← News

ResearchSources: arXiv · 18 Nov 2026

Surge AI reports RL adapters lift NVIDIA Nemotron 3.5 Super across five task areas

A paper on arXiv describes reinforcement learning runs that add small adapters to the 120B-parameter model, with benchmark gains of roughly 4.5 to 13.8 points.

In plain words

Surge AI researchers say they improved NVIDIA's Nemotron 3.5 Super model by training it with reinforcement learning, a method where the model learns by practicing tasks rather than by copying examples. They trained five separate small add-on modules, each focused on one job such as coding or reading documents. On their reported tests, each module scored higher than the base model on the same tasks.

What was announced

According to the paper, each run trained a fresh rank-32 LoRA adapter using a method called CISPO on the Tinker platform, with no supervised training stage beforehand. The five adapters target real-world software engineering, complex constraints, long-horizon tool use, professional chart understanding, and enterprise document reasoning. Every adapter was measured against the unmodified model on identical tasks, harness, and settings, the authors state.

Across twelve headline benchmarks, eight from other groups and four built by Surge AI, the reported improvements span from 4.5 points on IFBench to 13.8 points on GDP_pdf, where the score moves from 10.4% to 24.2%. Other cited gains include SWE-bench Pro V2 rising from 86.9% to 92.1%, Terminal-Bench 2.1 from 63.9% to 70.3%, MultiChallenge from 58.3% to 66.6%, AutomationBench from 14.8% to 27.5%, and Chartography from 39.4% to 50.6%. The authors also describe qualitative patterns: the tool-use adapter checks data and rules before acting, the coding adapter creates its own test or scratch files, the constraints adapter carries earlier conversation turns forward, the chart adapter measures chart marks in code, and the document adapter, trained without API tools, still gains 8.0 points on AutomationBench by reading its spreadsheets.

What it does not show

  • The evidence is the paper's abstract only; the full article's methods, setup details, and caveats are not available for verification.
  • All benchmark results and behavioral observations come from the authors themselves; no independent evaluation or confirmation is included in the supplied material.
  • The abstract does not state whether the adapters or training setup are publicly available.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.