Where Do Tokens Go? Understanding Pruning Behaviors in STEP at High Resolutions

Vision Transformers (ViTs) achieve state-of-the-art performance in semantic segmentation but are hindered by high computational and memory costs. To address this, we propose STEP (SuperToken and Early-Pruning), a hybrid token-reduction framework that combines dynamic patch merging and token pruning to enhance efficiency without significantly compromising accuracy. At the core of STEP is dCTS, a lightweight CNN-based policy network that enables flexible merging into superpatches. Encoder blocks integrate also early-exits to remove high-confident supertokens, lowering computational load. We evaluate our method on high-resolution semantic segmentation benchmarks, including images up to 1024 x 1024, and show that when dCTS is applied alone, the token count can be reduced by a factor of 2.5 compared to the standard 16 x 16 pixel patching scheme. This yields a 2.6x reduction in computational cost and a 3.4x increase in throughput when using ViT-Large as the backbone. Applying the full STEP framework further improves efficiency, reaching up to a 4x reduction in computational complexity and a 1.7x gain in inference speed, with a maximum accuracy drop of no more than 2.0%. With the proposed STEP configurations, up to 40% of tokens can be confidently predicted and halted before reaching the final encoder layer.

ResidualViT for Efficient Temporally Dense Video Encoding

Several video understanding tasks, such as natural language temporal video grounding, temporal activity localization, and audio description generation, require "temporally dense" reasoning over frames sampled at high temporal resolution. However, computing frame-level features for these tasks is computationally expensive given the temporal resolution requirements. In this paper, we make three contributions to reduce the cost of computing features for temporally dense tasks. First, we introduce a vision transformer (ViT) architecture, dubbed ResidualViT, that leverages the large temporal redundancy in videos to efficiently compute temporally dense frame-level features. Our architecture incorporates (i) learnable residual connections that ensure temporal consistency across consecutive frames and (ii) a token reduction module that enhances processing speed by selectively discarding temporally redundant information while reusing weights of a pretrained foundation model. Second, we propose a lightweight distillation strategy to approximate the frame-level features of the original foundation model. Finally, we evaluate our approach across four tasks and five datasets, in both zero-shot and fully supervised settings, demonstrating significant reductions in computational cost (up to 60%) and improvements in inference speed (up to 2.5x faster), all while closely approximating the accuracy of the original foundation model.

Adaptive Pareto-Optimal Token Merging for Edge Transformer Models in Semantic Communication

Large-scale transformer models have emerged as a powerful tool for semantic communication systems, enabling edge devices to extract rich representations for robust inference across noisy wireless channels. However, their substantial computational demands remain a major barrier to practical deployment in resource-constrained 6G networks. In this paper, we present a training-free framework for adaptive token merging in pretrained vision transformers to jointly reduce inference time and transmission resource usage. We formulate the selection of per-layer merging proportions as a multi-objective optimization problem to balance accuracy and computational cost. We employ Gaussian process-based Bayesian optimization to construct a Pareto frontier of optimal configurations, enabling flexible runtime adaptation to dynamic application requirements and channel conditions. Extensive experiments demonstrate that our method consistently outperforms other baselines and achieves significant reductions in floating-point operations while maintaining competitive accuracy across a wide range of signal-to-noise ratio (SNR) conditions. Additional results highlight the effectiveness of adaptive policies that adjust merging aggressiveness in response to channel quality, providing a practical mechanism to trade off latency and semantic fidelity on demand. These findings establish a scalable and efficient approach for deploying transformer-based semantic communication in future edge intelligence systems.

KVCompose: Efficient Structured KV Cache Compression with Composite Tokens

Large language models (LLMs) rely on key-value (KV) caches for efficient autoregressive decoding; however, cache size grows linearly with context length and model depth, becoming a major bottleneck in long-context inference. Prior KV cache compression methods either enforce rigid heuristics, disrupt tensor layouts with per-attention-head variability, or require specialized compute kernels. We propose a simple, yet effective, KV cache compression framework based on attention-guided, layer-adaptive composite tokens. Our method aggregates attention scores to estimate token importance, selects head-specific tokens independently, and aligns them into composite tokens that respect the uniform cache structure required by existing inference engines. A global allocation mechanism further adapts retention budgets across layers, assigning more capacity to layers with informative tokens. This approach achieves significant memory reduction while preserving accuracy, consistently outperforming prior structured and semi-structured methods. Crucially, our approach remains fully compatible with standard inference pipelines, offering a practical and scalable solution for efficient long-context LLM deployment.

Systematic Optimization of Open Source Large Language Models for Mathematical Reasoning

This paper presents a practical investigation into fine-tuning model parameters for mathematical reasoning tasks through experimenting with various configurations including randomness control, reasoning depth, and sampling strategies, careful tuning demonstrates substantial improvements in efficiency as well as performance. A holistically optimized framework is introduced for five state-of-the-art models on mathematical reasoning tasks, exhibiting significant performance boosts while maintaining solution correctness. Through systematic parameter optimization across Qwen2.5-72B, Llama-3.1-70B, DeepSeek-V3, Mixtral-8x22B, and Yi-Lightning, consistent efficiency gains are demonstrated with 100% optimization success rate. The methodology achieves an average 29.4% reduction in computational cost and 23.9% improvement in inference speed across all tested models. This framework systematically searches parameter spaces including temperature (0.1-0.5), reasoning steps (4-12), planning periods (1-4), and nucleus sampling (0.85-0.98), determining optimal configurations through testing on mathematical reasoning benchmarks. Critical findings show that lower temperature regimes (0.1-0.4) and reduced reasoning steps (4-6) consistently enhance efficiency without compromising accuracy. DeepSeek-V3 achieves the highest accuracy at 98%, while Mixtral-8x22B delivers the most cost-effective performance at 361.5 tokens per accurate response. Key contributions include: (1) the first comprehensive optimization study for five diverse SOTA models in mathematical reasoning, (2) a standardized production-oriented parameter optimization framework, (3) discovery of universal optimization trends applicable across model architectures, and (4) production-ready configurations with extensive performance characterization.

Set Block Decoding is a Language Model Inference Accelerator

Autoregressive next token prediction language models offer powerful capabilities but face significant challenges in practical deployment due to the high computational and memory costs of inference, particularly during the decoding stage. We introduce Set Block Decoding (SBD), a simple and flexible paradigm that accelerates generation by integrating standard next token prediction (NTP) and masked token prediction (MATP) within a single architecture. SBD allows the model to sample multiple, not necessarily consecutive, future tokens in parallel, a key distinction from previous acceleration methods. This flexibility allows the use of advanced solvers from the discrete diffusion literature, offering significant speedups without sacrificing accuracy. SBD requires no architectural changes or extra training hyperparameters, maintains compatibility with exact KV-caching, and can be implemented by fine-tuning existing next token prediction models. By fine-tuning Llama-3.1 8B and Qwen-3 8B, we demonstrate that SBD enables a 3-5x reduction in the number of forward passes required for generation while achieving same performance as equivalent NTP training.

ThinkDial: An Open Recipe for Controlling Reasoning Effort in Large Language Models

Large language models (LLMs) with chain-of-thought reasoning have demonstrated remarkable problem-solving capabilities, but controlling their computational effort remains a significant challenge for practical deployment. Recent proprietary systems like OpenAI's gpt-oss series have introduced discrete operational modes for intuitive reasoning control, but the open-source community has largely failed to achieve such capabilities. In this paper, we introduce ThinkDial, the first open-recipe end-to-end framework that successfully implements gpt-oss-style controllable reasoning through discrete operational modes. Our system enables seamless switching between three distinct reasoning regimes: High mode (full reasoning capability), Medium mode (50 percent token reduction with <10 percent performance degradation), and Low mode (75 percent token reduction with <15 percent performance degradation). We achieve this through an end-to-end training paradigm that integrates budget-mode control throughout the entire pipeline: budget-mode supervised fine-tuning that embeds controllable reasoning capabilities directly into the learning process, and two-phase budget-aware reinforcement learning with adaptive reward shaping. Extensive experiments demonstrate that ThinkDial achieves target compression-performance trade-offs with clear response length reductions while maintaining performance thresholds. The framework also exhibits strong generalization capabilities on out-of-distribution tasks.

SemToken: Semantic-Aware Tokenization for Efficient Long-Context Language Modeling

Tokenization plays a critical role in language modeling, yet existing approaches such as Byte-Pair Encoding (BPE) or WordPiece operate purely on frequency statistics, ignoring the underlying semantic structure of text. This leads to over-tokenization of semantically redundant spans and underutilization of contextual coherence, particularly in long-context scenarios. In this work, we propose \textbf{SemToken}, a semantic-aware tokenization framework that jointly reduces token redundancy and improves computation efficiency. SemToken first extracts contextual semantic embeddings via lightweight encoders and performs local semantic clustering to merge semantically equivalent tokens. Then, it allocates heterogeneous token granularity based on semantic density, allowing finer-grained tokenization in content-rich regions and coarser compression in repetitive or low-entropy spans. SemToken can be seamlessly integrated with modern language models and attention acceleration methods. Experiments on long-context language modeling benchmarks such as WikiText-103 and LongBench show that SemToken achieves up to $2.4\times$ reduction in token count and $1.9\times$ speedup, with negligible or no degradation in perplexity and downstream accuracy. Our findings suggest that semantic structure offers a promising new axis for optimizing tokenization and computation in large language models.

Language-Guided Temporal Token Pruning for Efficient VideoLLM Processing

Vision Language Models (VLMs) struggle with long-form videos due to the quadratic complexity of attention mechanisms. We propose Language-Guided Temporal Token Pruning (LGTTP), which leverages temporal cues from queries to adaptively prune video tokens, preserving contextual continuity while reducing computational overhead. Unlike uniform pruning or keyframe selection, LGTTP retains higher token density in temporally relevant segments. Our model-agnostic framework integrates with TimeChat and LLaVA-Video, achieving a 65% reduction in computation while preserving 97-99% of the original performance. On QVHighlights, LGTTP improves HIT@1 by +9.5%, and on Charades-STA, it retains 99.6% of R@1. It excels on queries with explicit temporal markers and remains effective across general video understanding tasks.

TreePO: Bridging the Gap of Policy Optimization and Efficacy and Inference Efficiency with Heuristic Tree-based Modeling

Recent advancements in aligning large language models via reinforcement learning have achieved remarkable gains in solving complex reasoning problems, but at the cost of expensive on-policy rollouts and limited exploration of diverse reasoning paths. In this work, we introduce TreePO, involving a self-guided rollout algorithm that views sequence generation as a tree-structured searching process. Composed of dynamic tree sampling policy and fixed-length segment decoding, TreePO leverages local uncertainty to warrant additional branches. By amortizing computation across common prefixes and pruning low-value paths early, TreePO essentially reduces the per-update compute burden while preserving or enhancing exploration diversity. Key contributions include: (1) a segment-wise sampling algorithm that alleviates the KV cache burden through contiguous segments and spawns new branches along with an early-stop mechanism; (2) a tree-based segment-level advantage estimation that considers both global and local proximal policy optimization. and (3) analysis on the effectiveness of probability and quality-driven dynamic divergence and fallback strategy. We empirically validate the performance gain of TreePO on a set reasoning benchmarks and the efficiency saving of GPU hours from 22\% up to 43\% of the sampling design for the trained models, meanwhile showing up to 40\% reduction at trajectory-level and 35\% at token-level sampling compute for the existing models. While offering a free lunch of inference efficiency, TreePO reveals a practical path toward scaling RL-based post-training with fewer samples and less compute. Home page locates at https://m-a-p.ai/TreePO.