously. Utilize Iterative DPO for training, allowing the model to refine instruction-following and reward-modeling capabilities across iterations. Fine-tune Llama 2 70B over three such iterations for enhanced performance.

Results

After three rounds of Iterative DPO training, the refined Llama 2 70B outperforms established AI systems like Claude 2, Gemini Pro, and GPT-4 0613, showcasing the potential of self-rewarding mechanisms for continual, autonomous improvement in language models.

Read the paper

TOP PUBLICATION

Tuning Language Models by Proxy

Alisa Liu, Xiaochuang Han, Yizhong Wang, Yulia Tsvetkov, Yejin Choi, Noah A. Smith

What’s New

Proxy-tuning significantly narrows the performance gap between large and smaller models, achieving up to 88% effectiveness compared to full model tuning, by adjusting output logits at decoding-time.

Problem

Finetuning large language models (LLMs) is resource-intensive and often impractical when model weights are inaccessible. There's a need for a less resource-heavy, more accessible method to adapt LLMs to specific tasks or domains.

Solution

Implement proxy-tuning by training a smaller LLM and using the logit differences between this trained model and its untrained counterpart to adjust the predictions of a larger, black-box LLM. This method modifies output logits at decoding-time, leveraging the smaller model’s adaptability to guide the larger model.

Results

Applied to Llama2-70B, proxy-tuning with a 7B model closed 88% of the performance gap to its fully-tuned counterpart, showing particular strength in benchmarks involving knowledge, reasoning, and safety. Notably, on TruthfulQA, proxy-tuned models outperformed directly-tuned models, suggesting retention of factual knowledge through decoding-time guidance.

Read the paper

NOTABLE PAPERS

OMG-Seg: Is One Model Good Enough For All Segmentation?

OMG-Seg is a unified framework handling ten segmentation tasks, outperforming task-specific models in efficiency, reducing parameter size and computational needs. It employs a transformer-based architecture, shared decoder, and innovative query representation.

Code Generation with AlphaCodium: From Prompt Engineering to Flow Engineering

AlphaCodium, an iterative, test-based code generation flow, enhances GPT-4's coding accuracy from 19% to 44% on the CodeContests dataset. This approach, available as open-source, refines code solutions through a multi-stage process, significantly outperforming standard direct prompts.

Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model

Vision Mamba (Vim) introduces a bidirectional state space model for efficient visual representation, outperforming DeiT with 2.8x speed and saving 86.8% GPU memory for high-res image tasks. Open-source code available, demonstrating notable improvements in ImageNet, COCO, and ADE20k benchmarks.

Not Great   Good    Amazing

Want to promote your company, product, job, or event to 150,000+ AI researchers and engineers? You can reach out here.

Thank You