KVpop -- Key-Value Cache Compression with Predictive Online Pruning

Key-value (KV) cache growth is a major bottleneck in autoregressive decoding, as memory and bandwidth scale linearly with context length. Existing KV eviction methods often rely on static heuristics or proxy scores, which poorly track future token utility and cause brittle eviction as relevance shifts. To address this, we introduce KVpop, which learns a fixed-budget KV eviction policy by directly supervising the keep-or-drop decision. The scorer is trained against a novel future-attention target, computed efficiently without materializing dense attention maps. We further introduce a delayed memory-based scorer that, uniquely among learned eviction methods, defers scoring for a fixed number of steps to exploit near-future context. On AIME and HMMT mathematical reasoning, KVpop retains 98% of full-attention performance on Qwen3-4B at 75% KV cache compression and 97% at 88% compression, consistently outperforming established eviction baselines. Qwen3-8B shows even stronger results, reaching near-full teacher performance. These results show that supervising eviction with future-attention signals cuts memory costs while maintaining quality.

On Subquadratic Architectures: From Applications to Principles

Transformers dominate modern sequence modeling, but their quadratic attention incurs substantial computational cost. Subquadratic architectures offer a scalable alternative. However, it remains unclear which designs yield the most effective sequence models. We compare three leading approaches: xLSTM, Mamba-2, and Gated DeltaNet. We evaluate these models on tasks with complex dependencies: (1) code-model pre-training, (2) distillation of code models from large language models, and (3) pre-training of time-series foundation models. Across these settings, xLSTM delivers the strongest overall performance. To explain xLSTM's advantage, we present a unified formulation and analyze the underlying architectural mechanisms, focusing on state tracking and memory dynamics. Our results show that xLSTM enables more flexible and stable memory correction via its gating scheme. We corroborate these findings on controlled synthetic length-generalization tasks. Overall, our findings indicate that xLSTM's gains on complex tasks stem from robust state tracking and accumulation.

Effective Distillation to Hybrid xLSTM Architectures

There have been numerous attempts to distill quadratic attention-based large language models (LLMs) into sub-quadratic linearized architectures. However, despite extensive research, such distilled models often fail to match the performance of their teacher LLMs on various downstream tasks. We set out the goal of lossless distillation, which we define in terms of tolerance-corrected Win-and-Tie rates between student and teacher on sets of tasks. To this end, we introduce an effective distillation pipeline for xLSTM-based students. We propose an additional merging stage, where individually linearized experts are combined into a single model. We show the effectiveness of this pipeline by distilling base and instruction-tuned models from the Llama, Qwen, and Olmo families. In many settings, our xLSTM-based students recover most of the teacher's performance, and even exceed it on some downstream tasks. Our contributions are an important step towards more energy-efficient and cost-effective replacements for transformer-based LLMs.

Adaptive Retrieval helps Reasoning in LLMs -- but mostly if it's not used

Large Language Models (LLMs) often falter in complex reasoning tasks due to their static, parametric knowledge, leading to hallucinations and poor performance in specialized domains like mathematics. This work explores a fundamental principle for enhancing generative models: treating retrieval as a form of dynamic in-context learning. We test an adaptive retrieval-augmented architecture where an LLM agent actively decides when to query an external knowledge base during its reasoning process. We compare this adaptive strategy against a standard Chain-of-Thought (CoT) baseline and a static retrieval approach on the GSM8K and MATH-500 benchmarks. Although our experiments show that static retrieval is inferior to CoT, the adaptive retrieval shows interesting behavior: While traces including retrieved results show slightly worse performance compared to CoT, traces that do not include retrieval actually perform better compared to CoT. This suggests that: (a) retrieval only rarely helps reasoning (we show a few counterexamples, e.g. using useful theorems) and (b) actively not using retrieval is indicative of good model performance. Furthermore, we find that the model scales its retrieval frequency with the difficulty of the problem, reinforcing that the decision to retrieve is a crucial metacognitive signal. The agent's ability to self-assess its knowledge and selectively engage with external information represents a key principle for building more robust and reliable generative models.