5arXiv cs.CL (Computation and Language)·
PyroDash: Token-level SLM-LLM collaborative inference with cost-aware policy learning
PyroDash is a framework that trains a small language model to emit control tokens during generation, triggering selective handoffs to a large LLM only when needed. The SLM policy is learned via a three-stage process culminating in Group Relative Policy Optimization with a reward balancing accuracy against normalized inference cost. On five mathematical reasoning benchmarks, the system achieves accuracy above LLM-only baselines while reducing cost by 20%+ at moderate settings, or cuts LLM calls to 0.012 per example at aggressive cost targets. The approach requires no separate router, LLM retraining, or access to LLM logits.
Related guides (3)
Related events (8)
Mask-Aware Policy Gradients improve RL training for Masked Diffusion Language Models
A new arXiv preprint introduces a two-stage action MDP formalization for applying reinforcement learning to Masked Diffusion Language Models (MDLMs), decomposing the policy gradient into a token prediction term and a masking order term. Prior approaches ignored the position-unmasking decision, leading to intractable log-likelihood estimates; the proposed method optimizes both terms jointly. The approach achieves 87.1% on GSM8K and 53.4% on MBPP, claiming state-of-the-art results for MDLM-based reasoning and coding.
LamPO: Lambda-Style Policy Optimization with Pairwise Decomposed Advantage for Reasoning LMs
LamPO proposes a new RLVR training objective that replaces GRPO's scalar group-relative advantages with a Pairwise Decomposed Advantage, aggregating pairwise reward gaps within response groups and weighting comparisons by confidence-aware log-probability differences. The method retains a critic-free, clipped-update PPO-style structure and optionally adds a ROUGE-L-based dense auxiliary reward to reduce sparsity. Experiments on AIME24, AIME25, MATH-500, and GPQA-Diamond using Qwen3-1.7B, Qwen3-4B, and Phi-4-mini show consistent improvements over GRPO and other RLVR variants with more stable training dynamics.
d-OPSD: First on-policy self-distillation framework tailored for diffusion LLMs
Researchers introduce d-OPSD, the first on-policy self-distillation (OPSD) framework designed specifically for diffusion large language models (dLLMs). The method addresses a fundamental mismatch between existing autoregressive OPSD approaches and dLLMs' arbitrary-order generation by using suffix conditioning on self-generated answers and step-level rather than token-level divergence supervision. Across four reasoning benchmarks, d-OPSD outperforms RLVR and SFT baselines while requiring only ~10% of the optimization steps of RLVR, suggesting strong sample efficiency gains for dLLM post-training.
AIR: Adaptive Interleaved Reasoning with Code in Multimodal LLMs via Reinforcement Learning
Researchers propose AIR, a system that trains multimodal large language models to adaptively interleave reasoning with code execution for numerical computation tasks, going beyond prior work that focused only on visual operations. The approach combines a two-stage cold-start data pipeline, RL dataset filtering, and a group-constrained reward function for tool-invocation decisions. Experiments show a 6.1 percentage point average improvement on evaluation benchmarks, with interleaved reasoning samples gaining 9.9 pp and tool-use success exceeding 95%.
TokenPilot: Dual-granularity context management cuts LLM agent inference costs by up to 87%
TokenPilot is a cache-efficient context management framework for LLM agents that addresses the trade-off between token sparsity and prompt cache continuity. It combines Ingestion-Aware Compaction (global prefix stabilization) with Lifecycle-Aware Eviction (local segment offloading) to reduce inference costs by 56–87% across benchmarks while maintaining competitive task performance. The system is evaluated on PinchBench and Claw-Eval and has been integrated into the open-source LightMem2 library.
Survey proposes four-layer architecture for token-operations-oriented LLM inference optimization
A new arXiv preprint introduces a four-layer technical architecture—Multi-model Fusion, Model Optimization, Compute-Model Fusion, and Compute-Network-Model Fusion—for systematically organizing LLM inference optimization techniques. The paper reviews key technologies and industry status at each layer and analyzes their application in real-world business scenarios. The framing around 'token operations' positions inference optimization as an operational discipline analogous to traditional IT operations.
ExpRL: RL-based mid-training using human QA data as reward scaffolds for LLM reasoning
ExpRL proposes an automated approach to LLM mid-training that replaces manually curated reasoning traces with large corpora of human-written QA data used as reward scaffolds rather than imitation targets. Reference solutions are hidden from the policy and used only to construct problem-specific grading rubrics, enabling dense process-level rewards that reinforce partial progress and intermediate reasoning steps. On challenging math reasoning benchmarks, ExpRL outperforms SFT, sparse-reward GRPO, and self-distillation as an RL initialization strategy, with additional mixed-domain experiments suggesting broader applicability.
GraphPO: Graph-based Policy Optimization reduces redundancy in LLM reasoning RL
GraphPO is a new reinforcement learning framework that represents reasoning rollouts as directed acyclic graphs rather than independent chains or trees, merging semantically equivalent reasoning paths into equivalence classes to share suffixes and reduce redundant exploration. The approach assigns efficiency advantages to incoming edges and correctness advantages to outgoing edges, deriving process supervision from outcome rewards. Experiments on three LLMs across reasoning and agentic search benchmarks show consistent improvements over chain- and tree-based baselines under equal token or response budgets. The method also provides theoretical guarantees on reduced advantage-estimation variance.