AWS engineers fine-tuned a Qwen3.6-27B search agent using Amazon SageMaker AI multi-turn reinforcement learning (MTRL), optimizing full agent trajectories with nDCG@10 as the trajectory-level reward. The fine-tuned model improved retrieval quality on three of four held-out benchmarks and sharply cut failure rates, matching small-model cost with frontier-model-like reliability. The writeup serves as a practical recipe for enterprises training agents on proprietary tools.
An arxiv paper introduces OPDSearch+, which combines on-policy distillation with reinforcement learning refinement to address two core challenges in equipping small language models with search-augmented reasoning: the high cost of collecting multi-turn search trajectories and training instability when using off-the-shelf teachers without task-specific fine-tuning.
An arxiv paper introduces OPDSearch+, which combines on-policy distillation with reinforcement learning refinement to address two core challenges in equipping small language models with search-augmented reasoning: the high cost of collecting multi-turn search trajectories and training instability when using off-the-shelf teachers without task-specific fine-tuning.
An arXiv paper proposes replacing naive repeated sampling in LLM test-time compute with a learned search policy: a small concept generator, trained via reinforcement learning, emits diverse problem-specific concepts that condition a larger frozen answer generator. On hard math reasoning problems it improves pass@k at the same answer-generation budget, beats concepts from much larger untuned models, and transfers across model families.
An arXiv paper proposes replacing naive repeated sampling in LLM test-time compute with a learned search policy: a small concept generator, trained via reinforcement learning, emits diverse problem-specific concepts that condition a larger frozen answer generator. On hard math reasoning problems it improves pass@k at the same answer-generation budget, beats concepts from much larger untuned models, and transfers across model families.
Researchers propose Abstention-Aware Reinforcement Learning (AWA-RL) to mitigate hallucinations in LLM-based search agents. The method dynamically shapes abstention rewards based on the model's query-specific prior capabilities, addressing the failure of current RL training to penalize fabricated answers when retrieval fails.
Researchers propose Abstention-Aware Reinforcement Learning (AWA-RL) to mitigate hallucinations in LLM-based search agents. The method dynamically shapes abstention rewards based on the model's query-specific prior capabilities, addressing the failure of current RL training to penalize fabricated answers when retrieval fails.
SCOPE-RL (Scaffolded Chain Optimization with Process Efficiency) is a two-stage reinforcement learning framework that densifies sparse verifiable final-answer rewards to provide feedback on reasoning paths both before and after success. It addresses the limitation of RLVR where no reward signal is given for prerequisite progress on hard problems before success, and outcome rewards cannot distinguish well-organized correct trajectories from redundant or locally flawed ones after success.
SCOPE-RL (Scaffolded Chain Optimization with Process Efficiency) is a two-stage reinforcement learning framework that densifies sparse verifiable final-answer rewards to provide feedback on reasoning paths both before and after success. It addresses the limitation of RLVR where no reward signal is given for prerequisite progress on hard problems before success, and outcome rewards cannot distinguish well-organized correct trajectories from redundant or locally flawed ones after success.
AWS published a walkthrough building an AI product tagging system using Amazon SageMaker serverless model customization. It fine-tunes Qwen3-8B with supervised fine-tuning (SFT) and then reinforcement learning with verifiable rewards (RLVR) via GRPO on an Amazon sales dataset, raising the overall tagging score from 0.354 to 0.6941. The approach targets high-volume retail catalog enrichment where a stable schema and programmatically scorable outputs favor a small custom model over a frontier LLM.
AWS published a walkthrough building an AI product tagging system using Amazon SageMaker serverless model customization. It fine-tunes Qwen3-8B with supervised fine-tuning (SFT) and then reinforcement learning with verifiable rewards (RLVR) via GRPO on an Amazon sales dataset, raising the overall tagging score from 0.354 to 0.6941. The approach targets high-volume retail catalog enrichment where a stable schema and programmatically scorable outputs favor a small custom model over a frontier LLM.