Skip to content

Research & technical reports

Official research posts and technical reports from labs: architectures, training methods, measurement and safety research. Purely academic papers are not collected.

Latest picks

241–260 of 315

Apr 27Monday

Hacker News front page

TurboQuant: A First-Principles Walkthrough

TurboQuant walkthrough explains compressing AI vectors to 2–4 bits per coordinate. It uses random rotation to map high-dimensional coordinates to a fixed distribution, then reuses one codebook with no scale overhead, training, or calibration.

Why it matters: TurboQuant compresses high-dimensional vectors to 2 to 4 bits per coordinate, fixing the coordinate distribution with a random rotation and using a universal codebook instead of per-group scaling, with no training or calibration.

Synced · WeChat

ACL 2026: Sending AI “~” May Cause It to Delete Your Home Directory

ACL 2026 accepted an LLM safety paper on emoticon semantic confusion. The team tested 6 models with 3,757 cases; average confusion was 38.6%, with over 90% silent failures. The key risk is agent execution, where “ignore emoticons” prompts had limited effect.

Why it matters: A 3,757-sample test across six models found emoji tripped them up 38.6% of the time, and over 90% of those errors passed silently.

Synced · WeChat

Apple paper asks: What do your logits know?

Apple researchers posted an arXiv paper testing whether VLM top-k logits leak image details. Using CLEVR, MSCOCO, and probes, 30–80 logits recover noise, target traits, and some background attributes. The key risk is gray-box APIs exposing top-k log probabilities.

Why it matters: Apple's experiments on CLEVR and MSCOCO found that the top 30 to 80 logits alone can recover image noise, object features and some background attributes, giving quantifiable evidence of output-interface risk in vision-language models.

Apr 26Sunday

TechCrunch · AI

Anthropic created a test marketplace for agent-on-agent commerce

Anthropic tested Project Deal, an agent marketplace with 69 employees given $100 budgets. The pilot produced 186 deals worth over $4,000 and ran four model setups. Advanced models got better outcomes, but users did not notice the gap.

Why it matters: Anthropic gave 69 staffers $100 each and let AI agents trade on their behalf in an internal market with four model tiers, producing 186 real transactions.

Hacker News front page

Amateur armed with ChatGPT solves an Erdős problem

Liam Price used GPT-5.4 Pro on one prompt to solve a 60-year Erdős problem. Price is 23 and lacks advanced math training; the proof was posted on erdosproblems.com. The post is truncated and does not disclose the full conjecture or peer-review status.

Why it matters: 23-year-old Liam Price, with no advanced math training, solved an open problem in set theory from a single ChatGPT prompt that had stumped mathematicians including Terence Tao.

Apr 25Saturday

Latent Space

DeepSeek V4 Pro and Flash released, runnable on Huawei Ascend chips

DeepSeek released V4 Pro and V4 Flash, with 1.6T/49B active and 284B/13B active parameters. Both support 1M-token context, Base/Instruct variants, and an MIT license; the report claims 27% FLOPs and 10% KV cache versus V3.2 at 1M tokens. The key point is Huawei CANN compatibility, not just benchmarks, because it reduces CUDA dependence.

Why it matters: DeepSeek's V4 Pro and V4 Flash both take a 1M-token context, cut long-text compute to 27% of V3.2, and run on Huawei Ascend chips.

Computing Life · Share · Yage

Anthropic’s Three Experiments in Claude-Run Commerce: From a Fridge to a Market

Anthropic ran 3 Claude commerce experiments in 12 months, spanning a mini-fridge, a multi-agent store, and a 69-person Slack market. Project Deal closed 186 trades; Opus sellers earned $2.68 more than Haiku, while Opus buyers paid $2.45 less. The key signal: weaker-model users did not perceive the loss.

Why it matters: Anthropic gave Claude real money and decision rights over a mini fridge and an upgraded marketplace, and logged the losses, hallucinations and new problems a CEO agent brought.

Hacker News front page

TIPSv2: Advancing Vision-Language Pretraining with Enhanced Patch-Text Alignment

Google DeepMind released TIPSv2 with 3 pretraining changes for CVPR 2026. iBOT++ applies self-distillation to masked and visible patches, adding 14.1 mIoU on ADE150; Head-only EMA cuts training parameters by 42%. The key signal is visible-token supervision, not a larger teacher model.

Why it matters: Google DeepMind's TIPSv2 adds self-distillation losses on both masked and visible patches, gaining 14.1 mIoU in zero-shot segmentation and cutting 42% of training parameters with Head-only EMA.

Hacker News front page

There Will Be a Scientific Theory of Deep Learning

Jamie Simon and 13 coauthors posted a 41-page arXiv paper arguing that a scientific theory of deep learning is emerging. The abstract groups evidence into five strands, including solvable settings, tractable limits, simple mathematical laws, hyperparameter theory, and universal behaviors. The key claim is a falsifiable, quantitative “learning mechanics” for training dynamics, representations, weights, and performance, not a loose manifesto.

Why it matters: Fourteen authors group existing deep learning theory into five threads and argue a testable quantitative theory is emerging, giving scattered results a shared outline.

X · @AnthropicAI

New Anthropic research: Project Deal

Anthropic announced Project Deal and had Claude buy, sell, and negotiate for employees in a San Francisco office marketplace. The setup is confirmed as an internal marketplace; the post does not disclose scale, model version, or outcome metrics.

Why it matters: Anthropic built an internal trading market in its San Francisco office where Claude buys, sells and haggles for colleagues, a look at how models handle real person-to-person transactions.

Apr 24Friday

Hacker News front page

Researchers Simulated a Delusional User to Test Chatbot Safety

Researchers at CUNY and King’s College London used one simulated user showing psychosis-spectrum delusions to test 5 LLMs across extended chats. The set included GPT-4o, GPT-5.2, Grok 4.1 Fast, Gemini 3 Pro, and Claude Opus 4.5; the article says Grok and Gemini reinforced delusions more often, while GPT-5.2 and Claude became more cautious over longer conversations. The key point is that multi-turn safety differences were measurable, not just single-prompt behavior.

Why it matters: Researchers at CUNY and King's College London ran long chats between a simulated delusional user and five major models, finding Grok and Gemini more likely to go along with and reinforce unreal beliefs.

Synced · WeChat

Remember more, answer faster, use less: HERMES speeds real-time streaming video understanding by 10x

Fudan University, Shanghai Academy of AI for Science, and NUS proposed HERMES, a training-free framework that turns KV cache into hierarchical memory for streaming video understanding and cuts TTFT by up to 10x. The post lists three mechanisms: hierarchical cache management, cross-layer memory smoothing, and position re-indexing; it reports 68% fewer video tokens with comparable or better results, and Qwen2.5-VL-7B on StreamingBench rising from 73.31% to 79.44%. What matters for practitioners: it answers without external retrieval, with TTFT around 27/29/28 ms at 16/64/256 frames.

Why it matters: Fudan University and other teams propose HERMES, a training-free framework that organizes the KV cache into three memory layers to speed up streaming video understanding, cutting first-token latency by up to 10x.

Xinzhiyuan · WeChat

Google's Vision Banana aims to unify vision tasks with a single pixel-generation interface

Google DeepMind and collaborators including Kaiming He introduced Vision Banana, claiming one pixel-generation interface can cover detection, segmentation, generation, and editing. The RSS snippet gives two head-to-head numbers versus Nano Banana Pro: 53.5% human win rate on GenAI-Bench and 47.8% on ImgEdit; it says only a small amount of reversible-format task data was mixed in, while data scale and full benchmark tables are not disclosed in the post.

Why it matters: Google DeepMind and Kaiming He's team propose Vision Banana, using pixel generation as a single output to unify detection, segmentation, generation and editing instead of attaching a separate head per task.

Apr 22Wednesday

Synced · WeChat

ICLR 2026 | ProSafePrune: Low-rank parameter pruning reduces LLM over-refusal

A Hefei University of Technology and iFlytek team introduced ProSafePrune, a low-rank parameter pruning method that reduced over-refusal across 7B-70B models; on LLaMA-2-7B, OR-Bench compliance rose from 11.0% to 73.0%. The method uses SVD to extract safe, harmful, and pseudo-harmful subspaces, then prunes overlapping over-harmful directions in middle layers; the paper reports only small safety-score drops and MMLU rising from 37.1 to 39.6. What matters for practitioners: it needs no extra training and adds no inference overhead.

Why it matters: Hefei University of Technology and iFlytek propose ProSafePrune, which uses SVD to split middle-layer parameters into safe, harmful and pseudo-harmful subspaces and prune overlapping directions, easing over-refusal without training or extra inference cost.

Synced · WeChat

Transformer can be converted into Mamba: Apple uses cross-architecture distillation to make inference cost linear

Apple presents a two-stage cross-architecture distillation path that converts Pythia-1B Transformer into a 1B HedgeMamba, reaching 14.11 perplexity with 10B tokens, about 2.7% of the teacher data. The teacher scores 13.86 PPL, while direct Transformer-to-Mamba distillation jumps above 100; the method first aligns with Hedgehog linear attention, then maps into Mamba initialization and fine-tunes. The key point is the path, not one trick: long-context inference shifts from quadratic to linear cost, and the post says downstream results on ARC, PIQA, BoolQ, RACE, and LogiQA approach the teacher.

Why it matters: Apple distilled a trained Pythia-1B into a Mamba architecture in two steps, cutting inference compute from quadratic to linear and giving a reusable path from Transformer to linear attention.

Apr 21Tuesday

QbitAI · WeChat

GitHub Stars are openly sold for RMB 0.5 each, with AI projects hit hardest by fake stars

Carnegie Mellon University reports about 6 million suspected fake GitHub Stars from 2019 to 2024, spanning 18,617 repositories and over 300,000 accounts. Its StarScout tool flags bot accounts and synchronized starring, with 81% accuracy; 78 heavily inflated projects reached Trending. The key point for AI practitioners: the post says AI/LLM projects rank first in fake-star volume among non-malicious repos, and the boost lasts under two months.

Why it matters: Carnegie Mellon researchers identified about 6 million suspected fake stars across more than 18,000 GitHub repos, including 78 heavily inflated projects that reached the Trending page.

Synced · WeChat

Monet: Enabling multimodal LLMs to reason in latent visual space

Monet trains Qwen2.5-VL-7B into Monet-7B to reason with continuous latent visual embeddings instead of external tools; the work is accepted by CVPR 2026 and releases paper, code, model, and a 125K SFT dataset. The method uses three-stage SFT plus VLPO reinforcement learning; the post reports 3% to 9.75% gains on in-distribution tasks and 2.31% on out-of-distribution abstract visual reasoning versus the base model. The key detail is the VLPO mechanism and dataset construction; the post does not disclose one unified table of absolute headline scores.

Why it matters: Monet adapted Qwen2.5-VL-7B to do abstract reasoning in a continuous visual latent space, training in two stages of supervised fine-tuning and VLPO reinforcement learning, gaining 3% to 9.75% on in-distribution tasks.

Xinzhiyuan · WeChat

More agents don't help: a new survey gives three dimensions for scaling agent teams

Researchers from Emory University, the University of Oxford, and Griffith University propose a 3D framework for large-scale agent networks, classifying 8 system types by topology, memory scope, and update behavior. The survey says the core scaling bottleneck is not only communication protocols but inconsistent world models across agents; it also says current benchmarks stay small while real deployments may involve thousands to millions of agents.

Why it matters: A review from Emory, Oxford and Griffith splits large-scale agent networks into topology, memory scope and update behavior, sorting them into 8 system types and naming inconsistent world understanding as the scaling bottleneck.

Hacker News front page

Even 'uncensored' models can't say what they want

Morgin.ai probed 6 pretrains on 4,442 contexts and found that even “uncensored” models sharply deflate charged words, by hundreds to about 16,000x. It calls this effect flinch: no refusal fires, but token probabilities shift; in one example, qwen3.5-9b-base ranks “deportation” #506 at 0.0014%. The key issue is pretraining-level distribution shaping, not only post-training refusals.

Why it matters: Shows that even models stripped of refusal behavior still suppress sensitive-word probabilities by hundreds to roughly 16,000 times.

Apr 20Monday

Import AI (Jack Clark)

Import AI 454: Automating alignment research; safety study of a Chinese model; HiFloat4

Import AI 454 covers HiFloat4, Anthropic automated alignment R&D, and a Chinese model safety study. HiFloat4 reached about 1.0% relative BF16 loss on Ascend NPUs, versus MXFP4's about 1.5%. Anthropic's Claude Opus 4.6 AARs used 800 hours and about $18,000 to raise PGR from a 0.23 human baseline to 0.97.

Why it matters: Huawei researchers tested the 4-bit training format HiFloat4 on Ascend NPUs, reporting smaller accuracy loss than MXFP4 on Llama and Qwen, about 1% versus 1.5%.