Challenge: Recent advances in mathematical reasoning typically rely on massive scale . yet, can strong reasoning capabilities be induced in small language models under extreme constraints?
Approach: They train small language models with a single GPU for under 24 hours . they find that adapters unlock significant plasticity in standard instruction-tuned models .
Outcome: The proposed model training on a single GPU (48GB) achieves 40% Pass@1 on AIME 24 (an 11.1% improvement over baseline) the model training results show that the adapter capacity and initialization are critical factors.

Similar Papers

Low-probability Tokens Sustain Exploration in Reinforcement Learning with Verifiable Reward (2026.findings-acl)

Copied to clipboard

Challenge: Recent studies show that RLVR training is slow and results plateau as policy entropy collapses . low-probability regularization (Lp-Reg) reduces the number of low-quality exploratory tokens induced by RL training .
Approach: They propose a method to reduce RLVR over-penalization by eliminating low-probability exploratory tokens . they propose 'Low-provability Regularization' to reduce the gradual elimination of low-quality exploratory entropy tokens.
Outcome: The proposed method eliminates low-probability exploratory tokens and prevents suppression of potentially valuable low-property candidates.
Beyond High-Entropy Exploration: Correctness-Aware Low-Entropy Segment-Based Advantage Shaping for Reasoning LLMs (2026.findings-acl)

Copied to clipboard

Challenge: Recent work studies RLVR through token entropy, arguing that high-entropies drive exploration and should receive stronger updates.
Approach: They propose a correctness-aware reinforcement framework that performs fine-grained advantage modulation over low-entropy segments.
Outcome: The proposed framework improves accuracy over strong RL baselines across three backbones and six math benchmarks while maintaining high-entropy exploration.
Steering LLM Reasoning Through Bias-Only Adaptation (2025.emnlp-main)

Copied to clipboard

Challenge: Compared with LoRA and BitFit, training a single steering vector per layer with reinforcement learning requires orders of magnitude fewer resources and isolates a much smaller, more interpretable parameter set.
Approach: They propose to train a single steering vector per layer with reinforcement learning while freezing all base weights to match the accuracy of fully RL-tuned reasoning models.
Outcome: The proposed approach improves on an 8 billion-parameter model while keeping all base weights fixed.
Beyond Reasoning Gains: Mitigating General-Capability Forgetting in Large Reasoning Models (2026.findings-acl)

Copied to clipboard

Challenge: Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning . however, the recipe introduces a significant risk of capability regression, where models forget foundational skills after prolonged training without employing regularization strategies.
Approach: They propose a replay strategy with dynamic objective reweighting for general knowledge preservation using short-horizon signals of convergence and instability.
Outcome: The proposed method preserves general capabilities and improves reasoning . it can be applied to existing RLVR pipelines without training additional models or tuning .
GeoRA: Geometry-Aware Low-Rank Adaptation for RLVR (2026.acl-long)

Copied to clipboard

Challenge: Existing parameter-efficient methods for RLVR face limitations . low-rank adaptation methods do not account for the distinct optimization dynamics .
Approach: They propose a low-rank adaptation method tailored for RLVR that exploits the anisotropic structure of RL update subspace and extracts its principal directions via Singular Value Decomposition (SVD).
Outcome: Experiments on large reasoning models show that GeoRA outperforms strong low-rank baselines across RLVR settings while showing stronger generalization and less forgetting on out-of-domain tasks.
LoNAS: Elastic Low-Rank Adapters for Efficient Large Language Models (2024.lrec-main)

Copied to clipboard

Challenge: Large Language Models (LLMs) reach hundreds of billions of parameters and require resources for training and inference stages.
Approach: They propose a low-rank adapter to reduce the number of trainable parameters in a model and reduce memory requirements.
Outcome: The proposed approach reduces memory and compute requirements while preserving performance.
Beyond Full Fine-tuning: Harnessing the Power of LoRA for Multi-Task Instruction Tuning (2024.lrec-main)

Copied to clipboard

Challenge: Low-Rank Adaptation (LoRA) is a parameter-efficient fine-tuning algorithm for large-scale language models.
Approach: They conduct a systematic study of Low-Rank Adaptation (LoRA) on diverse tasks and rich resources with different learning capacities.
Outcome: The proposed algorithm can achieve remarkable performance in high-resource and multi-task scenarios, even comparable to full fine-tuning.
How Much Knowledge Can You Pack into a LoRA Adapter without Harming LLM? (2025.findings-naacl)

Copied to clipboard

Challenge: Low-rank adaptation (LoRA) is a popular training technique for updating or domain-specific adaptation of Large Language Models (LLMs).
Approach: They propose to use low-rank adaptation to incorporate new facts into the LLM without compromising previously learned knowledge.
Outcome: The proposed approach is harmful because the model's performance declines after such fine-tuning.
Revisiting Entropy Regularization: Adaptive Coefficient Unlocks Its Potential for LLM Reinforcement Learning (2026.findings-acl)

Copied to clipboard

Challenge: Reasoning ability is a defining capability of Large Language Models (LLMs), but RLVR training suffers from policy entropy collapse, hindering exploration and limiting reasoning performance.
Approach: They propose a framework that dynamically balances exploration and exploitation via three components: difficulty-aware coefficient allocation, initial-anchored target entropy, and dynamic global coefficient adjustment.
Outcome: The proposed framework outperforms baselines on multiple mathematical reasoning benchmarks.
Revisiting Entropy in Reinforcement Learning for Large Reasoning Models (2026.findings-acl)

Copied to clipboard

Challenge: Reinforcement learning with verifiable rewards (RLVR) has emerged as a paradigm for enhancing the reasoning capabilities of large language models.
Approach: They propose a positive-advantage reweighting approach that regulates model entropy by adjusting the loss weights assigned to tokens with positive advantages during RLVR training.
Outcome: The proposed approach regulates model entropy by adjusting loss weights assigned to tokens with positive advantages during RLVR training while maintaining competitive performance.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations