researchByteDance

Bytedance-led team argues three training steps are enough for AI agents, releases competitive open model

TL;DR

A team from Bytedance and several US universities argues in a new paper that three training steps are enough to build capable AI agents. The authors say stacked, multi-phase pipelines make it hard to tell which step helps and make reproduction expensive. According to The Decoder, the work is backed by a competitive open model.

1 min read
0

A team from Bytedance, the parent company of TikTok, and several US universities argues in a new paper that three training steps are enough to build capable AI agents. The team presents the claim alongside a competitive open model, according to The Decoder, which reported on the paper.

The problem the paper targets

AI models are increasingly expected to research the web on their own, run code, and solve multi-step tasks. According to the paper, the training methods behind these capabilities are becoming more nested, with many developers stacking numerous training phases one after another.

The authors argue this has created two problems:

  • It has become almost impossible to tell which training step actually contributes to the result.
  • Reproducing the work is getting expensive.

The proposed alternative

The team aims to show that a pipeline of three steps is sufficient. The Decoder's report says the authors back this with an open model that is competitive. The English-language report is behind a paywall on The Decoder's subscription tier, and the details of the three steps, the model's name, its size and its benchmark results are not covered here.

What this means

The paper's argument is about attribution and reproducibility. If a pipeline has many sequential phases, it is hard to isolate which one improves an agent's ability to browse, execute code and complete multi-step tasks, and the cost of repeating the work grows. A three-step recipe, if it holds up, would be easier to reproduce and to analyze. Because the authors released an open model alongside the claim, outside researchers can test it directly.

Related Articles

research

Nvidia team: huge context windows often overkill, as summaries and lookups matched them in many tests

An Nvidia research team argues that very large context windows are often overkill and expensive, according to The Decoder. In many tests, periodic summaries and targeted lookups worked about as well.

research

MiniMax reports its unreleased MIM model solved 94 TII McEliece challenge instances, 55 previously unsolved

MiniMax published research on November 20 describing cryptanalysis results from MIM, an unreleased internal system. The company says MIM solved 94 TII McEliece challenge instances, including 55 it describes as previously unsolved, and states the work does not affect deployed cryptographic standards.

research

Ai2 launches Theorizer in Asta and open-sources the AI agent that proposes scientific theories

Ai2 launched Theorizer, an AI agent that combines findings from scientific papers to propose explanations for observations, inside its Asta platform. The code is open-source, and Ai2 covers token costs up to a limit in the web interface. A lung cancer case study shows the workflow.

research

Salesforce's BLIP-Live video agent cuts unnecessary responses by over 50%, company claims

Salesforce AI Research introduced BLIP-Live, a general model for live video agents that decides when to speak and when to stay silent. According to Salesforce, in one evaluation across 326 ego-centric videos it reduced unnecessary responses of state-of-the-art models by over 50%.

Comments

Loading...