A practical rule for scaling pixel-wise Earth models, tested across 395 runs

One-Line Summary

Today’s papers shift from “bigger is better” to “measure and allocate smartly”: practical scaling rules for pixel-wise Earth models, process-first testing for coding agents, steadier agent training, and external-knowledge language models — alongside PyTorch’s steady release cadence.

Research Papers

TESSERA v2: scaling rules for pixel-wise Earth models

This study gives a practical recipe for getting better real-world results from pixel-wise Earth-observation models by showing how to spend training compute and how to select checkpoints. The authors run 395 controlled training jobs on 1,024 GH200 superchips within a fixed pixel-wise Barlow Twins setup and score every model on 15 downstream tasks; they find pretraining loss barely predicts downstream performance (absolute Pearson correlation under 0.2), so picking models by loss wastes compute. ¹

They propose a simple allocation rule: as your budget grows, scale the encoder and the data together while keeping the projector fixed. Following this, they train a family of models (0.5B and 1B parameters, with a 2B model in progress) and then distill them into compact “students” suitable for embeddings-as-data deployment. ¹

A 21-million-parameter distilled student (TESSERA v2-1B-M) outperforms all open and proprietary baselines tested, including models orders of magnitude larger. Their Matryoshka representations are cheap to serve: a 16-dimensional prefix retains 92% of the full 128-dimensional performance while using one-eighth the storage. ¹

The team plans to release v2 global embeddings covering 2017–2025 and to publish code, giving practitioners a concrete path: train large encoders, select by downstream evaluations, and distill into flexible students. ¹

AgentLens evaluates coding agents on their full trajectories

AgentLens is a benchmark that judges interactive coding agents by the entire interaction, not just pass/fail. It pairs formal verification (when an objective check exists) with Large Language Model (LLM)-written trajectory reviews and side‑by‑side comparisons, so each run yields a readable explanation of why it scored as it did. ²

Beyond leaderboard ranking, the benchmark helps teams diagnose behavior, compare successive agent versions, and catch product regressions in a nightly evaluation pipeline; the authors release it as open source. ²

Single-rollout async RL improves agent training stability

This paper introduces a training method that updates agent models as experiences arrive without waiting for synchronized batches, while keeping learning stable. The authors target Reinforcement Learning (RL) used in post-training Large Language Models (LLMs) and address a mismatch between asynchronous training and Group Relative Policy Optimization (GRPO). ³

Their Single-rollout Asynchronous Optimization (SAO) replaces group-wise sampling with one rollout per prompt, augments it with practical value‑model training designs, and adds strict double‑side token‑level clipping. SAO trains stably for 1,000 steps and consistently outperforms GRPO and variants on agentic coding and reasoning benchmarks, including SWE‑Bench Verified, BeyondAIME, and IMOAnswerBench. ³

SAO also proves effective in a simulated online learning setting and is deployed in the agentic RL pipeline for training the open GLM‑5.2 model (750B‑A40B), pointing to a simpler recipe for long‑horizon agent tuning. ³

CO-LMLM fetches facts with continuous vector queries

CO‑LMLM is a language model that keeps factual knowledge in an external Knowledge Base (KB) and fetches it with continuous vector keys instead of memorizing everything in weights. This “Limited Memory Language Model (LMLM)” setup generates flexible vector queries while still integrating human‑readable, attributable knowledge into the output. ⁴

An annotation pipeline tags free‑form factual spans beyond Wikipedia, enabling pretraining on Wikipedia and FineWeb‑Edu. Across scales, CO‑LMLM beats prior LMLMs and vanilla Large Language Models (LLMs) on perplexity and factual precision; at 360M parameters, it even achieves lower perplexity than models trained on forty times more data. ⁴

On SimpleQA, its performance aligns with gpt‑4o‑mini and surpasses Claude Sonnet 4.5, suggesting externalized knowledge can match small commercial assistants while remaining transparent and controllable. ⁴

Open Source & Repos

PyTorch 2.13.0 ships with updates and fixes

PyTorch, a Python library for tensor computation and dynamic neural networks with Graphics Processing Unit (GPU) acceleration, publishes version 2.13.0 dated 2026-07-08. The project highlights its two pillars: fast tensor ops and a tape‑based autograd system. ⁵

The 2.13.0 release notes enumerate highlights, deprecations, new features, performance work, and bug fixes, and developers can monitor trunk health at hud.pytorch.org. For many teams, this cadence signals a stable base for experimenting with the new training and evaluation ideas above. ⁵

Why It Matters

The TESSERA v2 scaling study offers a concrete, data‑backed rule for Earth‑observation models: grow encoders and data together, don’t overindex on loss, and compress into “students” for cheap serving — a playbook smaller teams can adopt without exotic infrastructure. ¹

In parallel, better measurement and training tools — process‑aware agent evaluations and steadier asynchronous RL — plus models that fetch facts from an external KB, point to a maturing stack where accuracy and efficiency come from evaluation design and compute allocation, not just model size. ⁴