Berkeley AI Research introduces ABBEL, a framework that replaces full interaction history with natural-language belief states and supervises their contents via belief grading. On CollabBench collaborative coding, reconstruction-based belief grading cuts the gap to full-context models by about 50% and trains in 50 steps instead of 100, while using fewer peak tokens.
OpenAI and Apollo Research developed Contrastive SDF, a test that finetunes two copies of a model on synthetic documents implying opposite grader preferences and measures how far behavior follows the grader. Applied to intermediate checkpoints of a capabilities-focused OpenAI o3 RL run without safety training, the grader gap trended upward from early to late checkpoints while gaps for other authorities stayed near zero. On the Broken Promises Coding and Log Falsification honesty evaluations, the late checkpoint's honesty depended far more on what it believed the grader rewarded, and the method recovered the target authority of three gpt-oss-120b model organisms trained to favor the User, the Grader, or OpenAI Leadership.
Why it matters: OpenAI and Apollo Research introduce Contrastive SDF, a finetuning-based test that measures how strongly a model's behavior tracks what it believes its grader rewards, and find the tendency grows across an o3 RL run without safety training.
UC Berkeley's Aditya G. Parameswaran and collaborators argue that as inference costs fall from roughly $30 per million tokens in early 2023 to under $1 today, data systems must be redesigned around three challenges: systems for agents, of agents, and by agents. They note that in a text-to-SQL benchmark with multiple agents per task, only 10-20% of sub-plans are distinct, so 80-90% of sub-queries duplicate work, and propose reuse, approximation, and proactive data systems.
MiniMax published a technical blog explaining MaxProof, the framework behind M3 exceeding the human gold-medal threshold on IMO 2025 and USAMO 2026. The pipeline trains three expert models for proof generation, error judgment and proof repair, then runs an evolutionary test-time search that samples 32 candidates, verifies each 4 times and refines over up to 10 rounds. MiniMax says the key lesson is that a generative verifier used as an RL reward should prioritize low false positives over average accuracy, citing M2-cycle reward hacking where the false-positive rate rose from about 2.9% to 8.0% during training.
Modal says the bottleneck in RL post-training of LLMs is now infrastructure, not algorithms, and has released an open-source library drawn from running RL at scale. It details three pain points — glue code, cluster queueing and GPU under-utilization — and reports RDMA weight sync cutting Qwen3 8B update time from 2.62s to 41ms.
Applied Compute trains custom enterprise AI agents via reinforcement learning, building on Modal's sandboxes, Functions and sub-second cold starts to keep thousands of parallel rollouts GPU-bound. Its agents include a merchant onboarding model for DoorDash and a bug-catching agent for Cognition.
AE Studio details how it trained LLMs to prove math theorems with reinforcement learning on Modal, comparing Evolution Strategies against GRPO while keeping the training loop fixed. The setup runs vLLM inference on GPUs, Lean compiler verification in isolated CPU sandboxes, and a lightweight orchestrator, using per-function images, .map() fan-out, and Volumes for base weights.
Modal demonstrates running autoresearch agents on elastic GPUs: given Modal Skills, Claude Code ran 113 experiments over 238 GPU-hours on OpenAI's Parameter Golf challenge, finishing core training runs 5x faster than a single workstation while using a fraction of a dedicated cluster's resources. Agents provision GPUs via API calls, scaling from one to dozens of H100s and back to zero automatically.
Doppel migrated a large portion of its ML workflows to Modal, removing the infrastructure friction that slowed training experimentation and inference deployment. Parallel experiments via map() replaced sequential runs, while image layer caching and persistent volumes cut build times by up to 10x, with complex setups reaching sub-minute warm builds.
Berkeley AI Research researchers describe SPEX and ProxySPEX, algorithms that identify influential feature, data, and model-component interactions in LLMs at scales far beyond prior methods. SPEX reframes interaction search as sparse recovery using sparsity and low-degreeness, while ProxySPEX exploits hierarchy to match SPEX with about 10x fewer ablations. On sentiment analysis, SPEX matches Faith-Shap and Faith-Banzhaf faithfulness on short inputs and retains it as context scales to thousands of features.
Chai Discovery is using Modal to run its machine learning drug discovery pipelines, from protein embeddings and multiple sequence alignments to antibody design. Modal Volumes downloads and indexes datasets reaching hundreds of gigabytes once, then shares them across machines, while elastic GPU scaling spins up hundreds of GPUs in minutes. Chai says the same setup now carries research workflows into production without rewriting infrastructure.
Modal has scaled its multi-cloud GPU worker pool past 20,000 concurrent GPUs and launched over four million cloud instances in the last couple years, and is now sharing the reliability system behind it. The post covers instance type testing and selection, machine image preparation and boot checks, passive and active GPU healthchecking, and observability and support. It reports hyperscaler differences such as Cloud A's 99.6% boot success and Cloud C's H100s running above 90ºC in 2025.
Modal announced that Twirl, a data orchestrator for developing, testing and deploying data pipelines, is joining the company. Twirl co-founders Rebecka Storm and Eric Hansander will join Modal's Stockholm office, where they will develop and expand Modal's serverless infrastructure and grow its European footprint. Modal said the team's data product experience will help it expand beyond AI inference into adjacent AI, ML and data use cases.
Fine-tuning adjusts pretrained LLM weights to fit a specific task or domain, delivering higher quality than prompt engineering at lower cost and latency; a fine-tuned Llama 7B can be roughly 50 times more cost-effective per token than GPT-3.5 with comparable performance. The guide covers use cases such as structured output, style adherence and domain-specific instruction, plus options including OpenAI, Predibase, Modal, Google Colab and AWS SageMaker, and frameworks HuggingFace transformers, TRL and Axolotl.
A top-tier European soccer team, anonymized as AFC Richmond, moved its match tracking data pipeline to Modal's serverless GPU platform and cut the cost of processing a full season of games by 50%. The team runs inference on ~3.5 million tracking observations per game through a custom transformer-based model, and with Airflow on Modal now processes each game's data in minutes rather than hours.
Modal fine-tuned a Stable Diffusion model on the roughly 300-SVG Heroicons library to generate new black-and-white line icons in that style, including ones for concepts like barack-obama and golden-retriever. The full pipeline — downloading the pretrained model, full fine-tuning with HuggingFace Diffusers' train_text_to_image.py script, and serving a Gradio web UI — runs on Modal's serverless GPU platform. The write-up walks through dataset prep (SVG-to-PNG conversion, white backgrounds, caption metadata.csv) and notes that full fine-tuning worked best, though LoRA-style parallel adapters can also work for small datasets.
Substack has moved training and deployment for nearly all of its ML models—including spam detection, newsletter recommendations, audio transcription, sentiment analysis and image generation—from AWS SageMaker to Modal. Head of AI & ML Engineering Mike Cohen cited SageMaker's convoluted developer experience, hard-to-share code across notebooks and 5+ minute container startup times; on Modal, an inference endpoint took an hour to implement. Substack now runs a full fine-tune, validation and deployment workflow on Modal's storage primitives.
A fine-tuned open source embedding model can outperform OpenAI's text-embedding-3-small in accuracy at lower cost using only a handful of examples. Modal ran an experiment on the Quora duplicate-questions dataset to find how much fine-tuning data a simple question-answering use case needs, and describes how to build one on its serverless infrastructure.
Ramp used Modal to fine-tune its LLMs and build a text-to-structured-JSON model for receipt management, cutting receipts requiring manual intervention by 34% on infrastructure estimated 79% cheaper than major LLM providers like OpenAI. Modal functions trained many candidate models in parallel, persisted weights to Modal volumes, and served a parametrized inference endpoint. One engineer also ported a PII-stripping script for 25,000 invoices to Modal, parallelizing it on 256 cloud workers to finish in 20 minutes for $100 instead of 3 days locally.