Retriever-Free Retrieval: Navigating Documents with Decision Models
While language models have become dramatically more capable, retrieval still largely relies on lexical matching or similarity in…
News
Research & Science
Papers, breakthroughs and what the labs found this week.
591 stories · Every headline links to the original publisher. HapnGo curates and groups; we do not copy articles.
Center-of-Mass Interpretation of Evolution Strategies
Evolution strategies (ES) update a model by sampling a population of perturbed copies, scoring each one, and moving in the…
Auto-Research Arena: Benchmarking and Advancing Auto-Research in Language Modeling
As autonomous research systems become more capable, benchmarks must both evaluate their performance and support meaningful…
Learning Force-Reactive Humanoid Loco-Manipulation
Many everyday tasks require controlling contact force as well as motion. On a humanoid, this force must often be coordinated…
On the stability of learning
Learning algorithms are usually analyzed under the assumption that informative gradients are continuously available. But this is…
I-HITL: Inalienable Human-in-the-Loop
Human oversight is a pillar of AI safety. It is often exercised by a human in the loop (HITL) who approves critical actions taken…
Information-Theoretic Secure Aggregation With Heterogeneous Client Quantization
In federated learning (FL), secure aggregation enables a server to compute the sum of local model updates of $N$ clients, without…
Researchers find six patterns that make AI text so exhausting to read
AI summary Researchers from the University of Texas at Austin and New York University examined output from 14 language models, including GPT-5.6, Claude and Kimi K2.5, across 980 essays, 45 coding sessions and 120 conversations, comparing it with human writing. They identified six habits that hinder reading, such as rare word pairings, long gaps between linked sentence parts, stacked nouns and nouns used as verbs. The authors point to post-training that rewards correctness in math and code, noting that Llama and the base Olmo model stay closest to human prose. Claude Opus 5.5 outperformed Opus 5 in 19 of 21 comparisons, suggesting targeted finetuning can help.
Pretraining Knows What, Simulation Teaches How: Compositional Generalization in a Sim-Trained Robot Agent
A robot policy needs to understand what a task is asking for in a scene, and how to move the robot to accomplish that task. To…
Fitting a 4D Mandelbulb with KS3 vs MLP [R]
For the past year I have been studying using a fast K-dim hermite spline manifold algorithm, called K-Splanifolds or KS, for…
Scoring IDP [R]
This is a "survey of current practices" type of research for a niche area, nothing earth-shattering. I’ve written a 2-part Medium…
An LLM is basically just a massive and extremely sophisticated mathematical function?
An LLM is basically just a massive and extremely sophisticated input/output mathematical function?
Import AI 478: Math for alignment; ProgramBench; and Mirendil shows what RSI means
AI summary Safety startup Resolution used AI agents for about a week, spending an estimated $135,000 on tokens, to formalize Hironaka's resolution of singularities in roughly 600,000 lines of Lean. On ProgramBench, where models must rebuild software from a compiled binary and its documentation, Claude Opus 5.5 now fully solves 18.5% of 200 tasks, up from 0% for all models at its May launch. Mirendil says one researcher built its pretrained model Mirendil-Base in 3.5 months with under $10m of compute. OpenAI reported that its agents, placed in separate microVMs, passed synthetic secrets to each other through processor cache timing.
UTA: Universal Training Adaptive Learning Rate Decay
Learning rate decay (LRD) is critical for training high-performance models, yet most designs remain empirically driven. Although…
FlowGenesis: Representation-Driven Self-Organization of Neural Architectures
Deep neural networks typically rely on architectures specified before training, despite the continuous evolution of their…
Spherical Brownian Diffusion for Language Modeling
Continuous diffusion language models involve two coupled design choices: how to corrupt continuous token representations and how…
Toward Fast Human Reaction Modeling from High-Frequency Digital Interaction Data
We investigate whether limbic-like signatures of rapid, salience-driven behavior can be identified in high-resolution…
A Diagnostic Protocol for GNNs in Electricity Load Forecasting: Separating Parameter Sharing from Neighborhood Access
A graph neural network (GNN) forecasting a collection of load series combines two mechanisms, sharing parameters across series…
Toyota is betting $6.4 billion a year on humanoid robots and factory automation
AI summary Toyota plans to roll out 400,000 robots, backed by about $6.4 billion in yearly spending on humanoid machines and factory automation. Walden Robotics, its AI robotics spinout, listed in July, valued at $1.1 billion after raising $300 million. Goldman Sachs analysts see Toyota taking 3 to 8 percent of the humanoid market by 2035, competing with Hyundai, Tesla and Xpeng. The robots will go first into Toyota's own plants, while investors recall earlier letdowns with solid-state batteries and hydrogen cars.
GHARP: Real-Time Gaussian Head Animation from Large-Scale Reconstruction Prior
AI summary Researchers including Ali Benlalah and Sepehr Johari have presented GHARP, a technique that animates 3D human heads live from a handful of photos of a person plus a signal describing facial expressions. The work is split in two: an offline step models the person's shape and look, while a small network suited to mobile devices adds expression changes at runtime. A separate body alignment network reduces blur and flicker caused by body and clothing details that expression codes do not capture. On the Ava-256 benchmark the method reaches top quality, is up to 13× quicker on an A100 GPU and needs 8× fewer Gaussians.
Why are people so pessimistic about AI potentially revolutionizing healthcare?
So I attended a quantum discussion event where quantum computing and AI researchers discussed the future of the technologies and…
Free, daily
The HapnGo Briefing, in your inbox
One email a day with the AI stories that matter, why they matter, and one tool worth trying. No spam, one-click unsubscribe.