This week in deep learning, we bring you Opik’s Guardrails, Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning, and a paper on DeepEyes: Incentivizing “Thinking with Images” via Reinforcement Learning.
You may also enjoy Codestral Embed, a paper on Emerging Properties in Unified Multimodal Pretraining, and more!
As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.
Until next week!
Industry
Announcing Opik’s Guardrails Beta: Moderate LLM Applications in Real-Time
Comet unveiled a public beta of Guardrails, a lightweight system designed to help developers build safer, more robust LLM applications.
The Mistral team released Codestral Embed, their first embedding model specialized for code.
Security startup Horizon3.ai is raising $ 100M in new round
Horizon3.ai, a cybersecurity startup that provides tools like autonomous penetration testing, is raising $ 100 million in a new funding round.
Google CEO Sundar Pichai says AI will be ‘bigger than the internet’
The head of Google discusses the next AI platform shift and how it could change how we use the internet forever.
LMArena raises $ 100M at $ 600M valuation to expand AI benchmarking platform
LMArena, the company behind Chatbot Arena, has raised $ 100 million in initial funding.
MLOps & LLMOps
Architecting a Multi-Agent System with Google A2A and ADK
A post that explores the architectural concepts behind ADK and A2A using a trading simulator as an example, focusing on how these technologies enable effective multi-agent systems.
How I built an agent with Pydantic AI and Google Gemini
A blog post that walks you through the process of building a SWOT analysis agent using Pydantic AI and Gemini.
CodeAgents + Structure: A Better Way to Execute Actions
An article demonstrating that forcing AI agents to generate thoughts and code in a structured JSON format significantly improves performance and reliability on various benchmarks.
Learning
Fine-tuning LLMs with user-level differential privacy
Google researchers investigate and improve algorithms for fine-tuning large models with user-level differential privacy.
Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning
An article highlighting that current approximate machine unlearning methods for LLMs primarily obfuscate information, making them vulnerable to benign relearning attacks.
A blog post that examines some latest developments in image generation methods that inherently support multi-scale generation, including AR-Transformers and hybrid approaches.
The Misguided Quest for Mechanistic AI Interpretability
An article arguing that mechanistic AI interpretability has largely failed to provide useful insights over a decade of research, and proposing a shift towards a top-down approach.
Reinforcement learning with random rewards actually works with Qwen 2.5
An article presenting research findings that Reinforcement Learning with Verifiable Rewards (RLVR) applied to Qwen 2.5 models can boost performance on math tasks even with random rewards.
Exploring Quantization Backends in Diffusers
An article exploring various quantization backends available in Hugging Face Diffusers, including bitsandbytes, torchao, Quanto, GGUF, and FP8 Layerwise Casting, to reduce the memory and compute requirements of large diffusion models.
Libraries & Code
An open-source LLM evaluation tool used to debug, evaluate, monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.
A lightweight, high-performance library for building agents
Papers & Publications
Emerging Properties in Unified Multimodal Pretraining
Abstract:
Unifying multimodal understanding and generation has shown impressive capabilities in cutting-edge proprietary systems. In this work, we introduce BAGEL, an open-source foundational model that natively supports multimodal understanding and generation. BAGEL is a unified, decoder0only model pre-trained on trillions of tokens curated from large-scale interleaved text, image, video, and web data. When scaled with such diverse multimodal interleaved data, BAGEL exhibits emerging capabilities in complex multimodal reasoning. As a result, it significantly outperforms open-source unified models in both multimodal generation and understanding across standard benchmarks, while exhibiting advanced multimodal reasoning abilities such as free-form image manipulation, future frame prediction, 3D manipulation, and world navigation. In the hope of facilitating further opportunities for multimodal research, we share the key findings, pre-training details, data creation protocol, and release our code and checkpoints to the community.
DeepEyes: Incentivizing “Thinking with Images” via Reinforcement Learning
Abstract:
Large Vision-Language Models (VLMs) have shown strong capabilities in multimodal understanding and reasoning, yet they are primarily constrained by text-based reasoning processes. However, achieving seamless integration of visual and textual reasoning which mirrors human cognitive processes remains a significant challenge. In particular, effectively incorporating advanced visual input processing into reasoning mechanisms is still an open question. Thus, in this paper, we explore the interleaved multimodal reasoning paradigm and introduce DeepEyes, a model with “thinking with images” capabilities incentivized through end-to-end reinforcement learning without the need for cold-start SFT. Notably, this ability emerges natively within the model itself, leveraging its inherent grounding ability as a tool instead of depending on separate specialized models. Specifically, we propose a tool-use-oriented data selection mechanism and a reward strategy to encourage successful tool-assisted reasoning trajectories. DeepEyes achieves significant performance gains on fine-grained perception and reasoning benchmarks and also demonstrates improvement in grounding, hallucination, and mathematical reasoning tasks. Interestingly, we observe the distinct evolution of tool-calling behavior from initial exploration to efficient and accurate exploitation, and diverse thinking patterns that closely mirror human visual reasoning processes.
