In another reckless expansionist provocation, the Philippine government has delimited the so-called "baselines of the territorial sea" of China's Huangyan Dao, a move that flagrantly violates China's sovereignty and international law and is doomed to fail.

One of the interesting properties we’ve observed around schema-guided, complex document extraction tasks 📑 is that coding agent harnesses are good baselines (in terms of cost/accuracy) for longer documents. We tested Claude Code and Codex, along with specialized OCR tools (including LlamaParse) and raw VLMs. * On short documents, specialized OCR tools are generally a fraction of the cost of coding agents, with equivalent or higher accuracy * On longer documents, coding agents are a bit closer to the cost/accuracy Pareto curve (see bottom graph) It’s an interesting result, though ultimately not surprising. Complex document extraction is a specialized reasoning task, and coding agents are effectively generalized reasoning harnesses. Over long documents, coding agents have more room to use a variety of tools to search snippets of the document instead of loading the entire document into context. They can also make use of prompt caching to reduce total token cost even as it expands multi-step reasoning. On the flip side, they do generate a baseline degree of token usage that proves to be wasteful for shorter docs compared to specialized extractors. This specific graph is in our Appendix D in the ParseBench paper, come check it out! ArXiv: ExtractBench:

Cool paper showing how effective tuning a system prompt for an agent can be. Recommended paper if you tune agent harnesses. This paper presents EvolveTrade, which treats a trading agent's system prompt as its policy. After each trading interval, a separate Policy Agent reads the decision traces and the realized returns and rewrites the prompt. The backbone model stays frozen. Across several market regimes and two backbone models, the evolved agent beats fixed-prompt baselines on Sharpe ratio and cumulative return in most settings. The rewritten prompts also led the agent to run more code-based analysis and to compute signals that fit the current market regime. Paper:

Nice paper showing a better way to evolve agent skills. And they achieve 40–70% less token cost compared to frontier evolving methods. The idea is to let agents improve their skill prompts by ranking candidates with a learned rubric instead of running a full rollout to score every revision. Rollout cost is the reason skill self-evolution usually only patches observed failures. Every candidate edit needs a real agent run to evaluate. SkillLift trains a rubric to agree with real outcomes on which of two skills is better, since ranking needs fewer oracle runs than predicting each score. An inner loop revises skills against the frozen rubric at no rollout cost. An outer loop spends a few real rollouts to re-align the rubric by rank correlation. On SkillsBench and WildClawBench (147 tasks) with three models, it beats SkillOpt and CoEvoSkills in all six combinations, even when those baselines get twice the token budget. It reaches target performance with 40 to 70% fewer tokens. Paper:

Reinforcement Learning for Active Perception in Autonomous Navigation. [📍GitHub & Paper ] Most robots navigate as if their cameras were nailed in place. But perception is not passive. Animals move their heads and eyes constantly to decide where to go next. Robots should do the same. That is the idea behind “Reinforcement Learning for Active Perception in Autonomous Navigation,” which has just been accepted at ICRA 2026. Instead of treating navigation and perception as two separate problems, this work trains flying robots to do both at once. The robot does not only decide where to move. It also decides where to look. Using reinforcement learning, the robot learns to: •fly safely through cluttered environments, •actively reorient its onboard camera to reduce uncertainty, •balance reaching a goal with gathering better visual information. The key result is that actively controlling perception makes navigation safer. In simulation and on a real flying robot, this approach consistently outperforms static-camera baselines. The sim-to-real transfer holds up, which is usually where things break. This is a small but important shift in how we think about autonomy. Better planning does not always come from better maps or bigger models. Sometimes it comes from simply looking in the right direction at the right time. Thanks for sharing, Kostas Alexis! 📍Code: Paper: Video: —— Weekly robotics and AI insights. Subscribe free:

Recent research has highlighted the promise of scaling memory embeddings in LLM training. While Engram and STEM index memory by token identity or local n-grams, can we design more flexible memory routing that captures how each token’s meaning changes with context? This paper introduces Mixture-of-Memory Embeddings (MoME), which uses a learned router to sparsely select among multiple memory slots for each token. The architecture outperforms strong memory baselines across three model families. Interestingly, both qualitative and quantitative analyses show that the learned routers’ activations correlate with the context-dependent senses of polysemous words. The authors also train a sub-billion-parameter model that achieves competitive CORE-22 performance against similarly sized base models, including Qwen3-0.6B and Llama 3.2-1B. Code and pretrained models are publicly available!

🌍 TL;DR: A training-free way to re-explore a recorded video from new camera angles just dropped. It handles viewpoint changes, hole-filling, and recovering earlier appearances entirely through a frozen model's own attention. Title: World in World: Explore the World with World Models URL: Points 🎥 Four evidence sources (source video, target-view scene, rendered geometry, retrieved history) get converted into clean K/V fed straight into the frozen model 🧭 CGAR routes queries toward reliable source-video tokens using persistent point correspondence and camera geometry 🎚 EWA independently tunes each auxiliary channel's contribution using existing attention outputs, with zero extra network evaluations 🏆 Top VBench score of 85.192 and lowest rotation error (2.8326°) against 6 baselines including ReCamMaster 🔧 Same framework also powers bullet-time rendering, video stabilization, editing, and motion transfer Getting this much versatility with zero retraining feels like a win for evidence engineering over model adaptation. #WorldModels# #VideoGeneration#

“BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure” Reward hacking in agent benchmarks is often an infrastructure problem, not just a model-behavior problem. So this paper formalizes the full reward path and instruments runs to distinguish vulnerable tasks from actual exploit use, reaching 96% runtime detection accuracy and much higher exploit-chain recall than prior scanning baselines.

Most "streaming" TTS systems still wait for a complete sentence before speaking. X2Streaming-TTS removes that wait. It consumes text tokens as they arrive and generates speech with strict zero lookahead. The challenge: once speech is played, it cannot be revised. Take “He finished 3…”—should “3” become “three” in “3 laps,” or “third” in “3rd place”? Speaking too early risks an error; waiting defeats the purpose of streaming. X2Streaming-TTS addresses this with: 1️⃣ Causal commitment Ambiguous numbers, units, and symbols are held until their pronunciation becomes clear. Segments are closed using both punctuation and acoustic capacity. 2️⃣ Speech-state inheritance Waveform-decoder state and acoustic history are carried across segments, preserving pitch, timbre, and continuity. Results: ⚡ 15.8 ms median TTFT for one request ⚡ Under 120 ms at 64 concurrent requests 🎯 Lowest recognition error in 6/8 evaluated streaming conditions 🔢 0% CER on numeric and streaming-ambiguity tests 🗣️ Quality comparable to evaluated offline baselines Low-latency voice AI is not just about speaking faster—it is about knowing when it is safe to speak. Paper: Code: #TTS# #VoiceAI#