Lightnews — Scholar-powered news

sakanaai.bsky.social @sakanaai.bsky.social · 7h

There’s a fairly wide gulf in capabilities both among different LLMs and different linguistic specifications, with it being notably easier for systems to deal with settings that are commoner cross-linguistically than those that are rarer.

PDF arxiv.org/abs/2510.07591
Code github.com/SakanaAI/IASC

IASC: Interactive Agentic System for ConLangs

We present a system that uses LLMs as a tool in the development of Constructed Languages. The system is modular in that one first creates a target phonology for the language using an agentic approach ...

arxiv.org

4

sakanaai.bsky.social @sakanaai.bsky.social · 7h

Our goals with IASC:

1/ We hope that these tools will be fun to use for creating artificially constructed languages.

2/ We are interested in exploring what LLMs ‘know’ about language—not what they know about any particular language, but how much they know about and understand linguistic concepts.

1 4

sakanaai.bsky.social @sakanaai.bsky.social · 7h

We are happy to announce the release of IASC, an Interactive Agentic System for ConLangs (Constructed Languages).

GitHub: github.com/SakanaAI/IASC

GitHub - SakanaAI/IASC: LLMs for Constructed Languages

LLMs for Constructed Languages. Contribute to SakanaAI/IASC development by creating an account on GitHub.

github.com

1 3

sakanaai.bsky.social @sakanaai.bsky.social · 7h

IASC: Interactive Agentic System for ConLangs

arxiv.org/abs/2510.07591

If you’re a fan of science fiction or fantasy, you’ve probably heard of made-up languages like Elvish from “The Lord of the Rings” or Klingon from “Star Trek.”

Can LLM agents create new artificial languages?

IASC: Interactive Agentic System for ConLangs

We present a system that uses LLMs as a tool in the development of Constructed Languages. The system is modular in that one first creates a target phonology for the language using an agentic approach ...

arxiv.org

1 1 6

sakanaai.bsky.social @sakanaai.bsky.social · 15d

By making ShinkaEvolve open-source, our goal is to democratize access to advanced discovery tools. We envision it as a companion to help scientists and engineers, building efficient, nature-inspired systems to unlock the future of AI research.

GitHub Project: github.com/SakanaAI/Shi...

GitHub - SakanaAI/ShinkaEvolve

Contribute to SakanaAI/ShinkaEvolve development by creating an account on GitHub.

github.com

6

sakanaai.bsky.social @sakanaai.bsky.social · 15d

ShinkaEvolve's efficiency comes from three key innovations:

1) Adaptive parent sampling to balance exploration and exploitation.

2) Novelty-based rejection filtering to avoid redundant work.

3) A bandit-based LLM ensemble that dynamically picks the best model for the job.

1 3

sakanaai.bsky.social @sakanaai.bsky.social · 15d

3/ LLM Training: It discovered a novel load balancing loss for MoE models, improving model performance and perplexity.

1 2

sakanaai.bsky.social @sakanaai.bsky.social · 15d

2/ Competitive Programming: On ALE-Bench, it improved an existing agent's solution, turning a 5th place result into a 2nd place leaderboard rank for one task.

1 2

sakanaai.bsky.social @sakanaai.bsky.social · 15d

We applied ShinkaEvolve to a diverse set of hard problems:

1/ AIME Math Reasoning: It evolved sophisticated agentic scaffolds that significantly outperform strong baselines, discovering a Pareto frontier of solutions trading performance for efficiency.

1 3

sakanaai.bsky.social @sakanaai.bsky.social · 15d

On the classic circle packing optimization problem, ShinkaEvolve discovered a new state-of-the-art solution using only 150 samples. This is a massive leap in efficiency compared to previous methods that required thousands of evaluations.

1 2

sakanaai.bsky.social @sakanaai.bsky.social · 15d

Many evolutionary AI systems are powerful but act like brute-force engines, burning thousands of samples to find good solutions. This makes discovery slow and expensive. We took inspiration from the efficiency of nature. ‘Shinka’ (進化) is Japanese for evolution.

1 2

sakanaai.bsky.social @sakanaai.bsky.social · 15d

We’re excited to introduce ShinkaEvolve: An open-source framework that evolves programs for scientific discovery with unprecedented sample-efficiency. It leverages LLMs to find state-of-the-art solutions, orders of magnitude faster!

Blog: sakana.ai/shinka-evolve/
Paper: arxiv.org/abs/2509.19349

1 2 13

sakanaai.bsky.social @sakanaai.bsky.social · Aug 30

How Sakana AI’s new evolutionary algorithm builds powerful AI models without expensive retraining
venturebeat.com/ai/how-sakan...

How Sakana AI’s new evolutionary algorithm builds powerful AI models without expensive retraining

M2N2 is a model merging technique that creates powerful multi-skilled agents without the high cost and data needs of retraining.

venturebeat.com

2 8

sakanaai.bsky.social @sakanaai.bsky.social · Aug 29

We are honored that Sakana AI’s CEO David Ha (@hardmaru.bsky.social) has been named to the TIME 100 AI 2025 list. Full List: time.com/time100ai

We’re truly grateful for the recognition and will continue our mission to build a frontier AI company in Japan.

Thank you for your support!

1 12

sakanaai.bsky.social @sakanaai.bsky.social · Aug 25

We hope this work sparks more interest in these under-explored ideas!

Please read our paper: arxiv.org/abs/2508.16204

Code: github.com/SakanaAI/nat...

Published in ACM GECCO’25: Proceedings of the Genetic and Evolutionary Computation Conference. DOI: doi.org/10.1145/3712...

Competition and Attraction Improve Model Fusion

Model merging is a powerful technique for integrating the specialized knowledge of multiple machine learning models into a single model. However, existing methods require manually partitioning model p...

arxiv.org

7

sakanaai.bsky.social @sakanaai.bsky.social · Aug 25

This approach is central to our mission. Rather than scaling monolithic models, we envision a future where ecosystems of specialized models co-evolve and combine, leading to more adaptive, robust, and creative AI. 🐙

1 4

sakanaai.bsky.social @sakanaai.bsky.social · Aug 25

Does it work on multimodal models?

We merged several text-to-image models by adapting them only for Japanese prompts. The resulting model not only improved on Japanese but also retained its strong English capabilities—a key advantage over fine-tuning, which can suffer from catastrophic forgetting.

1 3

sakanaai.bsky.social @sakanaai.bsky.social · Aug 25

Does it scale?

We used M2N2 to merge a math specialist LLM with an agentic specialist LLM. The resulting model excelled at both math and web shopping tasks, significantly outperforming other methods. The flexible split-point was crucial.

1 3

sakanaai.bsky.social @sakanaai.bsky.social · Aug 25

Does it work?

This is the first time model merging has been used to evolve models entirely from scratch, outperforming other evolutionary algorithms. In one experiment, M2N2 evolved an MNIST classifier from random networks that achieved performance comparable to CMA-ES, but more efficiently.

1 1

sakanaai.bsky.social @sakanaai.bsky.social · Aug 25

3/ Attraction and Mate Selection 💏: Merging is computationally expensive. M2N2 introduces an “attraction” heuristic that intelligently pairs models based on their complementary strengths, making the evolutionary search much more efficient.

1 2

sakanaai.bsky.social @sakanaai.bsky.social · Aug 25

2/ Diversity through Competition 🐠: To ensure a rich pool of models to merge, M2N2 makes them compete for limited resources (i.e., data points). This forces models to specialize and find their own “niche,” creating a population of diverse, high-performing specialists.

1 2

sakanaai.bsky.social @sakanaai.bsky.social · Aug 25

1/ Evolving Merging Boundaries 🌿: Instead of merging models using pre-defined, static boundaries, M2N2 dynamically evolves the “split-points” for merging. This allows a more flexible exploration of parameter combinations, like swapping variable-length segments of DNA.

1 2

sakanaai.bsky.social @sakanaai.bsky.social · Aug 25

Our new paper proposes M2N2 (Model Merging of Natural Niches), a more fluid method which overcomes this with 3 key, nature-inspired ideas:

1/ Evolving Merging Boundaries 🌿

2/ Diversity through Competition 🐠

3/ Attraction and Mate Selection 💏

1 3

sakanaai.bsky.social @sakanaai.bsky.social · Aug 25

A key limitation in earlier work remained: model merging required manually defining how models should be partitioned (e.g., by fixed layers) before they could be combined. What if we could let evolution figure that out too?

1 2

sakanaai.bsky.social @sakanaai.bsky.social · Aug 25

Our research follows an evolutionary path. We started with using evolution to find merge "recipes" (Nature Machine Intelligence), then explored maintaining diversity for new skills (ICLR 2025). Now, we're combining these ideas into a full evolutionary system.

1 2