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

SABER: Switchable and Balanced Training for Efficient LLM Reasoning

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: RL with verifiable rewards 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.

SABER trains DeepSeek-R1-Distill-Qwen models with GRPO on 2K math/code examples, assigning per-example thinking-token budget tiers and mixing in no-think examples, yielding four switchable reasoning modes (NoThink, FastThink, CoreThink, DeepThink).

Problem

Reasoning LLMs overthink, producing excessively long chain-of-thought for all problems regardless of difficulty, and lack user control over reasoning depth or a reliable no-thinking mode.

Approach

Profile each training example's base-model thinking-token usage to assign tiered token budgets (128, 4096, 16384, unlimited), guide the policy with budget system prompts and length-aware rewards under GRPO without SFT warmup, and add no-think examples so one model supports four inference modes.

Key results

• SABER-1.5B-FastThink reduces reasoning length by 72.7% vs DeepSeek-R1-Distill-Qwen-1.5B with ~3.0% average accuracy gain across benchmarks (MATH: 83.5 vs 80.6).
• SABER-1.5B-DeepThink improves MATH500 accuracy from 78.9 to 83.2 while cutting length from 8042 to 5353 tokens.
• On 7B, SABER-FastThink cuts reasoning length 82.1% but loses 2.5% accuracy; SABER-DeepThink reduces length 33.2% with 1.9% accuracy gain.

Surprising findings

• Removing budget downgrade or accuracy filtering yields higher accuracy but longer outputs, while aggressive downgrade of all samples reduces accuracy.
• NoThink accuracy of SABER (76.3 on MATH500) is substantially higher than the base model with reasoning disabled (65.1).

Limitations

• Depends on an assessor LLM for budget estimation and manual grouping into difficulty bands.
• Accuracy trade-off on 7B FastThink (2.5% drop) shows degradation at larger scale.
• Evaluated only on a limited set of tasks and a single base model family; scaling to heterogeneous real-world workloads untested.

Experimental setup

Datasets: OpenR1-Math-220k KodCode-Light-RL-10K MATH500 GSM8K MATH MBPP LiveBench-Reasoning

Benchmarks: MATH500 GSM8K MATH MBPP LiveBench-Reasoning

Metrics: accuracy average response length (tokens)

Baselines: L1 SelfBudgeter DeepSeek-R1-Distill-Qwen-1.5B DeepSeek-R1-Distill-Qwen-7B

Builds on: GRPO L1 SelfBudgeter DeepSeek-R1-Distill-Qwen verl

Closest papers by meaning

• General-Reasoner: Advancing LLM Reasoning Across All Domains (NeurIPS 2025 · similarity 0.91)
• Generative Adversarial Reasoner: Enhancing LLM Reasoning with Adversarial Reinforcement Learning (ICLR 2026 · similarity 0.91)
• In-Token Rationality Optimization: Towards Accurate and Concise LLM Reasoning via Self-Feedback (AAAI 2026 · similarity 0.90)
• Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective (NeurIPS 2025 · similarity 0.90)
• SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs (ICLR 2026 · similarity 0.90)
• Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning (ICML 2025 · similarity 0.90)