AI archive
Every AI story, newest first. 91 in this archive.
Jul 28, 2026
The Möbius Bridge: Claude Mythos Finds Cryptographic Weaknesses NIST Review Missed
Anthropic reports Claude Mythos improved attacks on the HAWK post-quantum signature candidate and reduced-round AES, releases CryptanalysisBench, and warns verification — not discovery — may become the bottleneck.
Jul 28, 2026
The Surrogate Policy: SLPO Brings Outcome-Reward RL to Latent Reasoning and Unlocks Test-Time Scaling Below the Token Layer
SLPO brings outcome-reward reinforcement learning to latent reasoners via a surrogate policy interface and adaptive stopping head, improving Pass@8 and Pass@16 in all 12 evaluated backbone–dataset settings.
Jul 28, 2026
The Marathon Grade: Long-Horizon Terminal-Bench Exposes Where Agent Demos Stop and Real Work Begins
Long-Horizon Terminal-Bench tests 46 stateful terminal tasks requiring hundreds of steps and hours of execution. Even frontier models pass fewer than one-third of tasks under strict criteria.
Jul 28, 2026
The Reactive Fault Line: Why LLM Agent Bugs Live at the Model-Harness Boundary
An arXiv empirical study of 255 agent-reactive bugs across Codex, Gemini-CLI, LangChain, and CrewAI finds failures that only appear when a specific LLM response meets a specific harness reaction—exposing gaps in detection, reproduction, and fault attribution.
Jul 27, 2026
The Drifting Target: Why LLMs Lose the Plot When User Intent Evolves Mid-Conversation
A July 2026 arXiv paper shows strong single-turn benchmark scores collapse when user intent evolves across a conversation — a failure mode static evaluation cannot see.
Jul 27, 2026
The Relation Graph: GraphVid Moves Video Control From Pixels to Structured Interactions
GraphVid, posted to arXiv on July 23, replaces trajectory drawing with structured interaction graphs for multi-object video control — and reports large quality gains over Motion-I2V with fewer parameters.
Jul 27, 2026
The Analogy Layer: Meta's RA-RFT Teaches Models to Retrieve by Reasoning Benefit, Not Surface Similarity
Meta Superintelligence Labs introduces Retrieval-Augmented Reinforcement Fine-Tuning, a three-stage framework that ranks exemplars by transferable reasoning structure rather than semantic overlap—lifting AIME 2025 scores by up to 7.1 points over GRPO on small Qwen3 models.
Jul 27, 2026
The Verification Loop: How BAAI's AREX Recursively Improves Deep Research Instead of Searching Longer
Beijing Academy of Artificial Intelligence's AREX treats deep research as a discovery–verification asymmetry: an inner loop gathers evidence, an outer loop audits constraints, and a learned context-update tool keeps verified findings alive across rounds. On six benchmarks, the 122B-A10B MoE variant reaches 82.5% on BrowseComp and 82.0 Item-F1 on WideSearch-en with only 10B activated parameters.
Jul 26, 2026
The Expanding Canvas: How Expanding Flow Maps Make Output Size a Learnable Variable
Sophia Tang and Pranam Chatterjee's Expanding Flow Maps framework decouple generative modeling from fixed canvas sizes, pairing an expand operator with a transport map to grow and denoise state across molecular, graph, and language domains.
Jul 26, 2026
The Width Bet: How Google’s PoTRE Turns Four Reasoning Topologies Into a 49.92% HLE Score
Google Cloud researchers propose PoTRE, a test-time framework that runs four structurally different reasoning agents in parallel—and reports state-of-the-art results on Humanity’s Last Exam, ARC-AGI-2, and PRBench Finance without scaling parameters.
Jul 25, 2026
The Explorer and the Late Landing: Google DeepMind Maps Why Reasoning Models Overthink
A Google DeepMind ACL 2026 study introduces TRACE, a structural analyzer that decomposes chain-of-thought traces into sub-thought graphs—and finds that over-verification and over-exploration, not raw length, drive most wasted inference on simple queries.