QUOKA: Query-Oriented KV Selection For Efficient LLM Prefill

We present QUOKA: Query-oriented KV selection for efficient attention, a training-free and hardware agnostic sparse attention algorithm for accelerating transformer inference under chunked prefill. While many queries focus on a smaller group of keys in the attention operator, we observe that queries with low cosine similarity with respect to the mean query interact more strongly with more keys and have the greatest contribution to final attention logits. By prioritizing these low cosine similarity queries, the behavior of full attention during the prefill stage can be closely approximated. QUOKA leverages this observation, accelerating attention by (1) first retaining a small set of representative queries and (2) then subselectin the keys most aligned with those queries. Through experiments on Needle-In-A-Haystack, LongBench, RULER, and Math500, we show that, while realizing a 3x reduction in time-to-first-token, 5x speedup in attention on Nvidia GPUs and up to nearly a 7x speedup on Intel Xeon CPUs, QUOKA achieves near-baseline accuracy, utilizing 88% fewer key-value pairs per attention evaluation.

Fast Forward: Accelerating LLM Prefill with Predictive FFN Sparsity

The prefill stage of large language model (LLM) inference is a key computational bottleneck for long-context workloads. At short-to-moderate context lengths (1K--16K tokens), Feed-Forward Networks (FFNs) dominate this cost, accounting for most of the total FLOPs. Existing FFN sparsification methods, designed for autoregressive decoding, fail to exploit the prefill stage's parallelism and often degrade accuracy. To address this, we introduce FastForward, a predictive sparsity framework that accelerates LLM prefill through block-wise, context-aware FFN sparsity. FastForward combines (1) a lightweight expert predictor to select high-importance neurons per block, (2) an error compensation network to correct sparsity-induced errors, and (3) a layer-wise sparsity scheduler to allocate compute based on token-mixing importance. Across LLaMA and Qwen models up to 8B parameters, FastForward delivers up to 1.45$\times$ compute-bound speedup at 50% FFN sparsity with $<$ 6% accuracy loss compared to the dense baseline on LongBench, substantially reducing Time-to-First-Token (TTFT) for efficient, long-context LLM inference on constrained hardware.

A.X K1 Technical Report

We introduce A.X K1, a 519B-parameter Mixture-of-Experts (MoE) language model trained from scratch. Our design leverages scaling laws to optimize training configurations and vocabulary size under fixed computational budgets. A.X K1 is pre-trained on a corpus of approximately 10T tokens, curated by a multi-stage data processing pipeline. Designed to bridge the gap between reasoning capability and inference efficiency, A.X K1 supports explicitly controllable reasoning to facilitate scalable deployment across diverse real-world scenarios. We propose a simple yet effective Think-Fusion training recipe, enabling user-controlled switching between thinking and non-thinking modes within a single unified model. Extensive evaluations demonstrate that A.X K1 achieves performance competitive with leading open-source models, while establishing a distinctive advantage in Korean-language benchmarks.

KeyDiff: Key Similarity-Based KV Cache Eviction for Long-Context LLM Inference in Resource-Constrained Environments

In this work, we demonstrate that distinctive keys during LLM inference tend to have high attention scores. We explore this phenomenon and propose KeyDiff, a training-free KV cache eviction method based on key similarity. This method facilitates the deployment of LLM-based application requiring long input prompts in resource-constrained environments with limited memory and compute budgets. Unlike other KV cache eviction methods, KeyDiff can process arbitrarily long prompts within strict resource constraints and efficiently generate responses. We demonstrate that KeyDiff computes the optimal solution to a KV cache selection problem that maximizes key diversity, providing a theoretical understanding of KeyDiff. Notably,KeyDiff does not rely on attention scores, allowing the use of optimized attention mechanisms like FlashAttention. We demonstrate the effectiveness of KeyDiff across diverse tasks and models, illustrating a performance gap of less than 0.04\% with 8K cache budget ($\sim$ 23\% KV cache reduction) from the non-evicting baseline on the LongBench benchmark for Llama 3.1-8B and Llama 3.2-3B.

KeDiff: Key Similarity-Based KV Cache Eviction for Long-Context LLM Inference in Resource-Constrained Environments

In this work, we demonstrate that distinctive keys during LLM inference tend to have high attention scores. We explore this phenomenon and propose KeyDiff, a training-free KV cache eviction method based on key similarity. This method facilitates the deployment of LLM-based application requiring long input prompts in resource-constrained environments with limited memory and compute budgets. Unlike other KV cache eviction methods, KeyDiff can process arbitrarily long prompts within strict resource constraints and efficiently generate responses. We demonstrate that KeyDiff computes the optimal solution to a KV cache selection problem that maximizes key diversity, providing a theoretical understanding of KeyDiff. Notably,KeyDiff does not rely on attention scores, allowing the use of optimized attention mechanisms like FlashAttention. We demonstrate the effectiveness of KeyDiff across diverse tasks and models, illustrating a performance gap of less than 0.04\% with 8K cache budget ($\sim$ 23\% KV cache reduction) from the non-evicting baseline on the LongBench benchmark for Llama 3.1-8B and Llama 3.2-3B.

CAOTE: KV Caching through Attention Output Error based Token Eviction

While long context support of large language models has extended their abilities, it also incurs challenges in memory and compute which becomes crucial bottlenecks in resource-restricted devices. Token eviction, a widely adopted post-training methodology designed to alleviate the bottlenecks by evicting less important tokens from the cache, typically uses attention scores as proxy metrics for token importance. However, one major limitation of attention score as a token-wise importance metrics is that it lacks the information about contribution of tokens to the attention output. In this paper, we propose a simple eviction criterion based on the contribution of cached tokens to attention outputs. Our method, CAOTE, optimizes for eviction error due to token eviction, by seamlessly integrating attention scores and value vectors. This is the first method which uses value vector information on top of attention-based eviction scores. Additionally, CAOTE can act as a meta-heuristic method with flexible usage with any token eviction method. We show that CAOTE, when combined with the state-of-the-art attention score-based methods, always improves accuracies on the downstream task, indicating the importance of leveraging information from values during token eviction process.

Retrieval-Augmented Generation with Estimation of Source Reliability

Retrieval-augmented generation (RAG) addresses key limitations of large language models (LLMs), such as hallucinations and outdated knowledge, by incorporating external databases. These databases typically consult multiple sources to encompass up-to-date and various information. However, standard RAG methods often overlook the heterogeneous source reliability in the multi-source database and retrieve documents solely based on relevance, making them prone to propagating misinformation. To address this, we propose Reliability-Aware RAG (RA-RAG) which estimates the reliability of multiple sources and incorporates this information into both retrieval and aggregation processes. Specifically, it iteratively estimates source reliability and true answers for a set of queries with no labelling. Then, it selectively retrieves relevant documents from a few of reliable sources and aggregates them using weighted majority voting, where the selective retrieval ensures scalability while not compromising the performance. We also introduce a benchmark designed to reflect real-world scenarios with heterogeneous source reliability and demonstrate the effectiveness of RA-RAG compared to a set of baselines.

Expanding Search Space with Diverse Prompting Agents: An Efficient Sampling Approach for LLM Mathematical Reasoning

Large Language Models (LLMs) have exhibited remarkable capabilities in many complex tasks including mathematical reasoning. However, traditional approaches heavily rely on ensuring self-consistency within single prompting method, which limits the exploration of diverse problem-solving strategies. This study addresses these limitations by performing an experimental analysis of distinct prompting methods within the domain of mathematical reasoning. Our findings demonstrate that each method explores a distinct search space, and this differentiation becomes more evident with increasing problem complexity. To leverage this phenomenon, we applied efficient sampling process that uniformly combines samples from these diverse methods, which not only expands the maximum search space but achieves higher performance with fewer runs compared to single methods. Especially, within the subset of difficult questions of MATH dataset named MATH-hard, The maximum search space was achieved while utilizing approximately 43% fewer runs than single methods on average. These findings highlight the importance of integrating diverse problem-solving strategies to enhance the reasoning abilities of LLMs.

Towards Efficient Visual-Language Alignment of the Q-Former for Visual Reasoning Tasks

Recent advancements in large language models have demonstrated enhanced capabilities in visual reasoning tasks by employing additional encoders for aligning different modalities. While the Q-Former has been widely used as a general encoder for aligning several modalities including image, video, audio, and 3D with large language models, previous works on its efficient training and the analysis of its individual components have been limited. In this work, we investigate the effectiveness of parameter efficient fine-tuning (PEFT) the Q-Former using InstructBLIP with visual reasoning benchmarks ScienceQA and IconQA. We observe that applying PEFT to the Q-Former achieves comparable performance to full fine-tuning using under 2% of the trainable parameters. Additionally, we employ AdaLoRA for dynamic parameter budget reallocation to examine the relative importance of the Q-Former's sublayers with 4 different benchmarks. Our findings reveal that the self-attention layers are noticeably more important in perceptual visual-language reasoning tasks, and relative importance of FFN layers depends on the complexity of visual-language patterns involved in tasks. The code is available at https://github.com/AttentionX/InstructBLIP_PEFT.

PARCO: Learning Parallel Autoregressive Policies for Efficient Multi-Agent Combinatorial Optimization

Multi-agent combinatorial optimization problems such as routing and scheduling have great practical relevance but present challenges due to their NP-hard combinatorial nature, hard constraints on the number of possible agents, and hard-to-optimize objective functions. This paper introduces PARCO (Parallel AutoRegressive Combinatorial Optimization), a novel approach that learns fast surrogate solvers for multi-agent combinatorial problems with reinforcement learning by employing parallel autoregressive decoding. We propose a model with a Multiple Pointer Mechanism to efficiently decode multiple decisions simultaneously by different agents, enhanced by a Priority-based Conflict Handling scheme. Moreover, we design specialized Communication Layers that enable effective agent collaboration, thus enriching decision-making. We evaluate PARCO in representative multi-agent combinatorial problems in routing and scheduling and demonstrate that our learned solvers offer competitive results against both classical and neural baselines in terms of both solution quality and speed. We make our code openly available at https://github.com/ai4co/parco.