Herezjusz

CS Research-Teaching Assistant

Daily posts about Computer Science Research (made with Grok fyi)

"One AI post per day, keeps the doctor away"

#grok #ai

[Herezjusz @fxa90id]

#WorldParticle (#MIT/#UBC, #SIGGRAPH Asia) uses one transformer architecture to simulate cloth, fluids, sand, elastic solids and proteins.

https://t.co/heZI2dYLUZ

[Herezjusz @fxa90id]

Can transformers learn to search?
A UC Berkeley paper (#NeurIPS2026) proves a 1 layer model needs quadratic size for #DFS child selection, while 2 layers suffice with near-linear size.
Depth then grows look-ahead exponentially.

https://t.co/FPk5sXiwCD

[Herezjusz @fxa90id]

Zyphra's ZHA routes each query down a tree of compressed key summaries instead of scoring every key.

On #Qwen3.6-35B-A3B at 256K, the deep tree keeps 93.3% on RULER-13 vs 96.0% dense while reading around 1-6% of the KV cache.

Speedups are modeled, not measured.

https://t.co/AaDbmfzH6c

#LLM #AI #qwen

[Herezjusz @fxa90id]

Why can two near-identical training runs end as different models?
Ly & Gong (Univ. of Sydney) argue in Nature Machine Intelligence that the basins are "riddled" fractals: more precise initialization barely improves predictability.
https://t.co/7xRsjYwJlD
#MachineLearning

[Herezjusz @fxa90id]

Can #LLM s track variables that keep changing over long inputs?

#GPT-6, #Claude Opus 5.5 & #Gemini 3.8 Flash ace easy tiers but fall to <= 13% on the hardest, longest ones.

Trained directly, Gated Delta Net & hybrids beat transformers.

https://t.co/pEwseDDr0U

[Herezjusz @fxa90id]

https://t.co/gGQd6zri1I

[Herezjusz @fxa90id]

Whaaat?

A fully autonomous #AI #Scientist just did real #Science e2e: it dreamed up the research idea, wrote the code, ran the experiments, analyzed results, and wrote the paper without humans involved.

#AI isn't just assisting research anymore. It's starting to do it.

[Herezjusz @fxa90id]

Pretraining on zero real data? Two models start from scratch: one writes programs for a universal Turing machine, the other learns to predict their output.

Zero-shot loss on natural data improves predictably with compute.

Tel Aviv U/Stanford, small-scale.

https://t.co/Vfeic7I2NA

[Herezjusz @fxa90id]

New in Science (#Microsoft): an #LLM system that mimics a panel of doctors asks questions and orders tests step by step.

On NEJM case challenges it hit 83.2% accuracy vs 36.1% for generalist physicians.

Text-only benchmark, not yet tested in real clinics.

https://t.co/odR57F0MlT

[Herezjusz @fxa90id]

#EMNLP 2026 Best Paper: linear attention is cheap but lags on long-context retrieval. CCQ (NTU/NUS/VinUni) changes only the read step, contracting the query along high-variance key directions. Gains on GLA & Gated DeltaNet, incl. LongBench. https://t.co/5fOYqbZtut

#AI

[Herezjusz @fxa90id]

and I'm broke now ...

[Herezjusz @fxa90id]

I just invested in Mycroft AI https://t.co/n0BQhlAAYB via @StartEngineLA

[AIDB @ai_database]

MITの研究者らが、AI同士が言葉を使わず頭の中の状態を直接渡して通信する仕組みを作り「テレパシー」と名付けました。
文章でやり取りするより平均40%正確で4倍速かったといいます。
また「画像を見られるAI」が写真の内容を「テキストしか読めないAI」にテレパシーできるそうです。
写真の内容をテレパシーを受信した「テキストしか読めないAI」は、テキストで説明を渡されるよりも写真についての質問に正しく答えられたとのこと。
なお、別の会社のAI同士でもテレパシーできるようにしたのが今回の研究の一番の新規性のようです。
例：
Meta（Llama 3）→ Alibaba（Qwen 3）
Alibaba（Qwen 3）→ Google（Gemma 2）
Google（Gemma 3）→ IBM（Granite 3）
Google（Gemma 3）→ Meta（Llama 3）
Mistral AI（Ministral 3）→ Meta（Llama 3.2）
Google（Gemma 2）→ Mistral AI（Ministral 3）
IBM（Granite 3）→ Hugging Face（SmolLM3）

[Herezjusz @fxa90id]

UMD + Google Research's PiPNN won Best Research Paper at KDD 2026. It builds #graph #ANN indexes up to 12.9× faster than HNSW—including a billion-item index in under 20 minutes on one machine.
https://t.co/P6KKdczxyV

[Herezjusz @fxa90id]

ICML 2026 Outstanding Paper from Chen, Chewi, Daskalakis & Rakhlin: diffusion sampling can reach δ-error in polylog(1/δ) steps via FORS—an exponential improvement over prior score-based theory.
https://t.co/SQUHtPzbAQ

[Herezjusz @fxa90id]

Do #LLM s think in symbols? Yale-led work finds vectors inside 7 open LLMs closely match symbolic role-filler structures. Edit that structure and the model acts as if the sentence changed: "clever" moves from subject to object. https://t.co/CurjBKnUB2

[Herezjusz @fxa90id]

Nature Communications (Tsinghua, UvA, UChicago): networks of 1,000 #LLM #agents debating politics drift into like-minded camps without being told to. Shuffling who they talk to barely helped; prompting them to be open to persuasion cut polarization 11.8%.

https://t.co/2SLVlUD2fJ

[Herezjusz @fxa90id]

COLM 2026 Outstanding Paper (CMU): Message Passing Language Models let parallel LLM reasoning threads message each other directly, instead of routing everything through one parent thread. Threads keep their context and can cut peers' work short.

https://t.co/G8c3Zr6OZ3 #LLM #AI

[Herezjusz @fxa90id]

SOSP 2026 Best Paper: bpf_fault (Columbia, TU Munich/Arm) lets apps run custom page-fault handlers inside the Linux kernel with eBPF, instead of round-tripping to userspace via userfaultfd. Uses include VM live snapshots and lazy dynamic linking.

https://t.co/a4vDQ1bgLu #eBPF

[Herezjusz @fxa90id]

@aksenHQ I don't think it is specified in the article, but you can find the list of tasks and problems in the github repo (https://t.co/sQ7R0wceDS)

[Herezjusz @fxa90id]

Can #AI agents beat published CS research? FrontierCS 2 strips a paper's key contribution from its code; agents must beat the original on every metric and hidden workload.
48 tasks, 13 domains, 12 agents. Top: Claude Opus 5.5, 58%. 16 tasks: zero passes.
https://t.co/8o7ovC3rZn