LookaheadKV: Fast and Accurate KV Cache Eviction by Glimpsing into the Future without Generation

Transformer-based large language models (LLMs) rely on key-value (KV) caching to avoid redundant computation during autoregressive inference. While this mechanism greatly improves efficiency, the cache size grows linearly with the input sequence length, quickly becoming a bottleneck for long-context tasks. Existing solutions mitigate this problem by evicting prompt KV that are deemed unimportant, guided by estimated importance scores. Notably, a recent line of work proposes to improve eviction quality by "glimpsing into the future", in which a draft generator produces a surrogate future response approximating the target model's true response, and this surrogate is subsequently used to estimate the importance of cached KV more accurately. However, these approaches rely on computationally expensive draft generation, which introduces substantial prefilling overhead and limits their practicality in real-world deployment. To address this challenge, we propose LookaheadKV, a lightweight eviction framework that leverages the strength of surrogate future response without requiring explicit draft generation. LookaheadKV augments transformer layers with parameter-efficient modules trained to predict true importance scores with high accuracy. Our design ensures negligible runtime overhead comparable to existing inexpensive heuristics, while achieving accuracy superior to more costly approximation methods. Extensive experiments on long-context understanding benchmarks, across a wide range of models, demonstrate that our method not only outperforms recent competitive baselines in various long-context understanding tasks, but also reduces the eviction cost by up to 14.5x, leading to significantly faster time-to-first-token. Our code is available at https://github.com/SamsungLabs/LookaheadKV.

Fine-Tuning Without Forgetting In-Context Learning: A Theoretical Analysis of Linear Attention Models

Transformer-based large language models exhibit in-context learning, enabling adaptation to downstream tasks via few-shot prompting with demonstrations. In practice, such models are often fine-tuned to improve zero-shot performance on downstream tasks, allowing them to solve tasks without examples and thereby reducing inference costs. However, fine-tuning can degrade in-context learning, limiting the performance of fine-tuned models on tasks not seen during fine-tuning. Using linear attention models, we provide a theoretical analysis that characterizes how fine-tuning objectives modify attention parameters and identifies conditions under which this leads to degraded few-shot performance. We show that fine-tuning all attention parameters can harm in-context learning, whereas restricting updates to the value matrix improves zero-shot performance while preserving in-context learning. We further show that incorporating an auxiliary few-shot loss enhances in-context learning primarily on the target task, at the expense of degraded in-context learning ability on tasks not seen during fine-tuning. We empirically validate our theoretical results.

TAPE: Tool-Guided Adaptive Planning and Constrained Execution in Language Model Agents

Language Model (LM) agents have demonstrated remarkable capabilities in solving tasks that require multiple interactions with the environment. However, they remain vulnerable in environments where a single error often leads to irrecoverable failure, particularly under strict feasibility constraints. We systematically analyze existing agent frameworks, identifying imperfect planning and stochastic execution as the primary causes. To address these challenges, we propose Tool-guided Adaptive Planning with constrained Execution (TAPE). TAPE enhances planning capability by aggregating multiple plans into a graph and employing an external solver to identify a feasible path. During execution, TAPE employs constrained decoding to reduce sampling noise, while adaptively re-planning whenever environmental feedback deviates from the intended state. Experiments across Sokoban, ALFWorld, MuSiQue, and GSM8K-Hard demonstrate that TAPE consistently outperforms existing frameworks, with particularly large gains on hard settings, improving success rates by 21.0 percentage points on hard settings on average, and by 20.0 percentage points for weaker base models on average. Code and data available at here.

TABED: Test-Time Adaptive Ensemble Drafting for Robust Speculative Decoding in LVLMs

Speculative decoding (SD) has proven effective for accelerating LLM inference by quickly generating draft tokens and verifying them in parallel. However, SD remains largely unexplored for Large Vision-Language Models (LVLMs), which extend LLMs to process both image and text prompts. To address this gap, we benchmark existing inference methods with small draft models on 11 datasets across diverse input scenarios and observe scenario-specific performance fluctuations. Motivated by these findings, we propose Test-time Adaptive Batched Ensemble Drafting (TABED), which dynamically ensembles multiple drafts obtained via batch inference by leveraging deviations from past ground truths available in the SD setting. The dynamic ensemble method achieves an average robust walltime speedup of 1.74x over autoregressive decoding and a 5% improvement over single drafting methods, while remaining training-free and keeping ensembling costs negligible through parameter sharing. With its plug-and-play compatibility, we further enhance TABED by integrating advanced verification and alternative drafting methods. Code and custom-trained models are available at https://github.com/furiosa-ai/TABED.

ParallelBench: Understanding the Trade-offs of Parallel Decoding in Diffusion LLMs

While most autoregressive LLMs are constrained to one-by-one decoding, diffusion LLMs (dLLMs) have attracted growing interest for their potential to dramatically accelerate inference through parallel decoding. Despite this promise, the conditional independence assumption in dLLMs causes parallel decoding to ignore token dependencies, inevitably degrading generation quality when these dependencies are strong. However, existing works largely overlook these inherent challenges, and evaluations on standard benchmarks (e.g., math and coding) are not sufficient to capture the quality degradation caused by parallel decoding. To address this gap, we first provide an information-theoretic analysis of parallel decoding. We then conduct case studies on analytically tractable synthetic list operations from both data distribution and decoding strategy perspectives, offering quantitative insights that highlight the fundamental limitations of parallel decoding. Building on these insights, we propose ParallelBench, the first benchmark specifically designed for dLLMs, featuring realistic tasks that are trivial for humans and autoregressive LLMs yet exceptionally challenging for dLLMs under parallel decoding. Using ParallelBench, we systematically analyze both dLLMs and autoregressive LLMs, revealing that: (i) dLLMs under parallel decoding can suffer dramatic quality degradation in real-world scenarios, and (ii) current parallel decoding strategies struggle to adapt their degree of parallelism based on task difficulty, thus failing to achieve meaningful speedup without compromising quality. Our findings underscore the pressing need for innovative decoding methods that can overcome the current speed-quality trade-off. We release our benchmark to help accelerate the development of truly efficient dLLMs.

Draft-based Approximate Inference for LLMs

Optimizing inference for long-context Large Language Models (LLMs) is increasingly important due to the quadratic compute and linear memory complexity of Transformers. Existing approximation methods, such as key-value (KV) cache dropping, sparse attention, and prompt compression, typically rely on rough predictions of token or KV pair importance. We propose a novel framework for approximate LLM inference that leverages small draft models to more accurately predict the importance of tokens and KV pairs. Specifically, we introduce two instantiations of our proposed framework: (i) SpecKV, which leverages a draft output to accurately assess the importance of each KV pair for more effective KV cache dropping, and (ii) SpecPC, which uses the draft model's attention activations to identify and discard unimportant prompt tokens. To the best of our knowledge, this is the first work to use draft models for approximate LLM inference acceleration, extending their utility beyond traditional lossless speculative decoding. We motivate our methods with theoretical and empirical analyses, and show a strong correlation between the attention patterns of draft and target models. Extensive experiments on long-context benchmarks show that our methods consistently achieve higher accuracy than existing baselines, while preserving the same improvements in memory usage, latency, and throughput. Our code is available at https://github.com/furiosa-ai/draft-based-approx-llm.

In-Context Learning with Hypothesis-Class Guidance

Recent research has investigated the underlying mechanisms of in-context learning (ICL) both theoretically and empirically, often using data generated from simple function classes. However, the existing work often focuses on the sequence consisting solely of labeled examples, while in practice, labeled examples are typically accompanied by an instruction, providing some side information about the task. In this work, we propose ICL with hypothesis-class guidance (ICL-HCG), a novel synthetic data model for ICL where the input context consists of the literal description of a (finite) hypothesis class H and $(x,y)$ pairs from a hypothesis chosen from H. Under our framework ICL-HCG, we conduct extensive experiments to explore: (i) a variety of generalization abilities to new hypothesis classes; (ii) different model architectures; (iii) sample complexity; (iv) in-context data imbalance; (v) the role of instruction; and (vi) the effect of pretraining hypothesis diversity. As a result, we show that (a) Transformers can successfully learn ICL-HCG and generalize to unseen hypotheses and unseen hypothesis classes, and (b) compared with ICL without instruction, ICL-HCG achieves significantly higher accuracy, demonstrating the role of instructions.

VersaPRM: Multi-Domain Process Reward Model via Synthetic Reasoning Data

Process Reward Models (PRMs) have proven effective at enhancing mathematical reasoning for Large Language Models (LLMs) by leveraging increased inference-time computation. However, they are predominantly trained on mathematical data and their generalizability to non-mathematical domains has not been rigorously studied. In response, this work first shows that current PRMs have poor performance in other domains. To address this limitation, we introduce VersaPRM, a multi-domain PRM trained on synthetic reasoning data generated using our novel data generation and annotation method. VersaPRM achieves consistent performance gains across diverse domains. For instance, in the MMLU-Pro category of Law, VersaPRM via weighted majority voting, achieves a 7.9% performance gain over the majority voting baseline -- surpassing Qwen2.5-Math-PRM's gain of 1.3%. We further contribute to the community by open-sourcing all data, code and models for VersaPRM.

Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challenges

Large language models often struggle with length generalization and solving complex problem instances beyond their training distribution. We present a self-improvement approach where models iteratively generate and learn from their own solutions, progressively tackling harder problems while maintaining a standard transformer architecture. Across diverse tasks including arithmetic, string manipulation, and maze solving, self-improving enables models to solve problems far beyond their initial training distribution-for instance, generalizing from 10-digit to 100-digit addition without apparent saturation. We observe that in some cases filtering for correct self-generated examples leads to exponential improvements in out-of-distribution performance across training rounds. Additionally, starting from pretrained models significantly accelerates this self-improvement process for several tasks. Our results demonstrate how controlled weak-to-strong curricula can systematically teach a model logical extrapolation without any changes to the positional embeddings, or the model architecture.

Task Vectors in In-Context Learning: Emergence, Formation, and Benefit

In-context learning is a remarkable capability of transformers, referring to their ability to adapt to specific tasks based on a short history or context. Previous research has found that task-specific information is locally encoded within models, though their emergence and functionality remain unclear due to opaque pre-training processes. In this work, we investigate the formation of task vectors in a controlled setting, using models trained from scratch on synthetic datasets. Our findings confirm that task vectors naturally emerge under certain conditions, but the tasks may be relatively weakly and/or non-locally encoded within the model. To promote strong task vectors encoded at a prescribed location within the model, we propose an auxiliary training mechanism based on a task vector prompting loss (TVP-loss). This method eliminates the need to search for task-correlated encodings within the trained model and demonstrably improves robustness and generalization.