Fast-Slow Efficient Training for Multimodal Large Language Models via Visual Token Pruning

Multimodal Large Language Models (MLLMs) suffer from severe training inefficiency issue, which is associated with their massive model sizes and visual token numbers. Existing efforts in efficient training focus on reducing model sizes or trainable parameters. Inspired by the success of Visual Token Pruning (VTP) in improving inference efficiency, we are exploring another substantial research direction for efficient training by reducing visual tokens. However, applying VTP at the training stage results in a training-inference mismatch: pruning-trained models perform poorly when inferring on non-pruned full visual token sequences. To close this gap, we propose DualSpeed, a fast-slow framework for efficient training of MLLMs. The fast-mode is the primary mode, which incorporates existing VTP methods as plugins to reduce visual tokens, along with a mode isolator to isolate the model's behaviors. The slow-mode is the auxiliary mode, where the model is trained on full visual sequences to retain training-inference consistency. To boost its training, it further leverages self-distillation to learn from the sufficiently trained fast-mode. Together, DualSpeed can achieve both training efficiency and non-degraded performance. Experiments show DualSpeed accelerates the training of LLaVA-1.5 by 2.1$\times$ and LLaVA-NeXT by 4.0$\times$, retaining over 99% performance. Code: https://github.com/dingkun-zhang/DualSpeed

Conformal Thinking: Risk Control for Reasoning on a Compute Budget

Reasoning Large Language Models (LLMs) enable test-time scaling, with dataset-level accuracy improving as the token budget increases, motivating adaptive reasoning -- spending tokens when they improve reliability and stopping early when additional computation is unlikely to help. However, setting the token budget, as well as the threshold for adaptive reasoning, is a practical challenge that entails a fundamental risk-accuracy trade-off. We re-frame the budget setting problem as risk control, limiting the error rate while minimizing compute. Our framework introduces an upper threshold that stops reasoning when the model is confident (risking incorrect output) and a novel parametric lower threshold that preemptively stops unsolvable instances (risking premature stoppage). Given a target risk and a validation set, we use distribution-free risk control to optimally specify these stopping mechanisms. For scenarios with multiple budget controlling criteria, we incorporate an efficiency loss to select the most computationally efficient exiting mechanism. Empirical results across diverse reasoning tasks and models demonstrate the effectiveness of our risk control approach, demonstrating computational efficiency gains from the lower threshold and ensemble stopping mechanisms while adhering to the user-specified risk target.

Context Compression via Explicit Information Transmission

Long-context inference with Large Language Models (LLMs) is costly due to quadratic attention and growing key-value caches, motivating context compression. In this work, we study soft context compression, where a long context is condensed into a small set of continuous representations. Existing methods typically re-purpose the LLM itself as a trainable compressor, relying on layer-by-layer self-attention to iteratively aggregate information. We argue that this paradigm suffers from two structural limitations: (i) progressive representation overwriting across layers (ii) uncoordinated allocation of compression capacity across tokens. We propose ComprExIT (Context Compression via Explicit Information Transmission), a lightweight framework that formulates soft compression into a new paradigm: explicit information transmission over frozen LLM hidden states. This decouples compression from the model's internal self-attention dynamics. ComprExIT performs (i) depth-wise transmission to selectively transmit multi-layer information into token anchors, mitigating progressive overwriting, and (ii) width-wise transmission to aggregate anchors into a small number of slots via a globally optimized transmission plan, ensuring coordinated allocation of information. Across six question-answering benchmarks, ComprExIT consistently outperforms state-of-the-art context compression methods while introducing only ~1% additional parameters, demonstrating that explicit and coordinated information transmission enables more effective and robust long-context compression.

UniGeM: Unifying Data Mixing and Selection via Geometric Exploration and Mining

The scaling of Large Language Models (LLMs) is increasingly limited by data quality. Most methods handle data mixing and sample selection separately, which can break the structure in code corpora. We introduce \textbf{UniGeM}, a framework that unifies mixing and selection by treating data curation as a \textit{manifold approximation} problem without training proxy models or relying on external reference datasets. UniGeM operates hierarchically: \textbf{Macro-Exploration} learns mixing weights with stability-based clustering; \textbf{Micro-Mining} filters high-quality instances by their geometric distribution to ensure logical consistency. Validated by training 8B and 16B MoE models on 100B tokens, UniGeM achieves \textbf{2.0$\times$ data efficiency} over a random baseline and further improves overall performance compared to SOTA methods in reasoning-heavy evaluations and multilingual generalization.

Efficient Variance-reduced Estimation from Generative EHR Models: The SCOPE and REACH Estimators

Generative models trained using self-supervision of tokenized electronic health record (EHR) timelines show promise for clinical outcome prediction. This is typically done using Monte Carlo simulation for future patient trajectories. However, existing approaches suffer from three key limitations: sparse estimate distributions that poorly differentiate patient risk levels, extreme computational costs, and high sampling variance. We propose two new estimators: the Sum of Conditional Outcome Probability Estimator (SCOPE) and Risk Estimation from Anticipated Conditional Hazards (REACH), that leverage next-token probability distributions discarded by standard Monte Carlo. We prove both estimators are unbiased and that REACH guarantees variance reduction over Monte Carlo sampling for any model and outcome. Empirically, on hospital mortality prediction in MIMIC-IV using the ETHOS-ARES framework, SCOPE and REACH match 100-sample Monte Carlo performance using only 10-11 samples (95% CI: [9,11]), representing a ~10x reduction in inference cost without degrading calibration. For ICU admission prediction, efficiency gains are more modest (~1.2x), which we attribute to the outcome's lower "spontaneity," a property we characterize theoretically and empirically. These methods substantially improve the feasibility of deploying generative EHR models in resource-constrained clinical settings.

Beyond Tokens: Semantic-Aware Speculative Decoding for Efficient Inference by Probing Internal States

Large Language Models (LLMs) achieve strong performance across many tasks but suffer from high inference latency due to autoregressive decoding. The issue is exacerbated in Large Reasoning Models (LRMs), which generate lengthy chains of thought. While speculative decoding accelerates inference by drafting and verifying multiple tokens in parallel, existing methods operate at the token level and ignore semantic equivalence (i.e., different token sequences expressing the same meaning), leading to inefficient rejections. We propose SemanticSpec, a semantic-aware speculative decoding framework that verifies entire semantic sequences instead of tokens. SemanticSpec introduces a semantic probability estimation mechanism that probes the model's internal hidden states to assess the likelihood of generating sequences with specific meanings.Experiments on four benchmarks show that SemanticSpec achieves up to 2.7x speedup on DeepSeekR1-32B and 2.1x on QwQ-32B, consistently outperforming token-level and sequence-level baselines in both efficiency and effectiveness.

Agent Primitives: Reusable Latent Building Blocks for Multi-Agent Systems

While existing multi-agent systems (MAS) can handle complex problems by enabling collaboration among multiple agents, they are often highly task-specific, relying on manually crafted agent roles and interaction prompts, which leads to increased architectural complexity and limited reusability across tasks. Moreover, most MAS communicate primarily through natural language, making them vulnerable to error accumulation and instability in long-context, multi-stage interactions within internal agent histories. In this work, we propose \textbf{Agent Primitives}, a set of reusable latent building blocks for LLM-based MAS. Inspired by neural network design, where complex models are built from reusable components, we observe that many existing MAS architectures can be decomposed into a small number of recurring internal computation patterns. Based on this observation, we instantiate three primitives: Review, Voting and Selection, and Planning and Execution. All primitives communicate internally via key-value (KV) cache, which improves both robustness and efficiency by mitigating information degradation across multi-stage interactions. To enable automatic system construction, an Organizer agent selects and composes primitives for each query, guided by a lightweight knowledge pool of previously successful configurations, forming a primitive-based MAS. Experiments show that primitives-based MAS improve average accuracy by 12.0-16.5\% over single-agent baselines, reduce token usage and inference latency by approximately 3$\times$-4$\times$ compared to text-based MAS, while incurring only 1.3$\times$-1.6$\times$ overhead relative to single-agent inference and providing more stable performance across model backbones.

Bringing Reasoning to Generative Recommendation Through the Lens of Cascaded Ranking

Generative Recommendation (GR) has become a promising end-to-end approach with high FLOPS utilization for resource-efficient recommendation. Despite the effectiveness, we show that current GR models suffer from a critical \textbf{bias amplification} issue, where token-level bias escalates as token generation progresses, ultimately limiting the recommendation diversity and hurting the user experience. By comparing against the key factor behind the success of traditional multi-stage pipelines, we reveal two limitations in GR that can amplify the bias: homogeneous reliance on the encoded history, and fixed computational budgets that prevent deeper user preference understanding. To combat the bias amplification issue, it is crucial for GR to 1) incorporate more heterogeneous information, and 2) allocate greater computational resources at each token generation step. To this end, we propose CARE, a simple yet effective cascaded reasoning framework for debiased GR. To incorporate heterogeneous information, we introduce a progressive history encoding mechanism, which progressively incorporates increasingly fine-grained history information as the generation process advances. To allocate more computations, we propose a query-anchored reasoning mechanism, which seeks to perform a deeper understanding of historical information through parallel reasoning steps. We instantiate CARE on three GR backbones. Empirical results on four datasets show the superiority of CARE in recommendation accuracy, diversity, efficiency, and promising scalability. The codes and datasets are available at https://github.com/Linxyhaha/CARE.

TodyComm: Task-Oriented Dynamic Communication for Multi-Round LLM-based Multi-Agent System

Multi-round LLM-based multi-agent systems rely on effective communication structures to support collaboration across rounds. However, most existing methods employ a fixed communication topology during inference, which falls short in many realistic applications where the agents' roles may change \textit{across rounds} due to dynamic adversary, task progression, or time-varying constraints such as communication bandwidth. In this paper, we propose addressing this issue through TodyComm, a \textbf{t}ask-\textbf{o}riented \textbf{dy}namic \textbf{comm}unication algorithm. It produces behavior-driven collaboration topologies that adapt to the dynamics at each round, optimizing the utility for the task through policy gradient. Experiments on five benchmarks demonstrate that under both dynamic adversary and communications budgets, TodyComm delivers superior task effectiveness while retaining token efficiency and scalability.

Instruction Anchors: Dissecting the Causal Dynamics of Modality Arbitration

Modality following serves as the capacity of multimodal large language models (MLLMs) to selectively utilize multimodal contexts based on user instructions. It is fundamental to ensuring safety and reliability in real-world deployments. However, the underlying mechanisms governing this decision-making process remain poorly understood. In this paper, we investigate its working mechanism through an information flow lens. Our findings reveal that instruction tokens function as structural anchors for modality arbitration: Shallow attention layers perform non-selective information transfer, routing multimodal cues to these anchors as a latent buffer; Modality competition is resolved within deep attention layers guided by the instruction intent, while MLP layers exhibit semantic inertia, acting as an adversarial force. Furthermore, we identify a sparse set of specialized attention heads that drive this arbitration. Causal interventions demonstrate that manipulating a mere $5\%$ of these critical heads can decrease the modality-following ratio by $60\%$ through blocking, or increase it by $60\%$ through targeted amplification of failed samples. Our work provides a substantial step toward model transparency and offers a principled framework for the orchestration of multimodal information in MLLMs.