Step-GRPO: Internalizing Dynamic Early Exit for Efficient Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Large reasoning models that use long chain-of-thought excel at problem-solving but waste computational resources. |
| Approach: | They propose a framework that internalizes dynamic early-exit capabilities directly into the model. |
| Outcome: | The proposed framework reduces token consumption by 32.0% on a Qwen3-8B model compared to the vanilla model . |
Similar Papers
Full-Step-DPO: Self-Supervised Preference Optimization with Step-wise Rewards for Mathematical Reasoning (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to improve long-chain mathematical reasoning focus on the first erroneous step, but ignore all other steps and rely heavily on external signals. |
| Approach: | They propose a DPO framework that leverages step-wise rewards from the entire reasoning chain instead of optimizing only the first erroneous step. |
| Outcome: | The proposed framework improves on in-domain and out-of-domain mathematical reasoning benchmarks. |
Beyond Token Length: Step Pruner for Efficient and Accurate Reasoning in Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing reinforcement learning methods for large reasoning models suffer from excessive verbosity, known as "overthinking." Existing models penalize generated tokens to promote conciseness, but these methods encounter two challenges: they may develop hacking behavior in later stages of training by discarding reasoning steps. |
| Approach: | They propose a framework that steers large reasoning models toward more efficient reasoning . they prioritize correctness while imposing penalties for redundant steps . |
| Outcome: | The proposed framework reduces token usage by 69.7% on AIME24. |
Stabilizing Efficient Reasoning with Step-Level Advantage Selection (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models generate long and verbose reasoning traces at inference time . short context post-training alone induces substantial reasoning compression . |
| Approach: | They propose a step-level advantage selection approach that reduces reasoning length by over 30% . they propose to use GRPO without any length-aware objective to train models in a shorter context window . |
| Outcome: | The proposed approach reduces average reasoning length by over 30% while improving Pass@1 accuracy by 3.79 points over the strongest length-aware baseline. |
Anti-Length Shift: Dynamic Outlier Truncation for Training Efficient Reasoning Models (2026.acl-long)
Copied to clipboard
Wei Wu, Liyi Chen, Congxi Xiao, Tianfu Wang, Qimeng Wang, Chengqiang Lu, Yan Gao, null Yiwu, Yao Hu, Hui Xiong
| Challenge: | Existing efficient reasoning methods rely on explicit length penalties for excessive verbosity on simple queries. |
| Approach: | They propose a training-time intervention that selectively suppresses redundant tokens . they find length shift occurs when models generate unnecessary reasoning on trivial inputs - a phenomenon that is often unexplored . |
| Outcome: | The proposed method reduces inference token usage by 78% while increasing accuracy compared to the initial policy and surpasses state-of-the-art efficient reasoning methods. |
Efficiently Learning To Reason or Not to Reason: Root-token Policy Optimization for Adaptive Thinking (2026.acl-long)
Copied to clipboard
| Challenge: | Large reasoning models (LRMs) externalize explicit reasoning traces before producing the answer, yet suffer from overthinking challenge. |
| Approach: | They propose a framework that enables large reasoning models to self-determine when to reason by training only the initial root token via group relative reward and group-wise advantages. |
| Outcome: | The proposed framework reduces training overhead and VRAM usage by focusing on the root token . it learns difficulty-aware adaptive thinking at just 2% of the training compute of prior methods. |
Token-Level Policy Optimization: Linking Group-Level Rewards to Token-Level Aggregation via sequence-level likelihood (2026.acl-long)
Copied to clipboard
| Challenge: | Group Relative Policy Optimization (GRPO) has significantly advanced the reasoning ability of large language models (LLMs). |
| Approach: | They propose a token-level framework that leverages sequence-level likelihood to link group-level rewards with individual tokens via token- level aggregation and introduces a KL-Divergence mask constraint that targets tokens with positive advantages and decreasing entropy to mitigate abrupt policy updates. |
| Outcome: | Experiments show that TEPO achieves state-of-the-art performance on mathematical reasoning benchmarks and reduces convergence time by 50% compared with GRPO/DAPO. |
When Is Thinking Enough? Early Exit via Sufficiency Assessment for Efficient Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to early exit reasoning often rely on handcrafted or empirical indicators that are unreliable and impractical. |
| Approach: | They propose a framework that allows LRMs to assess the sufficiency of its chain-of-thought and determine the optimal point for early exit. |
| Outcome: | The proposed framework reduces reasoning length by 28.9%–34.9% with minimal performance loss, effectively mitigating overthinking. |
MMR-GRPO: Accelerating GRPO-Style Training through Diversity-Aware Reward Reweighting (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have demonstrated remarkable capabilities in mathematical reasoning tasks. |
| Approach: | They propose to use Maximal Marginal Relevance to reweigh rewards of multiple rollouts by balancing rollout quality with diversity to reduce rollout redundancy. |
| Outcome: | The proposed approach reduces training time and costs by 47.9% . evaluations across three model sizes, three GRPO variants, and five mathematical reasoning benchmarks show that it achieves comparable peak performance while requiring on average 70.2% less wall-clock time. |
From Tokens to Steps: Verification-Aware Speculative Decoding for Efficient Multi-Step Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Speculative decoding (SD) allows a lightweight draft model to propose outputs that a stronger target model verifies. |
| Approach: | They propose a verification-aware speculative decoding framework that performs step-level verification using only model-internal signals. |
| Outcome: | Experiments show that SpecGuard outperforms both SD and reward-guided SD in accuracy and reliability tests. |
ConCISE: Confidence-guided Compression in Step-by-step Efficient Reasoning (2025.emnlp-main)
Copied to clipboard
Ziqing Qiao, Yongheng Deng, Jiali Zeng, Dong Wang, Lai Wei, Guanbo Wang, Fandong Meng, Jie Zhou, Ju Ren, Yaoxue Zhang
| Challenge: | Existing methods for fine-tuning-based compression suffer from verbose outputs, increasing computational overhead. |
| Approach: | They propose a framework to generate concise reasoning chains using Confidence Injection and Early Stopping. |
| Outcome: | The proposed framework reduces the length of the model by up to 50% while maintaining high task accuracy. |