Research map / Large language models / Reasoning and chain of thought

Star Elastic: Many-in-One Reasoning LLMs with Efficient Budget Control

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Routing, pruning and distillation for reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Star Elastic post-trains a hybrid Mamba-Attention-MoE reasoning LLM (Nemotron Nano v3 30B) into nested 23B and 12B submodels with a learnable router and distillation, adds elastic budget control across thinking/answering phases, and produces nested FP8/NVFP4 checkpoints.

Problem

Training and storing a family of reasoning LLMs of different sizes is prohibitively expensive, and existing elastic/nested methods do not cover hybrid Mamba-Attention-MoE architectures or per-phase model selection during reasoning.

Approach

Extends Nemotron Elastic to MoE hybrids via importance ranking (including REAP for experts), a Gumbel-Softmax learnable router that picks nested sub-network configurations trained end-to-end with KD from the parent, two-stage short/long context curriculum, mixing submodels across thinking and answering phases, and nested quantization-aware distillation for NVFP4.

Key results

• Elastic-23B (2.8A) and Elastic-12B (2.0A) variants derived from one 160B-token run; the 30B elastic variant matches its parent (AIME-2025 88.54 vs 87.92, LiveCodeBench v5 72.70 vs 71.75).
• Elastic budget control yields up to 16% higher accuracy and up to 1.9x lower latency than standard Nano v3 budget control.
• Nested NVFP4 QAD recovers accuracy to 97.1-99.2% of BF16 per variant; elastic 12B NVFP4 reaches 7,426 tokens/s on RTX Pro 6000, 3.4x the 30B BF16 baseline.

Surprising findings

• Naive uniform budget sampling in extended-context training degrades the full-budget model while smaller budgets improve, motivating non-uniform sampling.
• Removing prompt masking improved performance by 1.5-2.5 points, and a 70% reasoning + 30% pretraining blend beat pure reasoning data.

Limitations

• Achieves only about 2.5x compression for the smallest variant; extreme compression ratios remain open.
• Wall-clock latency includes cache recomputation when switching nested models because inference frameworks lack cache reuse support.
• Heterogeneous per-layer configurations were not used in final models because vLLM does not support them.

Experimental setup

Datasets: open-source Nano v3 parent training data (reasoning post-training plus pretraining blend)

Benchmarks: AIME-2025 AIME-2024 GPQA LiveCodeBench v5 MMLU-Pro IFBench Tau Bench MATH-500

Metrics: accuracy latency throughput memory footprint

Baselines: Nemotron Nano v3 30B (independently trained parent) Qwen3-30B-A3B Minitron Minitron-SSM Nemotron Nano v3 default budget control

Builds on: Nemotron Elastic Minitron REAP Flextron MatFormer

Closest papers by meaning

• MILES: Modular Instruction Memory with Learnable Selection for Self-Improving LLM Reasoning (NeurIPS 2026 · similarity 0.88)
• Flextron: Many-in-One Flexible Large Language Model (ICML 2024 · similarity 0.88)
• Scalable Chain of Thoughts via Elastic Reasoning (ICLR 2026 · similarity 0.87)
• SABER: Switchable and Balanced Training for Efficient LLM Reasoning (AAAI 2026 · similarity 0.87)
• Puzzle: Distillation-Based NAS for Inference-Optimized LLMs (ICML 2025 · similarity 0.87)
• Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.87)