One-Line Summary

Structure-aware AI takes the spotlight: mapping narratives in training data, organizing long-document retrieval and agent memory, and speeding high-fidelity avatars — plus a practical coding agent update.

Research Papers

Researchers chart narratives across web-scale training data

This paper analyzes the story-like qualities — who does what, where — inside the massive web text used to train large language models (LLMs). Studying Dolma, a 3-trillion-token corpus, the authors design an 11-dimension framework spanning agency, setting, and events, train a predictor called NarraBERT, and release a large annotated dataset, NarraDolma. ¹

The team samples and labels 400 passages in detail, then fine-tunes a RoBERTa-based classifier to predict fine-grained narrative features and applies it to 3 million passages. They report that narrative structure is measurable at scale, forms a continuous multidimensional space, and is unevenly distributed across sources and topics — patterns that current curation practices do not capture. ¹

Why it matters: NarraBERT and NarraDolma offer a concrete way to gauge narrative composition in pretraining data, laying groundwork to study how data choices affect narrative reasoning tasks and to add narrative-aware checks to data pipelines. The authors publicly release both resources. ¹

SproutRAG improves long-document retrieval with learned sentence trees

SproutRAG makes Retrieval-Augmented Generation (RAG) better at handling long documents by organizing sentences into a binary chunking tree built from learned inter-sentence attention, so systems can retrieve at multiple granularities without extra large language model calls or lossy summaries. ²

Trained end-to-end, the method learns which attention heads capture document structure and uses hierarchical beam search at query time. Across four benchmarks (scientific, legal, open-domain), it improves information efficiency (IE) by 6.1% on average over the strongest baseline; the authors say code is available. ²

MemSlides separates long-term and working memory for slide agents

MemSlides is a presentation-generation agent that cleanly separates persistent user preferences (long-term memory) from session-level constraints (working memory) and reusable tool know-how (tool memory), and applies slide-local edits instead of regenerating whole decks. ³

In controlled studies, user profile memory improves persona alignment, injecting tool memory boosts closed-loop modify behavior, and qualitative cases show preferences carrying over across revision rounds — suggesting personalization hinges on distinct memory types and scoped, local updates. ³

SpatialAvatar-0 builds high-quality 4D head avatars faster

SpatialAvatar-0 produces high-quality 4D head avatars from one or a few portraits using 3D Gaussian Splatting (3DGS), bridging feed-forward predictors and per-subject refiners under a shared, mesh-bound Gaussian representation. ⁴

A two-phase schedule (monocular-temporal to multi-view-spatial) with a parameter-free K-source mean-pool helps prevent identity collapse, while a 10K-iteration per-subject refinement loop preserves the upstream layout and replaces densification with anti-spike regularization — cutting iteration counts versus common 300K–600K baselines. ⁴

On cross-domain zero-shot tests, the paper reports +1.5 dB Peak Signal-to-Noise Ratio (PSNR) over an in-domain leader (GAGAvatar), and on a monocular benchmark it leads all reported metrics, beating GeoAvatar by +1.3 dB PSNR with up to 60× shorter per-subject schedules. ⁴

Open Source & Repos

Forge: AI-enhanced terminal coding agent supporting 300+ models

Forge is an AI-enabled pair programmer that runs in the terminal and integrates with development workflows, connecting to Claude, GPT, O Series, Grok, Deepseek, Gemini and more than 300 models. The README describes it as a comprehensive coding agent that augments your environment. ⁵

Release v2.13.13 (Jun 22, 2026) ships a provider fix to sync Fireworks AI models with the current serverless catalog and updates Rust dependencies — an incremental but practical maintenance drop. The project also highlights a one-line installer in the README for quick setup. ⁵

Why It Matters

Across these papers, the throughline is structure: measuring narratives in training data gives builders a way to check “who/where/what happens” balance before training, hierarchical retrieval reduces wasted context in long-doc Q&A, and separating agent memories keeps personalization stable without full rewrites. For learners, the key terms to keep: agency/setting/events, hierarchical retrieval, and local edits. ¹

On the media side, SpatialAvatar-0 shows how a shared representation and tighter refinement loops can both improve quality and cut compute, while tooling like Forge reflects the steady, practical evolution of coding agents that track model catalogs. ⁴