Issue #7
Sent 2026-06-22
Looped World Models
On the ScienceWorld world-modeling task (five consecutive actions), a roughly 1B-parameter LoopWM reached 68.4% exact match versus 47.2% for claude-opus-4-6-max. Caveat: this is a single text-based benchmark compared against one general-purpose LLM used off-task, not against the RSSM or transformer world-model baselines the paper positions itself against, and the headline 100x parameter-efficiency and two-orders-of-magnitude FLOP claims are asserted rather than measured.
arXiv:2606.18208 · 476▲ · cs.LG, cs.AI, cs.CL, cs.CV
- RL
- Efficiency
- LLMs
The paper applies looped transformers (a shared transformer block iterated repeatedly across depth) to world models, calling the result LoopWM. It adds a spectrally constrained state-retention matrix to keep latent updates bounded, Poisson-sampled variable loop depth during training, an early-exit gate for adaptive inference, and a deferred-decoding mode that skips intermediate observation reconstruction during multi-step rollouts.
Full breakdown →LoopCoder-v2: Only Loop Once for Efficient Test-Time Computation Scaling
On SWE-bench Verified, the two-loop model scored 64.4 vs 43.0 for the non-looped baseline, while the three-loop model regressed to 27.6 and four-loop to 22.4 (all same 7B model, matched training and tuning). Caveat: these are the authors' own from-scratch models and the headline comparison is against their own R=1 baseline, not an independent equal-compute system.
arXiv:2606.18023 · 209▲ · cs.LG, cs.AI
- LLMs
- Efficiency
- Inference
The paper studies how many times to loop a Parallel Loop Transformer (PLT), a looped-Transformer variant that reuses shared blocks with cross-loop position offsets (CLP) and shared-KV gated sliding-window attention to keep latency and KV-cache memory flat regardless of loop count. It trains LoopCoder-v2, a 7B coder, from scratch on 18T tokens at loop counts of 1, 2, 3, and 4, and uses per-loop diagnostics (hidden-state dynamics, attention routing, output-distribution shift) to explain why performance peaks at two loops.
Full breakdown →MemSlides: A Hierarchical Memory Driven Agent Framework for Personalized Slide Generation with Multi-turn Local Revision
On a diagnostic matched-pair modify evaluation (only 9 pairs, controlled proxy tasks, not real users), tool-memory injection cut geometric-mean core tool time to 0.327x of the no-injection baseline and reduced time to first correct edit from 609.5s to 242.5s; the strongest paired evidence was Strict Verify (0.310 to 0.534) and tool time ratio, both with sign-test p=0.0195, while completion and latency gains were only directional.
arXiv:2606.17162 · 176▲ · cs.CL, cs.HC, cs.MA
- Agents
- LLMs
MemSlides is an agent framework for generating and editing presentation slides that splits memory into three parts: long-term user profile memory (persistent style/layout preferences per user and intent), tool memory (reusable execution traces for editing tools), and session-scoped working memory (temporary constraints across revision turns). It pairs this with a Plan-Act-Guard revision pipeline that scopes edits to the smallest affected slide region rather than regenerating the whole deck on each feedback turn.
Full breakdown →Moebius: 0.2B Lightweight Image Inpainting Framework with 10B-Level Performance
On Places2 (Small), Moebius scores FID 0.92 / LPIPS 0.091 versus FLUX.1-Fill-Dev's 0.94 / 0.099 and PixelHacker's 0.82 / 0.088, at 0.22B params, 0.154 TFLOPs, and 0.52s total inference versus FLUX's 8.05s. Note these are fine-tuned in-domain results on academic benchmarks, not zero-shot; on out-of-distribution sets Moebius (FID 17.81 natural) trails both PixelHacker (13.84) and FLUX (14.52).
arXiv:2606.19195 · 139▲ · cs.CV
- Vision
- Efficiency
Moebius is a 0.22B-parameter image inpainting model built on a latent diffusion U-Net whose transformer blocks are replaced with a Local-Lambda Mix Interaction (LLambdaMI) block that summarizes spatial context and semantic priors into fixed-size linear matrices instead of quadratic attention. It is trained with an adaptive multi-granularity knowledge distillation scheme that aligns the small student to the larger PixelHacker teacher entirely in latent space using gradient-norm-balanced losses.
Full breakdown →Geometric Action Model for Robot Policy Learning
On LIBERO-Plus (out-of-distribution perturbations of the LIBERO benchmark), GAM scored 85.5% overall vs. 84.6% for the next best baseline, with a 9.7 percentage point gain in the camera-perturbation setting, at 1.4B params and 6.9ms inference (about 55x faster than the 382ms diffusion-based Cosmos-Policy). Note the standard LIBERO scores are saturated near 97-98% across most methods, so the differentiation is almost entirely in the perturbed setting.
arXiv:2606.17046 · 117▲ · cs.RO, cs.CV, cs.LG
- Robotics
- Multimodal
- Inference
GAM (Geometric Action Model) is a robot manipulation policy that repurposes a pretrained geometric foundation model (a transformer that infers 3D depth/geometry from images) by splitting it at an intermediate layer, using the shallow layers as an observation encoder and inserting a causal transformer that predicts future latent tokens conditioned on language, proprioception, and action history. The remaining deep layers of the same backbone then decode those predicted tokens into both future depth maps and robot action chunks in a single forward pass.
Full breakdown →DreamX-World 1.0: A General-Purpose Interactive World Model
On the authors' own 5-second basic evaluation, DreamX-World-1.0-5B scored 73.75 camera control and 84.76 overall, versus 80.79 (HY-WorldPlay 1.5, 8B) and 80.45 (LingBot-World, 14B); note this is a self-devised benchmark and trajectory set, not an external standard, and the gap narrows on 30-second rollouts (70.41 vs 68.85 and 67.43).
arXiv:2606.16993 · 113▲ · cs.CV
- Vision
- Efficiency
- Inference
DreamX-World 1.0 is an interactive text/image-to-video world model built by fine-tuning Wan2.2 to support camera navigation, revisiting earlier scenes, and prompted multi-object events across photorealistic, game, and stylized domains. It combines a camera-conditioning method called E-PRoPE (projective positional encoding applied to spatially downsampled tokens), geometry-based memory retrieval, DMD distillation into a few-step autoregressive generator, and RL alignment, with serving optimizations to hit real-time streaming.
Full breakdown →Learning from the Self-future: On-policy Self-distillation for dLLMs
On GSM8K with LLaDA-8B-Instruct, d-OPSD reaches 81.0 (256 seq len) versus 79.8 for the diffu-GRPO RLVR baseline and 76.0 for the base model, while converging in about 425 optimization steps versus 7700 for diffu-GRPO. Caveat: gains are modest and evaluations cover only four reasoning tasks on a single 8B model, with best-checkpoint selection reported.
arXiv:2606.18195 · 76▲ · cs.CL
- LLMs
- Training
- RL
The paper introduces d-OPSD, an on-policy self-distillation method for diffusion language models (dLLMs) where a single model acts as both student and teacher. The teacher gets privileged information by revealing part of the model's own already-generated answer as a suffix condition, and training minimizes step-level KL divergence between student and teacher across the iterative denoising steps instead of the token-level divergence used for autoregressive models.
Full breakdown →PerceptionDLM: Parallel Region Perception with Multimodal Diffusion Language Models
On the authors' own ParaDLC-Bench, PerceptionDLM hits 62.4% average accuracy versus roughly 35% for diffusion baselines (LLaDA-V) and about 69% for AR region models (GAR, DAM), while cutting total benchmark inference time to 276s vs 479s for GAR. Caveat: accuracy is still below the AR models, the benchmark is new and self-constructed, and the 3.44x throughput speedup is measured under a fixed 4-masks-per-image workload.
arXiv:2606.19534 · 64▲ · cs.CV, cs.AI, cs.CL
- Multimodal
- Vision
- Efficiency
PerceptionDLM is a multimodal diffusion language model that captions multiple image regions at once inside a single denoising pass, instead of the region-by-region autoregressive decoding used by prior models. It combines a SigLIP-2 vision encoder and an 8B LLaDA diffusion backbone with region prompting (learnable per-region embeddings), RoI-aligned feature replay, and a structured attention mask that isolates each region's tokens while sharing global context.
Full breakdown →Subscribe
Get the next issue.
Free. One email a week. Unsubscribe any time: no account, no dark patterns.