All Updates
Page 1129 of 1528
March 19, 2026
RideJudge: AI Framework for Ride Disputes
RideJudge introduces a progressive framework aligning visual and logical reasoning for ride-hailing dispute adjudication, addressing limitations in multimodal LLMs. Key innovations include SynTraj for trajectory synthesis, Adaptive Context Optimization, and Ordinal-Sensitive RL. The 8B model achieves 88.41% accuracy, outperforming 32B baselines.
PIER RL Cuts Shipping Fuel Waste 9-Fold
PIER is a physics-informed offline RL framework that learns fuel-efficient, safety-aware maritime routing from historical AIS data and ocean reanalysis, without needing online simulators or forecasts. It reduces mean CO2 emissions by 10% versus great-circle routing and slashes catastrophic fuel waste from 4.8% to 0.5% of voyages on 2023 Gulf of Mexico routes. The method shows 3.5x lower fuel variance and transfers to other domains like wildfire evacuation.
LLMs Struggle in Clue Reasoning Test
Researchers built a text-based multi-agent Clue game to evaluate LLMs' multi-step deductive reasoning with GPT-4o-mini and Gemini-2.5-Flash agents. Across 18 simulated games, agents secured only four correct wins, struggling with consistent reasoning. Fine-tuning on logic puzzles did not reliably enhance performance and sometimes increased verbosity without precision gains.
Kumiho: Graph-Native Memory for AI Agents
Kumiho introduces a graph-native cognitive memory architecture for AI agents, grounded in formal AGM belief revision semantics. It unifies cognitive memory and versioned assets using primitives like immutable revisions and URI addressing. Achieves SOTA results on LoCoMo (0.565 F1) and LoCoMo-Plus (93.3% accuracy), outperforming Gemini.
InfoDensity Rewards Dense Reasoning Traces
Large Language Models generate verbose reasoning traces that waste compute. InfoDensity proposes an RL reward framework using AUC-based and monotonicity signals, weighted by length, to favor concise, high-quality steps. It matches SOTA accuracy on math benchmarks while cutting token usage.
Geometry-Switching Fixes Agent Cascade Failures
New method introduces cascade-aware routing for multi-agent AI systems using spatio-temporal sidecars and adaptive geometry switching between Euclidean and hyperbolic models. It boosts Genesis 3 benchmark win rates from 50.4% to 87.2%, especially in tree-like graphs prone to exponential failures. The lightweight 133-parameter MLP selector uses topology stats for geometry decisions.
GenAI Workflow for Participatory Planning Models
Researchers propose a templated LLM workflow to accelerate problem conceptualization in socio-environmental planning under deep uncertainty. It uses models like ChatGPT to extract components from stakeholder descriptions, unify perspectives, and generate Python implementations iteratively. Demonstrated successfully on lake and electricity market problems.
Draft-and-Prune Boosts Auto-Formalization Reliability
Draft-and-Prune (D&P) is an inference-time framework that improves auto-formalization by drafting diverse natural-language plans, pruning incorrect formalizations, and aggregating via majority voting. It achieves 78.43% accuracy on AR-LSAT with GPT-4 and 78% with GPT-4o, outperforming baselines like MAD-LOGIC and CLOVER. D&P reaches near-ceiling performance, including 100% on PrOntoQA and LogicalDeduction.
CRAFT: Hidden-State RL for Jailbreak Defense
CRAFT is a red-teaming framework that aligns reasoning models using hidden representations to boost robustness against jailbreaks. It combines contrastive learning with RL to separate safe/unsafe trajectories in latent space. Achieves 79% reasoning safety and 87.7% response safety gains on Qwen3-4B-Thinking and R1-Distill-Llama-8B, outperforming IPO and SafeKey.
AI Scientist via Synthetic Task Scaling
Researchers developed a synthetic environment pipeline for training machine learning agents using the SWE-agent framework. The pipeline generates verified ML challenges grounded in real Huggingface datasets with self-debugging. Student models trained on GPT-5 trajectories improved MLGym benchmark performance by 9% for Qwen3-4B and 12% for Qwen3-8B.
Anthropic's 80K Claude User Survey
Anthropic surveyed ~80,000 Claude users across languages, revealing strong demand for expert-level AI capabilities with 80%+ reporting practical value. Concerns include reliability issues and loss of autonomy, especially cognitive decline fears in East Asia like Japan.
Metcash Adds AI Search to B2B Platform
Metcash has introduced AI-powered search to its B2B marketplace. The SAP-based platform is now augmented with Coveo's AI technology. This upgrade aims to enhance search capabilities for wholesale users.
Aliyun Compute Shortage Sparks Token Surge
Alibaba Cloud faces acute computing power shortages amid the AI agent boom. Token prices are surging dramatically. The article analyzes Aliyun's valuation in this new era.
Frore Unicorn: $112B AI Chip Cooling Funding
Frore Systems raised $143M Series D at $1.64B valuation for liquid cooling tech serving Nvidia, Qualcomm, AMD chips. Pivoted to liquid cooling post-Nvidia CEO advice amid AI data center heat surge. Neutral positioning benefits from all AI chipmakers' growth.
CarbonBase and Wukong Burn Tokens for New AI
Carbon-based tech and Alibaba's Wukong collaborate on intensive token-burning compute tasks. This partnership yields a groundbreaking new AI species. Teases larger developments post-super individuals.
Xiaomi AI Models Spark Shares Surge, SU7 Facelift Nears
Xiaomi Corp. shares rallied Thursday after releasing its latest artificial intelligence models. The company is also preparing to launch a refreshed version of its SU7 electric vehicle.
Qwen-Code v0.13.0 Preview with Agent Arena
Qwen-Code releases v0.13.0-preview.0 featuring system prompt customization in SDK/CLI, hooks extension mechanism, and agent collaboration arena for multi-model competition. Performance boosts via concurrent task tool execution and truncation support for MCP outputs. VSCode enhancements include fuzzy search, token usage display, and Tab-fill completions.
China's First Robot-Majority City in Yizhuang?
Yizhuang is set to become China's first location where robots outnumber residents. The article recommends that every tech entrepreneur visit this emerging robotics hotspot.
Why AI Rationality Fails Leaders
Mearsheimer's prescient warnings on wars ignored due to arrogance, fear, and power's emotional logic. Rational analysts like him or AI provide clear insights but lose to emotion-driven politics. Humans' ancient emotional systems override rationality in high-stakes decisions.
Zhongqing & Duolun Launch Qing'an Robotics
ๆฑ่ๆๅฎๆฉๅจไบบๆ้ๅ ฌๅธ่ฟๆฅๆ็ซ๏ผ่จปๅ่ณๆฌ1ๅๅ ๏ผ็ถ็็ฏๅๆถต่ๆบ่ฝๆฉๅจไบบ็ ็ผ้ทๅฎใAIๆ็จ่ป้ซ้็ผ็ญใ่กๆฑ็บๅคๅซ็งๆ่ๆทฑๅณ็พๆๆฉๅจไบบ็งๆใ