RightSignal · Tuesday, October 20

Fine-tuning large AI models with small add-ons to predict brain activity

Zhuofan Josh Ying, Guanglong Sun, Nolan Ho, Kieran Liu, Rui Yang, Andrew Chen · Submitted Oct 19, 2026 · 231 upvotes yesterday

Preprint · not peer reviewedTrending

Brain encoding models try to predict how the brain responds to sights and sounds. Brain data is scarce, so big models are usually frozen and only a small readout is trained. The authors instead attach small trainable add-ons called LoRA adapters to models with up to 104 billion parameters. They also predict several time points in one pass to save compute.

Key finding: Their method, BrainLoRA, improved results across nine models and set the best published score on the Algonauts 2025 Challenge. An ensemble of more than 200 adapted models ranked first on the public leaderboard.

Why it matters: It shows that large AI models can be adapted cheaply even when training data is scarce and costly.

Read on arXiv

A benchmark built from real paid consulting jobs tests whether AI agents can do expert planning work

Peisong Wang, Feixue Liu, Yize Meng, Bowen Liu, Ruoyu Chen, Gaoyang Dai · Submitted Oct 16, 2026 · 151 upvotes yesterday

Preprint · not peer reviewedTrending

OR-License has 31 tasks from real operations research projects, such as scheduling and planning work, that clients paid $3.68 million for in total. Agents get the original client materials and a simulated client to question, and their answers are compared with the expert solutions the clients accepted.

Key finding: Across 11 frontier models, some beat the experts on solution quality when their answers worked, but none matched the experts overall. The best Claude Code setup produced valid solutions 88.39% of the time.

Why it matters: It tests AI agents on messy real business work instead of tidy textbook problems, and shows where they still fall short.

Read on arXiv

MemT-1.0 builds memory into the AI model itself instead of rereading the whole history

Zhiyu Li, Zeyu Zhang, Linyun Liu, Yiwei Bu, Chenyang Xi, Xichong Zhang · Submitted Oct 19, 2026 · 109 upvotes yesterday

Preprint · not peer reviewedTrending

The authors design a model family that stores what it has seen inside its own computation. It does not keep the full conversation history in a cache. They release two sizes, Nano and Flash, trained on 34 billion tokens.

Key finding: With a history of 128K tokens, MemT-1.0-Nano ran about 1.86 times faster end to end and needed 66 times less stored data than keeping the full cache. It stayed competitive with retrieval-based methods.

Why it matters: Long-running assistants and agents need memory that does not get slower and more expensive as the conversation grows.

Read on arXiv

In mixed AI model designs, one layer type can stop another from learning to look things up

Xingyu Shen, Yingfa Chen, Ziqing Qiao, Chaojun Xiao, Xu Han, Zhiyuan Liu · Submitted Oct 19, 2026 · 7 upvotes yesterday

Preprint · not peer reviewedTrending

Many new language models mix fast recurrent layers with attention layers. The authors find the two compete to learn the same skill. The recurrent layers learn to retrieve information first, so attention rarely learns it, even though attention is better at it.

Key finding: Pausing updates to the recurrent layers partway through training lets attention learn retrieval. In models up to 1 billion parameters, this improved results on language modeling, reasoning and retrieval tests.

Why it matters: A simple change to the training schedule could make faster hybrid models better at recalling facts from long inputs.

Read on arXiv

KronLSTM, a new update rule that lets AI models revise memories without overwriting others

Anamaria-Roberta Hartl, Levente Zólyomi, Michael List, Günter Klambauer, Sebastian Böck, Sepp Hochreiter · Submitted Oct 19, 2026 · 28 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

KronLSTM adds a new memory update rule to the xLSTM family of models. When it corrects one stored association, it changes the memory in only one direction, so it disturbs other stored information less. The authors prove the memory stays stable without extra normalization tricks.

Key finding: KronLSTM matches and often beats existing linear attention models, including the DeltaNet family. Its gains are largest on long-context tests.

Why it matters: Stable memory with a solid mathematical basis could make cheaper long-context models more reliable. Code and models are open source.

Read on arXiv

Your RightSignal

Topics in News

Fine-tuning large AI models with small add-ons to predict brain activity

A benchmark built from real paid consulting jobs tests whether AI agents can do expert planning work

MemT-1.0 builds memory into the AI model itself instead of rereading the whole history

In mixed AI model designs, one layer type can stop another from learning to look things up

KronLSTM, a new update rule that lets AI models revise memories without overwriting others

That’s every paper.