New 3D method sharpens local geometry in point maps
SurGe introduces a local-surface metric and two training components to reduce visible micro-geometry errors in point maps while maintaining top global accuracy. Companion papers spotlight agent ‘harness’ design and a one-step video generator, pointing to gains in precision and latency rather than just bigger models.
The business desk's take → Anthropic confidentially files for U.S. IPO ahead of OpenAI30-second briefing
- New papers advance local 3D, one-step video, harness.
- Enables lower latency and more reliable long-horizon agents.
- Watch adoption of normals metric and harness benchmarks.
One-Line Summary
Today's papers emphasize practical gains: cleaner local 3D geometry, stable one-step video generation for lower latency, and agent research that elevates the surrounding system or “harness” alongside the model.
Research Papers
SurGe improves local surface geometry in point maps
Point-map 3D reconstruction methods often get global shape right but local surface orientation wrong; SurGe makes those local errors measurable with a new “point map normal” metric and reduces them while preserving global accuracy. 1
To do this, it adds two parts: a point gradient matching loss that supervises depth-normalized 3D finite differences, and a Neighborhood Attention Decoder (NAD) that progressively upsamples features and mixes local context using Neighborhood Attention. 1
Across eight zero-shot monocular geometry benchmarks, SurGe achieves the best average rank for global point map Absolute Relative Error (AbsRel) and shows consistent improvements on local point map and point map-normal evaluations. 1
Agent self-evolution separates 'update skill' from 'benefit'
This study examines self-evolving Large Language Model (LLM) agents that update the “harness” around a base model — prompts, skills, memories, and tools — from execution evidence, asking which models write useful updates and which actually benefit from them at task time. 2
It disentangles two capabilities: harness-updating (producing persistent, useful updates) and harness-benefit (gaining from those updates during execution). Findings: harness-updating is largely flat across base capability — even Qwen3.5-9B’s updates yield gains comparable to Claude Opus 4.6 — while harness-benefit is non-monotonic, with mid-tier models gaining most and weak-/strong-tier models gaining less; weak-tier gains are limited by failures to invoke harness artifacts or to follow them faithfully. 2
One-Forcing stabilizes one-step autoregressive video generation
One-Forcing is a training recipe for high-quality “one-step” autoregressive video generation that augments the Self-Forcing distillation objective with an auxiliary Generative Adversarial Network (GAN) loss, aiming to cut latency without sacrificing sharpness or motion. 3
On VBench, One-Forcing scores 83.76 — the state of the art among one-step causal video methods — and remains competitive with strong many-step approaches; it also achieves stable one-step framewise autoregressive generation with about one-third of the training cost of the chunkwise model. 3
From model scaling to system scaling in agentic AI
This position paper argues the next bottleneck in agentic AI is the auditable, modular system around a foundation model — the “agent harness” spanning memory, retrieval, tool use, orchestration, verification, and governance — and that model-centric evaluation undershoots this interaction. 4
It outlines three bottlenecks (context governance, trustworthy memory, dynamic skill routing) and a research agenda for harness-level benchmarks measuring trajectory quality, memory hygiene, context efficiency, communication fidelity, verification cost, and safe evolution; the authors introduce CheetahClaws, a Python-native reference harness, and compare it with Claude Code and OpenClaw. 4
Open Source & Repos
InsForge: all-in-one backend for agentic coding
InsForge is an open-source backend platform that bundles database, authentication, storage, compute, hosting, and an AI gateway so a coding agent can ship end-to-end full-stack apps; it also provides a software development kit (SDK) on npm. 5
The repository highlights Apache-2.0 licensing and activity badges, and shows continued updates, including a v2.1.10 release dated 2026-05-29. 5
Why It Matters
Precision and latency are center stage: SurGe brings evaluation and training closer to the local surface details users notice, while One-Forcing collapses video sampling steps toward near-instant interaction without the usual blur. 1
Equally, the agent studies argue that progress depends on system design — the harness for memory, routing, and governance — as much as on raw model power, a cue for teams to invest in harness-level evaluation and reusable infrastructure alongside models. 4
Terms in This Article
- Attention
Attention is a neural mechanism that computes a weighted aggregation of information by scoring the similarity between a query representation and key–value representations of inputs, enabling long-range, distance-agnostic dependencies; in Transformers it is implemented as scaled dot-product and multi-head self-attention, replacing recurrence and convolution with highly parallelizable processing.
- LLM
A large language model is a deep learning system trained on vast text corpora to understand and generate natural language, typically built on Transformer architectures that capture token-to-token dependencies via self-attention in parallel—often using decoder-only autoregressive next-token prediction—and can extend versatility through encoder–decoder variants or multimodal inputs.
- Claude
Claude is Anthropic’s family of large language models delivered through a developer platform — exposed via the Messages API for direct calls and managed Agents for stateful workflows — that performs text- and code-centric tasks while governing usage through token-based pricing, spend tiers, and token-bucket rate limits.
Two desks on this story
New 3D method sharpens local geometry in point maps
Anthropic confidentially files for U.S. IPO ahead of OpenAI
Anthropic submitted a confidential U.S. IPO filing after a $65 billion raise that valued it at $965 billion. The move tests investor appetite alongside SpaceX’s planned $75 billion offering at a $1.75 trillion valuation and as OpenAI prepares its own filing, while Nvidia’s Cosmos 3 highlights the pivot to embodied AI.
Comments (0)