AI archive
Every AI story, newest first. 91 in this archive.
Aug 6, 2026
One Pass, Two Indexes: UEmbed Unifies Sparse and Dense Multimodal Retrieval in a Causal Decoder
UEmbed emits sparse lexical and dense vector embeddings from one decoder-only multimodal pass—closing the dense–sparse gap on MMEB-v2 without bidirectional retrofits or cross-modal bolt-ons.
Aug 5, 2026
Continuous Latents, Discrete Promise: AURORA-LM Tests Diffusion for Text Without Sacrificing Decode Fidelity
An arXiv paper introduces AURORA-LM, a continuous-latent diffusion language model that preserves decodable text representations and reports leading results among diffusion-based language models on OpenWebText and XSum.
Aug 4, 2026
Memory Is a Policy Problem: MemCon Learns When to Retrieve, Forget, and Inject Plans
UCLA-led MemCon treats agent memory as a learned MDP: a tabular bandit decides when to retrieve, inject plans, or forget — up to 15.2-point gains with 5–20% fewer tokens and zero extra LLM calls.
Aug 3, 2026
Memory Is Reconstructed, Not Retrieved: MRAgent Rewrites the Agent Memory Playbook
A June 2026 paper introduces MRAgent, a Cue–Tag–Content graph that lets LLM agents actively reconstruct long-horizon memory—lifting LoCoMo scores by up to 23% while cutting token use versus static retrieval baselines.
Aug 2, 2026
Inside the Model, Not Beside It: Metis Prototypes Native Memory for Foundation Models
MemTensor's Metis prototype embeds persistent memory states directly in a foundation model backbone — releasing code and checkpoints while acknowledging long-horizon degradation.
Aug 1, 2026
Ten Proofs in Lean: OpenAI's Astra Cracks a Decade of Stalled Mathematics
OpenAI's August 1 release documents ten stalled math and complexity problems solved by an internal Astra model, formalized in Lean — with explicit attribution ethics and a ~$2,000 inference cost estimate.
Aug 1, 2026
The Executable Loop: Frontis-MA1 and OpenMLE Put Recursive Self-Improvement on a Benchmark
FrontisAI's 35B Frontis-MA1 model and OpenMLE stack deliver open-weight, execution-grounded progress on recursive ML engineering—with weights and benchmarks released.
Aug 1, 2026
Open Weights at 2.8 Trillion Parameters: Kimi K3 Sets a New Bar for Deployable Frontier Models
Moonshot AI's Kimi K3 tech report on arXiv describes a 2.8T-parameter MoE model with open weights, 1M-token context, and native vision — trailing only Claude Fable 5 and GPT-5.6 Sol on the authors' eval suite.
Jul 31, 2026
Memory on Its Own Terms: Scaling Parametric Long-Term Memory Beyond the Base Model
A July 30 arXiv paper scales parametric long-term memory to 6.9B parameters, beating larger monolithic models on 17 benchmarks when indexing costs are tamed with distributed Faiss.
Jul 31, 2026
Bigger Models Build Fewer Tools: Lomekwi Paper Finds Inverse Scaling in Agent Recognition
A Yale-led arXiv preprint decomposes LLM tool discovery into curiosity, recognition, and efficiency — and finds recognition inversely scales with model size in controlled environments.
Jul 31, 2026
Skills That Evolve Themselves: Skill Self-Play Reconciles Verification and Open-Ended LLM Training
A new arXiv paper introduces Skill Self-Play, a co-evolutionary framework where proposers, solvers, and a dynamic skill library push LLM capability through verifiable agent skills rather than unbounded self-generation.
Jul 31, 2026
Open Frontier Intelligence: Moonshot's Kimi K3 Redefines the Open-Weight Ceiling
Moonshot AI's Kimi K3 ships as a 2.8T-parameter open MoE with 1M context, native vision, and 2.5× scaling efficiency over K2 — the first open 3T-class frontier model.
Jul 30, 2026
Timing Is a Measurement Problem: RL for Code Optimization Finally Works
Meta researchers show why RL for code speed fails under noisy timing rewards — and how DMC-Optim plus adapted GRPO lifts Qwen 7B and CWM 32B optimization pass rates without sacrificing correctness.
Jul 30, 2026
The Proof Breakers: Mythos, CryptanalysisBench, and AI's Cryptographic Turn
Anthropic's Mythos Preview found flaws in HAWK and reduced-round AES while launching CryptanalysisBench — a 191-task test of whether frontier models can do genuine cryptanalysis.
Jul 30, 2026
The Linguist Compressor: When Rules Beat Forward Passes on Inference Cost
An arXiv paper shows CPU-only linguistic rule compressors matching recent neural methods on multiple QA benchmarks — with honest degradation at aggressive ratios.
Jul 30, 2026
The Desktop World Simulator: ABot-World-0 Puts Interactive Rollout on One RTX 5090
ABot-World-0 claims 720p, 16 FPS interactive world rollout on a single RTX 5090 via LongForcing and unified keyboard control — a new desktop-scale world-model benchmark.
Jul 29, 2026
The Timing Signal: Reinforcement Learning for Code Optimization Crosses a Measurement Threshold
arXiv:2607.25970 shows how a three-stage RL pipeline makes execution-time rewards learnable, boosting Qwen 2.5 7B and CWM 32B optimization pass rates without sacrificing correctness.
Jul 29, 2026
The Context Bridge: Ellington et al. Unify Statistics, Meta-Learning, and Foundation-Model Routing
A July 25 arXiv paper by Caleb Ellington et al. proves that explicit parameter adaptation, meta-learning, and foundation-model routing can be equivalent to kernel ridge regression on joint input-context features — with audit metrics for deployment.
Jul 29, 2026
The Effect Cloud: FEGA Shows Sparse Autoencoder Features Rarely Steer Like Simple Directions
A July 27 arXiv paper introduces FEGA, showing that sparse autoencoder features split into value-like and pointer-like types — and that interpretable features rarely steer models along stable directions.
Jul 29, 2026
The Schedule Layer: LeSAMP Learns When Diffusion Models Should Change Their Sampling Knobs
LeSAMP, posted to arXiv on July 26, uses reinforcement learning to learn prompt-conditioned, timestep-varying diffusion sampling schedules—reporting up to 68% win rates over static baselines on Flux and SD3.5.