• Retrieved memory pushes agents into flattery. MemSyco-Bench argues memory isn't just a store-and-fetch accuracy problem — what a user said before can outweigh objective evidence, and current memory evals miss this blind spot entirely.

• Visual reasoning fails at seeing, not thinking. P2R splits perception out from reasoning — locate the evidence precisely first, then answer — and a 4B model hits 93.2% on V-Star. PixelEyes bet on the same path the same day.

• Every data-recipe change means retraining a proxy model. CausalMix wants one fit to cover them all. It treats data mixing as causal inference, so a changing data pool shouldn't force a full restart.

• Diffusion world models imagine many futures by nature, but they're too slow for online planning. Valdi uses single-step diffusion to cut latency, and exposes a tug-of-war between multimodal prediction and control performance.

Also Notable

• ByteDance Seed2.0's Real Story Is the Eval, Not the Model — it builds an evaluation close to real complex scenarios (long-tail knowledge plus complex instruction following) and reverse-engineers model goals from it; the methodology beats the model card, but discount the self-reported numbers.
• Another Route for Continuous Latent Reasoning: Skip the Language-Token Bottleneck — but training on the posterior of ground-truth answers causes a train-inference mismatch, which this patches with asymmetric mutual variational learning.
• Turning Video Retrieval From One-Shot Preprocessing Into Iterative Refinement — a failed first retrieval is no longer a dead end; soft query refinement drives inter-video and intra-video reasoning.
• A Unified World Action Model for Mobile Manipulation — it points out that current WAMs model over coarse video chunks, entangle navigation and manipulation actions, and train in a way that mismatches autoregressive inference.
• Traceable Hypothesis Generation for Materials Discovery — graph-native GRPO fine-tuning makes intermediate reasoning steps checkable, addressing LLMs that sound fluent while hiding whether the reasoning holds.
• Stanford's Multi-Turn Agentic Literature Search — for cases where user intent is itself vague and evolves through interaction, it replaces the fixed pipeline with workflow induction.

Read the full edition →