cs.LGNov 17, 2026

What RL on Expert-Built Tasks Teaches Nemotron 3.5 Super

Authors: Sushant Mehta, Edwin Chen

Organizations: Surge AI

Abstract

We post-train NVIDIA's Nemotron 3.5 Super (120B total, 12B active parameters) with reinforcement learning (RL) on five expert-built task datasets, one at a time: real-world software engineering, complex constraints, long-horizon tool use, professional chart understanding and enterprise document reasoning. Each run trains a fresh rank-32 LoRA adapter with CISPO on Tinker, with no supervised stage, and each adapter is compared with the base model on the same tasks, harness and settings. On twelve headline benchmarks, eight built by other groups and four by Surge AI, the gains range from +4.5 points on IFBench to +13.8 on GDP_pdf, which rises from 10.4% to 24.2%; SWE-bench Pro V2 rises 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 document adapter, trained without API tools, also gains +8.0 points on AutomationBench. In attempts read side by side with the base model's on tasks each adapter newly solves, the base model's failures recur in a few modes, and each adapter does more often what the base model missed: the tool-use adapter finds the data and the rules before acting, the coding adapter writes its own test or scratch files on SWE-bench Pro V2, the constraints adapter carries earlier turns into its answers, the chart adapter measures the marks in code, and the document adapter reads AutomationBench's spreadsheets.

Figures & tables

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Oct 28, 2026

Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI

We introduce Nemotron 3.5 Super and Nemotron 3.5 Lightning, open and efficient Mixture-of-Experts Hybrid Mamba-Transformer models designed for reasoning, long-context understanding, and agentic workloads. Super is a 120B-parameter, 12B-active multimodal model, while Lightning is a smaller 30B-parameter, 3B-active text model; both designed for efficient deployment. Both models combine large-scale distillation, long-context training, and native speculative decoding, while Super further incorporates Indic continued pretraining and agentic software-engineering mid-training. We develop a comprehensive post-training pipeline spanning supervised fine-tuning and reinforcement learning (RLVR and RLHF), complemented in Super by specialized teacher training and multi- teacher on-policy distillation (MOPD). This transforms Super from a text-only base model into a multimodal agent capable of reasoning over text, images, video, and interactive tool environments. We further explore test-time scaling through parallel agent execution and self-verification. Across training and deployment, we jointly optimize capability and efficiency. Indic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points, while agentic mid-training strengthens software-engineering capabilities and robustness across agent harnesses. The post-trained Super model achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra despite using approximately 4.6x fewer active parameters. Post-training MTP boosting improves speculative- decoding acceptance length by up to 21.8% under probabilistic draft sampling, while NVFP4 quantization achieves 99.62% median accuracy recovery relative to BF16. Inference optimizations enable up to 4,856 output tokens/s/GPU on GB200 while sustaining at least 50 tokens/s/user.
Apr 14, 2026

KnowRL: Boosting LLM Reasoning via Reinforcement Learning with Minimal-Sufficient Knowledge Guidance

RLVR improves reasoning in large language models, but its effectiveness is often limited by severe reward sparsity on hard problems. Recent hint-based RL methods mitigate sparsity by injecting partial solutions or abstract templates, yet they typically scale guidance by adding more tokens, which introduce redundancy, inconsistency, and extra training overhead. We propose KnowRL (Knowledge-Guided Reinforcement Learning), an RL training framework that treats hint design as a minimal-sufficient guidance problem. During RL training, KnowRL decomposes guidance into atomic knowledge points (KPs) and uses Constrained Subset Search (CSS) to construct compact, interaction-aware subsets for training. We further identify a pruning interaction paradox -- removing one KP may help while removing multiple such KPs can hurt -- and explicitly optimize for robust subset curation under this dependency structure. We train KnowRL-Nemotron-1.5B from OpenMath-Nemotron-1.5B. Across eight reasoning benchmarks at the 1.5B scale, KnowRL-Nemotron-1.5B consistently outperforms strong RL and hinting baselines. Without KP hints at inference, KnowRL-Nemotron-1.5B reaches 70.08 average accuracy, already surpassing Nemotron-1.5B by +9.63 points; with selected KPs, performance improves to 74.16, establishing a new state of the art at this scale. The model, curated training data, and code are publicly available at https://github.com/Hasuer/KnowRL.
Nov 6, 2026

From LoRA to Language: Recovering Skill Documents from Adapters

Agent skills are short documents that teach a language model how to handle a class of tasks. To save context, these documents are increasingly compiled into low-rank adapters (LoRA), which keep the behaviour but lose the text: an adapter is hard to read, edit, audit or move to another model. We ask whether the skill document can be written back from the adapter and its frozen base model, and we score each recovered document by the share of the original's key instructions it states and by the tasks it solves. Reading the weights alone recovered little with the methods we tried; a learned map from weights to text passed on about one bit of the document. Running the adapter worked much better. The Executing Writer, a small writer trained on top of the adapted model, recovers 39-44% of the key instructions of held-out skills, against 1.3-1.5% with another skill's adapter, and on adapters fitted to 41 external skills its documents restore 85-91% of the original's task gain. These strongest results are on adapters whose training stored the document; adapters trained on behaviour alone are much harder for trained writers (1-2% of instructions). For these adapters we test an agent that runs the base and adapted models and writes a skill from where they differ. On ALFWorld its skills recover 58-71% of a LoRA's gain and carry over to models that cannot load the LoRA, lifting Llama-3.1-8B from 5% to 55%. Whether such a skill comes from the adapter depends on what the adapter learned: when its behaviour cannot be guessed, the adapter is the source; when it can, a reward loop without the adapter does as well.