AdaSR: Adaptive streaming reasoning framework with Hierarchical Relative Policy Optimization
Researchers introduce AdaSR, a framework enabling large reasoning models to reason incrementally during streaming input (e.g., audio/video) rather than waiting for complete context, then perform final deliberation once the stream ends. The core contribution is Hierarchical Relative Policy Optimization (HRPO), which decomposes policy optimization into streaming and deep reasoning phases with fine-grained per-phase advantage assignment, integrating format, accuracy, and latency-aware rewards. Experiments show AdaSR improves the tradeoff among reasoning accuracy, computational efficiency, and streaming latency over supervised fine-tuning baselines. Code is publicly released.
Related guides (3)
Related events (8)
StreamMA: Streaming communication in multi-agent reasoning reduces latency and improves accuracy
Researchers introduce StreamMA, a multi-agent reasoning system that streams individual reasoning steps to downstream agents as they are generated, rather than waiting for a complete chain. This pipelining approach reduces end-to-end latency and also improves accuracy by shielding downstream agents from error-prone late reasoning steps. Evaluated across eight benchmarks, two frontier LLMs (Claude Opus 4.6 and GPT-5.4), and three topologies, StreamMA outperforms serial and single-agent baselines by an average of 7.3 percentage points. The paper also identifies a 'step-level scaling law' — a new scaling dimension orthogonal to agent-count scaling.
ReasonAlloc: Hierarchical KV Cache Budget Allocation for Long-CoT Reasoning Models
ReasonAlloc is a training-free framework that reframes decoding-time KV cache compression as a hierarchical budget allocation problem, operating at both layer-wise (offline) and head-wise (online) levels. The method identifies an architecture-driven pattern called the 'Reasoning Wave' to guide layer preallocation, then dynamically reallocates to information-rich heads during decoding. Evaluated on MATH-500 and AIME 2024 using DeepSeek-R1-Distill and AceReason models, it outperforms uniform-budget baselines (R-KV, SnapKV, Pyramid-RKV) especially at small budgets of 128–512 tokens, with negligible overhead.
Adaptive Parallel Reasoning: The Next Paradigm in Efficient Inference Scaling
A BAIR blog post surveys recent progress in parallel reasoning for LLMs, covering methods from simple self-consistency and Best-of-N sampling through structured search (Tree of Thoughts, MCTS) to newer adaptive approaches including ParaThinker, GroupThink, and Hogwild! Inference. The core motivation is that sequential reasoning scales linearly with exploration depth, causing latency, context-rot, and compute inefficiency. Adaptive parallel reasoning aims to let models themselves decide when and how to decompose tasks into concurrent threads, rather than imposing fixed parallel structure externally. The post frames this as an emerging inference-time scaling paradigm with implications for agentic and complex reasoning workloads.
RA-RFT: Retrieval-Augmented Reinforcement Fine-Tuning teaches LLMs to reason by analogy
Researchers propose Retrieval-Augmented Reinforcement Fine-Tuning (RA-RFT), a post-training framework that trains a retriever to rank contexts by expected reasoning benefit rather than semantic similarity, then fine-tunes a policy model via reinforcement learning using retrieved analogous demonstrations. The key insight is that reasoning-relevant retrieval surfaces complementary solution strategies rather than superficially similar problems. On mathematical reasoning benchmarks, RA-RFT improves AIME 2025 average@32 accuracy by 7.1 and 2.8 points over GRPO for Qwen3-1.7B and Qwen3-4B respectively, suggesting reasoning-aware retrieval is orthogonal to reward design and training curriculum improvements.
AXPO: Agent Explorative Policy Optimization Addresses Thinking-Acting Gap in Multimodal Agentic Reasoning
This paper identifies a structural asymmetry in agentic reasoning called the 'Thinking-Acting Gap,' where tool use is attempted in only ~30% of rollouts under standard RL training (GRPO), and all-wrong tool-using subgroups suppress learning signals. The authors propose AXPO (Agent eXplorative Policy Optimization), which fixes the thinking prefix and resamples tool calls for all-wrong subgroups, combined with uncertainty-based prefix selection. Evaluated across nine multimodal benchmarks on Qwen3-VL-Thinking at multiple scales, SFT+AXPO outperforms SFT+GRPO by +1.8pp on both Pass@1 and Pass@4 at 8B, with the 8B model surpassing the 32B baseline on Pass@4 using 4× fewer parameters.
N-GRPO: Semantic Neighbor Mixing for Improved Policy Optimization in LLM Reasoning
A new arXiv preprint introduces N-GRPO, an exploration strategy for the GRPO reinforcement learning framework that improves solution diversity during rollout by mixing embeddings of anchor tokens with their nearest semantic neighbors rather than using token-level sampling or random noise. The method is evaluated on DeepSeek-R1-Distill-Qwen models of various sizes and shows consistent improvements on math reasoning benchmarks plus out-of-distribution generalization. The work targets a known limitation in RLHF-style training: redundant rollout trajectories that reduce effective learning signal.
ProtoAda: Prototype-Guided Adaptive Adapter Expansion for Multimodal Continual Instruction Tuning
ProtoAda is a new framework for Multimodal Continual Instruction Tuning (MCIT) that addresses a key failure mode in sparse Mixture-of-LoRA-Experts architectures: image-text similarity routing is format-blind and incorrectly merges tasks with similar semantics but different output structures (e.g., coordinate prediction vs. VQA). The method introduces format-aware task prototypes to guide both routing and adapter expansion, then consolidates compatible updates geometrically to reuse and refine existing parameters. Experiments across multiple benchmarks show improved performance, particularly on tasks whose answer formats are vulnerable to corruption by sequential fine-tuning.
ACTS: Agentic Chain-of-Thought Steering for efficient and controllable LLM reasoning
Researchers introduce Agentic Chain-of-Thought Steering (ACTS), a framework that formulates inference-time reasoning control as a Markov decision process, where a controller agent adaptively steers a frozen reasoner by issuing reasoning strategy directives and steering phrases at each step. The controller is initialized from synthetic steering trajectories with multi-budget augmentation and further optimized via reinforcement learning with budget-conditioned reward shaping. ACTS matches full-thinking performance with significant token savings and enables controllable accuracy-efficiency trade-offs across multiple benchmarks and reasoner models.