• Latent reasoning gains come from side effects, not reasoning itself. Causal mediation analysis reveals a causal disconnect between latent tokens and both inputs and outputs. A simple text-based "imagination" baseline outperforms complex latent-space methods.

• Deep research agent cuts 70% of reasoning steps and gets more accurate. Parallel evidence gathering replaces serial reasoning chains. Search breadth beats reasoning depth.

• "Test-driven error correction" from education psychology enters multimodal training. A diagnostic-reinforcement loop automatically identifies model weaknesses and generates targeted data. Eleven benchmarks improve without interfering with each other.

• A framework for what makes a world model "qualified" arrives, but practical application is far off. Triple consistency (modality, spatial, temporal) provides a unified evaluation lens. The 184 HF upvotes reflect community anxiety more than a solved problem.

• Multi-agent error propagation gets a plug-and-play firewall. Test-time pruning intercepts misinformation flow without retraining or changing topology. Average improvement: 6.3 percentage points.

Also Notable

• Route Planning Agents Get a Real-World Benchmark — From Baidu, combining real map services with diverse travel scenarios. 94 HF upvotes confirm strong demand.
• Memory-Augmented Exploration Helps LLM Agents in Unfamiliar Environments — Hybrid on/off-policy RL framework. ICLR accepted.
• Omni-Modal Agent Evaluation: Vision + Audio + Language — OmniGAIA benchmark shows current models still fall short on cross-modal reasoning.
• Medical RL Framework Outputs Free-Text Diagnoses, Not Multiple Choice — Composite reward design moves closer to clinical utility.
• Token-Level Sparse Attention Breaks the Block-Granularity Ceiling — Long-context inference latency can drop further.
• Modeling Diffusion Denoising as Path Planning — DPCache, a training-free caching acceleration method. CVPR accepted.
• Membership Inference Without Captions — Uses the model's own embedding distribution to detect training data memorization. ICLR accepted.
• Second-Order Structure in RL Rollouts Improves Data Efficiency — Creates dependencies between responses instead of just generating more independent ones.
• Hierarchical GRPO: Step-Level + Group-Level Optimization — Targets long-sequence agent tasks. ICLR accepted.
• Two iPhones Capture Scene-Level 4D Human Motion Data — Dramatically reduces data collection cost for embodied agent training.
• VLM Reasoning Gaps May Trace Back to Reporting Bias in Training Data — Humans instinctively omit the obvious when describing images. Models learn the same omission.
• Single Forward Pass Edits Internal Representations to Reduce Hallucination — No reference model or multi-round inference needed. CVPR accepted.

Read the full edition →