• Understanding, generation, and editing in one autoregressive model: ARM bets everything on a discrete tokenizer trained against three goals at once — semantic discrimination, language alignment, and faithful reconstruction. RL on the 7B version aligns it to human preference and surfaces cross-task gains between generation and editing.

• Character animation drops the skeleton and the mask, going fully end-to-end: SCAIL-2 splices the whole driving video into the sequence as in-context conditioning, skipping every intermediate representation. Its 37 upvotes led the community that day, with code and weights released.

• RL credit assignment that traces information flow along attention: FlowTracer builds the reasoning chain into a directed acyclic graph and weights only the tokens that actually reach the answer. Accepted at ICML.

• One piece of multimodal evidence compressed into a single latent token: Latent Memory matches mainstream RAG on seven text and multimodal QA benchmarks while cutting generation-side tokens to a third or a tenth.

• A triathlon for video world models: WorldOlympiad runs them across physics, geometry, and interaction. SOTA models show holes in long-horizon interaction and 3D consistency. From Alibaba, with code.

Also Notable

• Linear Attention's State Merging Goes From Fixed to Dynamic — Dynamic Linear Attention adjusts multi-state memory merging by token importance. ICML.
• NVIDIA Replaces Quadratic Attention With Gated Sparse Memory — for long-context modeling, avoiding the loss of compressing history into a fixed size like state-space models.
• Flow Policies Improve at Test Time Through Gradient Guidance — sidestepping the old problem of backpropagating RL gradients through the whole denoising process, for diffusion policies.
• Workflow-GYM Tests Computer-Use Agents on Long-Chain Workflows — checking whether they complete high-value tasks in real professional domains, not fragmented actions.
• Huawei's ActiveMem Makes Long-Horizon Memory Distributed and Active — avoiding the capacity and interference trade-off of cramming centralized memory into one context.
• Two 2-Bit Quantization Papers Landed Today — UniSVQ unifies scalar and vector quantization, LC-QAT (2606.10531) takes data-efficient 2-bit QAT, both for ultra-low-bit deployment, both ICML.
• Multi-Turn Models Can Lock Into Unsafe Stances Early — final-turn refusal rates can't see it; this paper exposes the temporal failures that terminal-score evaluation masks.
• SSR-Merge Does Training-Free LoRA Merging — using subspace signal routing to avoid the parameter interference that wrecks multi-LoRA merges, for diffusion models.
• Low-Light Video Enhancement Goes Modality-Agnostic — AnyMod-LLVE uses auxiliary modalities like event streams or infrared when present and runs without them otherwise, no longer tied to one source.
• AI-Image Detectors' High Scores Come From a Bias Toward "Real" — sensitivity collapses once compression and other post-processing hit; this paper dissects and prunes the bias.

Read the full edition →