Understanding and Mitigating Spurious Signal Amplification in Test-Time Reinforcement Learning for Math Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | a framework to mitigate spurious optimization signals is proposed for test-time reinforcement learning (TTRL) Reinforcement learning with verifiable rewards (RLVR) is an effective paradigm for improving large language models on structured challenging reasoning tasks. |
| Approach: | They propose a framework to mitigate spurious optimization signals from label noise . they propose to use a frequency-based sampling strategy to exclude ambiguous samples . |
| Outcome: | The proposed framework outperforms existing TTRL baselines on three large language models across multiple mathematical reasoning benchmarks. |
Similar Papers
What If Consensus Lies? Selective-Complementary Reinforcement Learning at Test Time (2026.acl-long)
Copied to clipboard
| Challenge: | Existing TTRL methods rely on positive pseudo-labeling strategies to enhance reasoning capabilities. |
| Approach: | They propose a test-time reinforcement learning framework that mitigates label noise amplification by deriving pseudo-rewards from majority voting consensus. |
| Outcome: | The proposed framework mitigates label noise amplification by implementing selective positive pseudo-labeling and entropy-gated negative p-labeled pruning. |
Beyond Majority Voting: Towards Fine-grained and More Reliable Reward Signal for Test-Time Reinforcement Learning (2026.acl-long)
Copied to clipboard
| Challenge: | RLVR is a paradigm for improving reasoning ability of large language models . but voting results often induce confirmation bias and suffer from sparse rewards . |
| Approach: | They propose a framework integrating model confidence and dynamic subgroup partitioning to address these issues. |
| Outcome: | The proposed framework outperforms recent baselines on multiple models and benchmarks. |
Selective Test-Time Debiasing for CLIP via Reward Gating (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for debiasing use uniform bias corrections across all input queries . weak debiases retains bias in sensitive queries, while weak dealiases in biased ones . |
| Approach: | They propose a framework that selectively applies debiasing based on input sensitivity . RG-TTA adaptively triggers fairness regularization based upon bias sensitivity of each input . |
| Outcome: | Experiments show that debiasing improves zero-shot performance while maintaining fairness . weak debiased queries distort semantically meaningful information while weak ones fail to mitigate stereotypes . |
Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data (2026.acl-short)
Copied to clipboard
| Challenge: | Strong base models saturate benchmarks, resulting in weaker performance, a paradox . a new approach to Reinforcement Learning (RL) is needed to improve performance . |
| Approach: | They propose a method that uses constrained uniform top-k sampling to flatten the local optimization landscape by sampling uniformly from constrained high-confidence candidates. |
| Outcome: | Experiments show that the proposed approach prevents policy degeneration and boosts out-of-domain generalization. |
QaRL: Rollout-Aligned Quantization-Aware RL for Fast and Stable Training under Training–Inference Mismatch (2026.findings-acl)
Copied to clipboard
Hao Gu, Hao Wang, Jiacheng Liu, Lujun Li, Qiyuan Zhu, Bei Liu, Binxing Xu, Lei Wang, Xintong Yang, Sida Lin, Sirui Han, Yike Guo
| Challenge: | Recent work has shown that reinforcement learning with simple rule-based reward functions (RLVR) can induce emergent reasoning behaviors and yield gains in challenging domains such as math problem solving. |
| Approach: | They propose a rollout-alignment-quantization-aware RL which aligns training-side forward with the quantized rollout to minimize mismatch. |
| Outcome: | The proposed approach outperforms quantized-rollout training by +5.5 on Qwen3-30B-A3B MoE for math problems while maintaining low-bit throughput. |
Debiasing Reward Models via Causally Motivated Inference-Time Intervention (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches for mitigating spurious features in RMs focus on response length . Existing methods focus on RM activation, resulting in performance trade-offs . |
| Approach: | They propose a method that uses neurons to suppress spurious features in RMs at inference time. |
| Outcome: | The proposed method reduces sensitivity to spurious features without inducing performance trade-offs on RM benchmarks. |
ConciseRL: Conciseness-Guided Reinforcement Learning for Efficient Reasoning Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models excel at complex tasks by breaking down problems into structured reasoning steps. |
| Approach: | They propose a hyperparameter-free conciseness score to guide models toward generating correct reasoning traces. |
| Outcome: | The proposed method reduces token usage by 31x on simple problems while improving accuracy by 7% on the hardest problems. |
Rewarding the Unlikely: Lifting GRPO Beyond Distribution Sharpening (2025.emnlp-main)
Copied to clipboard
| Challenge: | Reinforcement learning is emerging as a primary driver for improving language model reasoning capabilities. |
| Approach: | They propose a method for explicitly up-weighting rare but correct solutions to overcome rank bias in group relative policy optimization (GRPO) . |
| Outcome: | The proposed method mitigates rank bias and improves pass@N across a large range of N in both synthetic and real theorem proving settings. |
Decorrelate Irrelevant, Purify Relevant: Overcome Textual Spurious Correlations from a Feature Perspective (2022.coling-1)
Copied to clipboard
| Challenge: | Existing methods to debiase samples with biased features obstructs the model in learning from non-biased parts of the samples. |
| Approach: | They propose to eliminate spurious correlations in a fine-grained manner from a feature space perspective by using Random Fourier Features and weighted re-sampling to decorrelate dependencies between features. |
| Outcome: | The proposed method eliminates spurious correlations in a fine-grained manner from a feature space perspective. |
Hard Sample Aware Prompt-Tuning (2023.acl-long)
Copied to clipboard
| Challenge: | Prompt-tuning based few-shot learning has garnered increasing attention in recent years due to its efficiency and promising capability. |
| Approach: | They propose a framework to distinguish informative hard samples from misleading ones in model training. |
| Outcome: | The proposed framework achieves new SOTA results on a series of NLP tasks pushing the SST-5 accuracy to 49.5% (1.1% point absolute improvement), QNLI accuracy to 74.6% (1.9% absolute improvement) |