Fig.1

Issue archive

Issue #4

Sent 2026-06-01

Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players

On synchronized two-agent Minecraft trajectories, Gamma-World reports FVD 184.1 / FID 24.8 on the Memory protocol versus Solaris at 333.8 / 51.7 and a frame-concat baseline at 450.6 / 69.8, with similar gains across Grounding, Movement, Building, and Consistency. Sparse Hub Attention cuts self-attention latency at 8 players to 4.5ms vs 17.6ms for dense attention. Caveat: evaluation is limited to Minecraft-style scenes plus a small real-world robot-arm demo, all against just two baselines.

arXiv:2605.28816 · 431▲ · cs.CV

  • Vision
  • Efficiency
  • Robotics

Gamma-World is a video world model that generates synchronized, action-controllable video streams for multiple agents (game players or robot arms) sharing one environment. It introduces two components: Simplex Rotary Agent Encoding, a parameter-free extension of 3D RoPE that assigns each agent a rotary phase at the vertex of a regular simplex so agents stay distinct but permutation-symmetric, and Sparse Hub Attention, where learnable hub tokens mediate cross-agent communication instead of dense all-to-all attention.

Full breakdown →
Fig. 1

Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments

On real-world ALOHA out-of-distribution tasks, the pretrained variant hit 76.9% average success versus 41.5% for the pi-0.5 baseline and 36.2% for the same architecture trained from scratch. Note most manipulation gains come from the SFT stage; the final RL step added only about +2.9 points on the SimplerEnv environment it was trained in and near-zero on held-out benchmarks, and cross-benchmark comparisons pit one generalist against separately-tuned specialists.

arXiv:2605.30280 · 146▲ · cs.RO, cs.AI, cs.CL

  • Robotics
  • Multimodal
  • RL

Qwen-VLA is a single vision-language-action model that puts a DiT-based flow-matching action decoder on top of the Qwen3.5-4B multimodal backbone, so one model handles robot manipulation, navigation, and trajectory prediction. It uses embodiment-aware text prompts (describing the robot platform, arm config, control frequency, and prediction horizon) instead of separate output heads per robot, and is trained in four stages: text-to-action decoder pretraining without images, multimodal continued pretraining, supervised fine-tuning, and PPO reinforcement learning.

Full breakdown →
Fig. 2

LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding

In the default Hybrid mode, LocateAnything-3B hits 12.7 boxes per second versus 5.0 for the comparable Rex-Omni-3B (2.5x) and about 1.1 for textual-coordinate Qwen3-VL (over 10x), while also improving mean F1 by +3.8 on LVIS and +1.8 on COCO. Caveat: the headline throughput numbers were all measured on COCO with batch size 1 on a single H100, and Fast mode alone drops accuracy in dense scenes (for example Dense200 F1 falls from 61.3 Hybrid to 46.8 Fast).

arXiv:2605.27365 · 145▲ · cs.CV, cs.AI, cs.LG, cs.RO

  • Multimodal
  • Vision
  • Inference

LocateAnything is a vision-language model for object detection and grounding that uses Parallel Box Decoding (PBD), which predicts all four coordinates of a bounding box in a single forward step instead of generating them token by token. It treats each box (or point) as an atomic block with bidirectional intra-block attention, trained jointly with a standard next-token prediction objective and offering three inference modes (Fast, Slow, Hybrid).

Full breakdown →
Fig. 3

AgentDoG 1.5: A Lightweight and Scalable Alignment Framework for AI Agent Safety and Security

The authors report their small models reach accuracy comparable to GPT-5.4 and Gemini-3.1-Pro on trajectory safety classification while training on roughly 1k samples, and claim 1/100 the memory and startup latency of Docker-level environments like SWE-Bench. Caveat: several headline benchmarks (ATBench, ATBench-Claw, ATBench-Codex) are built by the same group using the same taxonomy and synthetic data pipeline, so cross-vendor comparability is limited, and the referenced models and dates are forward-looking/fictional.

arXiv:2605.29801 · 144▲ · cs.AI, cs.CL, cs.CR, cs.CV, cs.LG

  • Agents
  • Safety
  • LLMs

A framework for training small guardrail models (0.8B to 8B parameters) that audit an AI agent's full execution trajectory and label it safe or unsafe, plus diagnose the risk source, failure mode, and real-world harm. It combines a taxonomy-guided synthetic data engine, influence-function-based data selection to shrink training to about 1k samples, and two-stage SFT plus RL (using a per-dimension variant of GRPO called GDPO).

Full breakdown →
Fig. 4

DVAO: Dynamic Variance-adaptive Advantage Optimization for Multi-reward Reinforcement Learning

On tool-use (BFCL-v4, Qwen2.5-7B-Instruct) DVAO reached 63.00% average accuracy and 79.21% format compliance, versus 60.13%/68.12% for GDPO and 44.25%/68.04% for advantage combination. Evaluations are limited to dual-objective settings (accuracy plus length or format) with equal-weight initialization across all baselines; multi-objective (3+ reward) behavior is untested.

arXiv:2605.25604 · 138▲ · cs.CL, cs.LG

  • RL
  • LLMs
  • Training

DVAO is a method for combining multiple reward signals in GRPO-style RL training of LLMs. Instead of using fixed weights to combine rewards or advantages, it scales each objective's contribution by that objective's empirical reward standard deviation within a rollout group, up-weighting objectives that show more variation (a stronger learning signal) and down-weighting noisy or saturated ones.

Full breakdown →
Fig. 5

COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation

The only quantitative evidence is deployment surface, not performance: as of 2026-05-28 the repo had roughly 18.5k GitHub stars, and the gallery listed 215 skills from 165 contributors with over 100k cumulative stars. The authors themselves caveat that these counts indicate distribution reach only, not behavioral fidelity, task impact, or adoption quality, and no accuracy or utility benchmark is reported.

arXiv:2605.31264 · 123▲ · cs.AI, cs.CL, cs.LG

  • Agents
  • LLMs
  • Safety

COLLEAGUE.SKILL is a pipeline that takes heterogeneous traces from a person or role (chat logs, docs, review comments, interviews) and distills them into a versioned skill package compatible with the Agent Skills standard (a folder centered on SKILL.md). Each package splits into a capability track (procedures, review criteria, decision heuristics) and a bounded behavior track (communication style, interaction rules, correction log), and supports inspection, natural-language correction, rollback, and installation across agent hosts.

Full breakdown →
Fig. 6

GrepSeek: Training Search Agents for Direct Corpus Interaction

Best overall token-level F1 (0.5691 micro-average) across seven open-domain QA benchmarks versus 0.5441 for the strongest baseline (Search-R1 with Qwen3-4B dense retriever), winning 4 of 7 datasets. Caveat: it significantly loses on PopQA (long-tail entities, diacritics) due to exact-match brittleness, and per-query latency is higher (8.67s vs 4.77s for E5) because of longer reasoning trajectories, though the sharded-parallel engine cuts tool execution from 5.39s to 0.71s (up to 7.6x).

arXiv:2605.29307 · 116▲ · cs.CL, cs.AI, cs.IR, cs.LG

  • LLMs
  • Agents
  • RL

GrepSeek trains a compact LLM (Qwen3.5-9B) to answer questions by issuing Unix shell commands (rg, grep, head, awk, etc.) directly against a raw text corpus instead of querying a precomputed retrieval index. It uses a two-stage pipeline: supervised fine-tuning on a synthetic cold-start dataset built by an answer-aware Tutor and answer-blind Planner, followed by GRPO reinforcement learning, plus a sharded-parallel execution engine that runs shell pipelines across corpus shards while preserving byte-exact output.

Full breakdown →
Fig. 7

WBench: A Comprehensive Multi-turn Benchmark for Interactive Video World Model Evaluation

Across 20 models on the 158-case navigation split, dedicated camera-controlled (76.0) and action-conditioned (77.7) models beat text-driven models (67.6) on navigation by about 10 points, yet text-driven models led on physics compliance (67.0 vs 64.2 and 61.7). Navigation degraded roughly 33 points from turn 1 to turn 4+, the fastest decay of any interaction type. Automated scores matched human blind pairwise rankings with Spearman rho >= 0.94 across all ten aspects, though this validates ranking order, not absolute accuracy.

arXiv:2605.25874 · 104▲ · cs.CV

  • Vision
  • Multimodal

WBench is a benchmark for evaluating interactive video world models (systems that generate the next video frame conditioned on observation history plus a user action). It defines 289 test cases with 1,058 multi-turn interaction turns across five dimensions (video quality, setting adherence, interaction adherence, consistency, physics compliance), scored by 22 automatic sub-metrics that combine specialist vision models (MegaSaM pose estimation, SAM2, Depth Anything 3, DINOv2) with VLM-based grading.

Full breakdown →
Fig. 8

Subscribe

Get the next issue.

Free. One email a week. Unsubscribe any time: no account, no dark patterns.