One-Line Summary

Robot learning is cohering around two paths — test-time fixes on frozen policies and reusable code skills — while new work shows how to compress skills, compile on-screen routines into agent memory, and benchmark data agents, plus a purpose-built Yiddish model.

Research Papers

Robot learning splits into ‘weights vs skills’

A new survey says robot learning is coalescing around two approaches: either embed behavior in frozen weights of vision-language-action (VLA) policies, or have agents write and refine executable skills as code. The authors examine 77 representative systems across six technique families and define self-improvement mechanisms from zero-shot program synthesis to closed-loop skill memory with evolutionary search, highlighting that only a few recent systems combine execution feedback, skill memory, and evolutionary search in one open-ended loop. They also map today’s “skill economy” and its open problems. ¹

On the “weights” side, Retrieve in Time, Correct in Frequency (RTCF) is a training-free, test-time correction for frozen VLA policies. Across four LIBERO suites and 2,000 episodes per condition, RTCF raises aggregate success from 86.4% to 88.4% and lifts LIBERO-Long from 61.6% to 68.6% by progressively aligning an episode with a successful memory and transferring only a clipped low-frequency residual of the retrieved motion; it needs no parameter updates, no repeated policy inference, and no extra graphics processing unit (GPU) resources, and runs on the client central processing unit (CPU) with a median 10.99 ms overhead per action chunk. ²

On the “skills” side, SkillZip compresses agent skill libraries while preserving each procedure’s executable contract. It rewrites recurring section-level motifs into reversible macros that retain boundary signatures, dependency closure, and verifier reachability; experiments report up to 12.2-point gains over the strongest baseline, a 3.46× compression ratio, 99.2% dependency preservation, and 98.7% verifier reachability, with robust retrieval from libraries spanning 200 to 100K skills. ³

The survey connects these techniques to emerging robot-skill marketplaces: many distribute one-tap skills as static playback, surfacing gaps in adaptation, cross-embodiment portability, provenance, safety verification, composition, and standardization. Together, the taxonomy and exemplars frame a near-term choice for teams between enhancing frozen policies with test-time memory and investing in executable, contract-bearing skills with ongoing maintenance. ¹

Activity Frames compiles screen activity into deterministic agent memory

This paper turns passively captured screen activity into typed, replayable “activity frames” so computer-use agents can reuse routines without paying full frontier inference. On a single-user corpus of 128,756 frames over 51 active days, the compiler reduces a day of raw capture to a prompt-ready context block 86× smaller in 68 ms, and an agent reading that block answers questions about the day at 98.4% accuracy (Wilson 95% CI 91.7–99.7%), versus 66–80% for a large language model (LLM) summary. ⁴

The same compiler measures demand-side costs before delegation, reporting a Routine Overhead Ratio R of 60–343× and delegable routine recurrence of 9.0% in-sample and 7.7% out-of-sample, implying a realistic all‑fleet token ceiling near 8%. Compiled routines replay deterministically with the model out of the loop at zero tokens, and the schema, compiler, and evaluation harness are open. ⁴

DataSpace tests data agents on verifiable analytics across mixed files

DataSpace asks an agent to produce complete, verifiable tables by discovering and joining evidence in a task-local workspace spanning CSV, JSON, SQLite, Markdown, PDF, and video. It contains 410 cross-language tasks and 7,439 artifacts totaling 15.01 GB, and it served as the official evaluation benchmark for the KDD Cup 2026 Data Agents for Complex Data Analysis competition. ⁵

A deterministic evaluator handles header-invariant alignment, type- and precision-aware normalization, and order-aware row comparison. Across six recently released frontier multimodal models and five widely used agent harnesses, the best accuracy reaches 66.34%, while the harness alone creates a 15.36‑point spread with the backbone fixed; integrating multimodal evidence and performing joins consistently reduces accuracy across all six backbones, indicating the benchmark remains unsaturated. ⁵

LLM & SOTA Models

MameLoshnLM: an 8B Yiddish model with its own corpus and benchmark

MameLoshnLM is the first open-source 8‑billion‑parameter large language model (LLM) built specifically for Yiddish, obtained by continuing pretraining of Llama 3.1 8B on a curated corpus (Oytser) and evaluating on a new multi‑task benchmark (Kashes). Across benchmark tasks, it outperforms open baselines of similar scale. ⁶

The authors report that gains are qualitative as well as quantitative: relative to general-purpose multilingual models, MameLoshnLM better captures language‑defining lexical and morphological patterns. They argue this exposes a broader failure mode of noisy web‑scale multilingual data for low‑resource languages and offer a practical template for building models in historically rich but digitally underrepresented languages. ⁶

Open Source & Repos

Ray: a distributed engine for scaling AI workloads

Ray is an AI compute engine that packages a core distributed runtime with AI libraries aimed at accelerating machine‑learning workloads. The project’s repository highlights documentation, a community Slack, and discussion links to help users get started. ⁷

For teams exploring distributed training or parallel compute for AI, Ray provides a unified runtime plus libraries rather than an ad‑hoc stack of scripts and services. The repo is the canonical entry point for installation, examples, and community support. ⁷

Why It Matters

Robot learning is moving from “train more, then deploy” toward “reuse what worked” — either by correcting frozen policies at test time with memory or by packaging procedures as executable, verifiable skills. This lowers retraining risk and shifts effort to retrieval, compression, verification, and maintenance — areas the new survey and papers lay out with concrete mechanisms and numbers. ¹

Meanwhile, practical tooling is catching up: deterministic activity-memory compilers and unified data‑agent benchmarks make agent costs and errors measurable, and specialized models like a Yiddish 8B show when domain‑specific data beats general multilingual corpora. Together, these trends point to more auditable, efficient agent stacks. ⁵

This Week, Try

1. RTCF in pictures: skim Figures and Method to see progress‑aware retrieval and frequency correction in action (arXiv). ²
2. Explore Ray’s README: browse the repo to see how the runtime and libraries are organized before trying a small job. ⁷