• δ-mem Bolts an 8×8 State Matrix Onto a Frozen Backbone. A delta-rule online update lifts memory-heavy tasks 10–15% over baseline. Reframes long context from "stretch the window" to "design a state machine."

• CausalCine Treats Shot Boundaries as First-Class Citizens. Native multi-shot training plus relevance-routed KV retrieval (CAMR) reaches bidirectional-model quality while keeping streaming interactivity.

• Multi-Turn Activation Steering Drift Has a Concrete Mechanism. Steered tokens land in KV cache and get re-consumed every step, compounding the drift. Move the intervention to the attention-delta layer and round-10 trait retention climbs from 78% to 93%.

• SenseNova-U1 Frames "Understand vs. Generate" as a Structural Problem. 157 upvotes show the community is waiting for a credible unified VLM, but the abstract alone can't tell you how it actually differs from Chameleon and Janus.

Also Notable

• Mela Borrows Neuroscience Consolidation for Hierarchical Memory Modules. Cross-frequency coupling stabilizes test-time experience into durable representations. Pairs with δ-mem as a second branch of state-machine memory. Mela
• Agent Memory Benchmarks Move From "Remember User History" to "Remember Environment, Interfaces, Failure Modes." LongMemEval-V2 evaluates long-term memory inside real web-agent usage. The target is now "an experienced colleague," not a chat log. LongMemEval-V2
• Research Agent Skill Library, Memory, and Policy Evolve Together. NanoResearch handles user differences in resources, methodology, and output format as a precondition rather than post-processing. NanoResearch
• Unreal Engine 5 Powers a Controllable Visual Simulation Framework. LychSim aims to lower the graphics-stack barrier for closed-loop optimization and OOD evaluation. LychSim
• Agent Knowledge Bases Get Polluted by Incompleteness, Errors, and Redundancy. DeepRefine uses RL to make refinement a learnable step instead of a manual cleanup. DeepRefine
• Train Agents to Learn Patterns From Cheap Small Experiments, Then Configure Expensive Large Ones. Large LLM experiment configs are too costly for trial-and-error AutoML. AutoLLMResearch routes around that. AutoLLMResearch
• A Single Neuron Can Bypass Safety Alignment. Mechanistic split between refusal neurons and concept neurons gives red teaming and alignment audits a concrete handle.
• Jina v5-omni Uses Locked Aligned Towers to Preserve Each Modality's Geometry. No brute-force collapse into a shared space. Useful reference for multimodal retrieval and embedding services. jina-embeddings-v5-omni
• Structural Causes Behind Performance Drops Under Complex Interleaved Instructions. Scaling laws for interleaved multi-image editing and generation, useful for unified visual-generation interface design. Images in Sentences
• Neural Decoders for Quantum Error Correction Get Re-Examined Under Throughput and Latency. Systematic discussion of what role neural networks should play in the QEC pipeline.

Read the full edition →