AI Research news and analysis on Web Pulse: 22652 curated articles tracking the latest AI Research developments, tools, and research, updated continuously from vetted sources.
A developer analysis of Artificial Analysis benchmarking data finds that GLM-5.3-Flash scores 42 at $0.25 per task while Kimi K3 scores 44 at $2.00 per task, an eight-fold cost gap for a two-point capability difference. …
DeepSeek is retiring its flagship deepseek-v4-pro model, routing all API requests to V4.1-Flash as of September 14, 2026, until V4.1-Pro ships. According to the company, testing by multiple parties found V4.1-Flash beats…
The Princeton Plasma Physics Laboratory (PPPL) announced a Pacman AI framework for controlling fusion systems safely that makes key decisions in milliseconds, according to a PPPL news post dated 2026. The framework is de…
AI tutoring startup Aristotle argues that most AI tutors are generic chatbots that fail students because they hand over answers rather than guide reasoning, and says its voice-first system is built on the ICAP engagement…
SkyDiscover-Synthesize (SkySynth), an autonomous pipeline that builds just-in-time specialized systems, synthesizes key-value stores up to 2.3× faster than Redis and FASTER, inference engines with up to 2.2× higher throu…
Researchers introduced the KuaiRP series of role-playing models in a technical report, using a multi-stage training pipeline to inject deep domain knowledge without losing general agent capabilities. The pipeline combine…
A multi-agent framework that automatically generates Quadratic Unconstrained Binary Optimization (QUBO) formulations from natural-language problem descriptions achieved 68% accuracy on QUBOBench, a new benchmark of 100 c…
A multi-stage rule-chaining framework solved 230 of 240 ARC-AGI-2 test tasks and exceeded 95 percent overall accuracy, according to an arXiv paper (2609.10654v1) presenting the system. The framework combines a determinis…
A study posted to arXiv (2609.10657v1) mapped the memorization-to-generalization boundary across 384 configurations of two-hidden-layer MLPs on modular arithmetic, fitting a power-law scaling relation for grokking onset …
A meta-agent variant achieved the highest Avg@3 reward on five of six heterogeneous benchmarks in a study of task-agnostic environment preprocessing, according to an arXiv paper (2609.10824v1). The research formalizes a …
A tail-risk-aware turn release scheduling method for agentic LLM workflows reduces the P95 of workflow flow time under contention by up to 3.50x, according to an arXiv paper (2609.10964v1). The method, which uses a mean-…
A new arXiv paper (2609.10992v1) analyzes the privacy-utility trade-off in Large Language Model interactions and introduces an intent-driven local protection framework built on a distilled lightweight model, Veilmind-4B.…
A new arXiv paper, arXiv:2609.11018v1, surveys AI agent evaluation across five dimensions of agenticness — environmental interaction, learning and adaptation, autonomy, goal-directed behavior, and temporal coherence — to…
Researchers introduced the Agent Incident Registry (AIR), a source-linked catalog of agent-related incidents disclosed from a first date through a last date, with each record carrying supporting evidence, a stable identi…
A new arXiv paper (2609.11061v1) proposes belief-shift branching, a fork-placement method for tree-structured reinforcement learning with verifiable rewards that forks a chain just before the step where the model's conse…
Researchers introduced Mosaic, a training-free framework that formulates GraphRAG retrieval as a per-query control problem, with an LLM analyzer converting query-specific evidence requirements into a bounded policy over …
A study of 70,500 X messages tied to securities class actions from 2002 to 2025 found that benchmark agreement between financial sentiment tools and human labels establishes semantic validity but does not by itself deter…
A study posted to arXiv (2609.11452v1) introduces RouteRepair, an instance-level failure diagnosis and targeted repair method for LLM-based automated heuristic design in routing optimization. On the traveling salesman pr…
Researchers proposed CoMA-DiT, a bidirectional cross-modal Diffusion Transformer for latent augmentation that treats paired modalities as mutual generative supervision, according to an arXiv paper (2609.11341v1). In expe…
Researchers introduced ActMap, a white-box representation that compresses a large language model's generation-time hidden-state trajectory into a fixed 12 × 32 × 128 tensor occupying 96 KiB, enabling single-pass uncertai…