All Updates

#exploration#bisimulation#visual-rl

TEB Boosts Visual RL Exploration

TEB introduces task-aware exploration for visual reinforcement learning under sparse rewards using a predictive bisimulation metric. It learns behaviorally grounded representations and measures intrinsic novelty in latent space. Experiments on MetaWorld and Maze2D show it outperforms baselines.

#energy-based-models#training-free

Spilled Energy Detects LLM Hallucinations

Researchers reinterpret LLM softmax classifiers as Energy-Based Models (EBMs) to track 'energy spills' during decoding, correlating with factual errors, biases, and hallucinations. They introduce two training-free metrics from output logits: spilled energy (discrepancy across steps) and marginalized energy (single-step). Evaluated on nine benchmarks with LLaMA, Mistral, Gemma, and Qwen3, it achieves robust detection without probes or training.

💼

VentureBeat•43d ago

Smarsh Archie AI Boosts Self-Service 59%

Smarsh launched Archie, an AI support agent built on Salesforce Agentforce, acting as a 'front door' for regulated industries to simplify support across products. It centralizes knowledge for plain-language queries, driving 59% self-service adoption. Results include 20% higher success rates, 25% faster resolutions, and 30% productivity boost.

#customer-service#agentic-ai#compliance

#fiber-bundle#landauers-principle#symbolic-ai

Physics Forces Symbolic AI Semantics

Challenges static semantics in visual AI as incomplete, proposing dynamic Observation Semantics Fiber Bundle for bounded agents. Proves thermodynamic limits via Landauer's Principle create Semantic Constant B, requiring phase transition to discrete symbolic structures. Language emerges as ontological necessity to prevent information collapse.

#chart-summarization#multi-agent#multimodal

Multimodal Agent Unlocks Deeper Chart Insights

Chart Insight Agent Flow is a new multi-agent framework that uses MLLMs to extract profound insights from chart images, surpassing basic descriptions. It introduces ChartSummInsights, a dataset of real-world charts with expert-written insightful summaries. Experiments show it significantly boosts MLLM performance on chart summarization.

#federated-learning#reasoning-patterns

LaDa: Federated Reasoning Distillation Framework

LaDa is a federated reasoning distillation framework addressing bidirectional learnability gaps between LLMs and SLMs via adaptive data allocation. It uses a model learnability-aware filter to select high-reward samples for effective knowledge transfer. Domain-adaptive distillation aligns reasoning paths through contrastive learning, acting as a plug-in for existing frameworks.

#rl-alignment#hierarchical-rewards#reward-design

HRDL: Language Rewards Align RL Agents

HRDL extends reward design to encode nuanced human preferences for hierarchical RL agents in complex tasks. L2HR translates language specs into hierarchical rewards. Experiments demonstrate superior task completion and adherence to human specifications.

#agentic-reasoning#ranking-systems#ml-framework

GEARS: Agentic Framework for Ranking Optimization

GEARS reframes large-scale ranking as an agentic discovery process in a programmable environment, addressing engineering bottlenecks over modeling limits. It encapsulates expert knowledge into reusable agent skills for high-level intent steering and includes validation hooks for robust, stable policies. Experiments across product surfaces show superior near-Pareto-efficient outcomes.

#agentic-ai#analytic-diversity#reproducibility

AI Analysts Reveal Data Analysis Diversity

Autonomous AI analysts powered by LLMs replicate human 'many-analyst' studies scalably, producing diverse outcomes on the same dataset. They independently build analysis pipelines, showing dispersion in effect sizes, p-values, and hypothesis support. Variability is structured and steerable via LLM models and personas, validated by an AI auditor.

#automated-research#ai-scientist#token-consumption

FARS Outputs 100 Papers in 228 Hours

Analemma's Fully Automated Research System (FARS) ran for 228 hours straight, generating 244 hypotheses and 100 short papers at a rate of one every 2 hours. It consumed 11.4 billion tokens during this public demo. The month-long live stream continues online.

#embodied-ai#funding#open-source-model

Qianxun Raises $2.8B, Valuation Tops $14B

Embodied AI firm Qianxun Intelligent secured nearly 20B RMB in two funding rounds from top VCs, industry giants, and state funds. Valuation exceeds 100B RMB, entering the 'billion club'. Spirit v1.5 model outperforms Pi0.5 with data pyramid training.

#multi-agent#scaling-laws#agent-benchmarks

DeepMind Debunks More Agents Always Better

DeepMind's paper evaluates 180 agent configs across 5 architectures, revealing scaling limits where more agents hurt performance if mismatched to tasks. Tested on Finance-Agent, web nav, planning benchmarks with GPT/Gemini/Claude. Introduces quantitative principles for agent systems.

⚛️

量子位•43d ago

DeepSeek GitHub Updates Trigger Wall Street Fear

DeepSeek is continuously updating its GitHub repositories, reigniting fears among Wall Street and the US AI community of being dominated. This activity signals the potential arrival of a 'DeepSeek second moment' as the company resumes work post-holiday.

#github-updates#ai-competition#china-ai