Long context compression is a critical research problem due to its significance in reducing the high computational and memory costs associated with LLMs. In this paper, we propose Activation Beacon, a plug-in module for transformer-based LLMs that targets effective, efficient, and flexible compression of long contexts. To achieve this, our method introduces the following technical designs. 1) We directly compress the activations (i.e. keys and values at every layer), rather than leveraging soft prompts to relay information (which constitute a major bottleneck to encapsulate the complex information within long contexts). 2) We tailor the compression workflow, where each fine-grained input unit is progressively compressed, enabling high-quality compression and efficient computation during both training and inference. 3) We train the model through compression-based auto-regression, making full use of plain texts and instructional data to optimize the model's compression performance. 4) During training, we randomly sample a compression ratio at each step, teaching the model to support a wide range of compression configurations. Extensive evaluations are conducted on various long-context tasks whose lengths (e.g., 128K) may far exceed the maximum training length (20K), such as document understanding, few-shot learning, and Needle-in-a-Haystack. Whilst existing methods struggle to handle these challenging tasks, Activation Beacon maintains a comparable performance to the uncompressed baseline across various scenarios, achieving a 2x acceleration in inference time and an 8x reduction of memory costs for KV cache.

Currently, short video platforms have become the primary place for individuals to share experiences and obtain information. To better meet users' needs for acquiring information while browsing short videos, some apps have introduced a search entry at the bottom of videos, accompanied with recommended relevant queries. This scenario is known as query recommendation in video-related search, where core task is item-to-query (I2Q) recommendation. As this scenario has only emerged in recent years, there is a notable scarcity of academic research and publicly available datasets in this domain. To address this gap, we systematically examine the challenges associated with this scenario for the first time. Subsequently, we release a large-scale dataset derived from real-world data pertaining to the query recommendation in video- related search on the Kuaishou app (KuaiRS). Presently, existing methods rely on embeddings to calculate similarity for matching short videos with queries, lacking deep interaction between the semantic content and the query. In this paper, we introduce a novel LLM-based framework named GREAT, which guides que ry g ener ation with a trie to address I2Q recommendation in related search. Specifically, we initially gather high-quality queries with high exposure and click-through rate to construct a query-based trie. During training, we enhance the LLM's capability to generate high-quality queries using the query-based trie. In the inference phase, the query-based trie serves as a guide for the token generation. Finally, we further refine the relevance and literal quality between items and queries via a post-processing module. Extensive offline and online experiments demonstrate the effectiveness of our proposed method.

The existing Retrieval-Augmented Generation (RAG) systems face significant challenges in terms of cost and effectiveness. On one hand, they need to encode the lengthy retrieved contexts before responding to the input tasks, which imposes substantial computational overhead. On the other hand, directly using generic Large Language Models (LLMs) often leads to sub-optimal answers, while task-specific fine-tuning may compromise the LLMs' general capabilities. To address these challenges, we introduce a novel approach called FlexRAG (Flexible Context Adaptation for RAG). In this approach, the retrieved contexts are compressed into compact embeddings before being encoded by the LLMs. Simultaneously, these compressed embeddings are optimized to enhance downstream RAG performance. A key feature of FlexRAG is its flexibility, which enables effective support for diverse compression ratios and selective preservation of important contexts. With these designs, FlexRAG achieves superior generation quality while significantly reducing running costs. The experiments across multiple QA datasets validate our approach as a cost-effective and flexible solution for RAG systems (codebase: https://github.com/wcyno23/FlexRAG).

With the popularity of online music services, personalized music recommendation has garnered much research interest. Recommendation models are typically trained on datasets constructed from user feedback, which includes both the active feedback (e.g., clicking the Like or Skip buttons) and passive feedback (e.g., auto-play), with passive feedback comprising the majority. Due to the unavailability of user attention, the massive amount of passive feedback is unreliable, significantly compromising the quality of the training data. How to estimate the user's attention on the target music has become a critical problem in music recommendation. Heuristic methods such as exponential decay and negative sampling have been proposed. However, they either neglect the sequential dependencies between feedback actions or utilize only a small fraction of passive samples, leading to inaccurate and biased attention estimation. In this paper, we naturally propose modeling user attention prediction as a positive-unlabeled (PU) learning problem, where active feedback is treated as positive samples and passive feedback is treated as unlabeled samples, as we can only ensure that the user's attention is focused when she provides active feedback. Then we propose an extended PU-learning model with sequential dependencies, called UAE, which contains an unbiased user attention estimator and an unbiased propensity estimator. Subsequently, a joint learning algorithm is developed in which the attention and propensity estimators are optimized in alternating fashion. Theoretical analysis shows the unbiasedness and variance of the attention estimator and the propensity estimator. Extensive experiments on two large-scale datasets demonstrate the proposed UAE's effectiveness and generality in enhancing downstream music recommendation. One week online A/B testing on Huawei Music App manifests that UAE can significantly increase the users' play count and time over 2%, further demonstrating the effectiveness of UAE in real-world music recommendation products.

Contextual batched bandit (CBB) is a setting where a batch of rewards is observed from the environment at the end of each episode, but the rewards of the non-executed actions are unobserved, resulting in partial-information feedback. Existing approaches for CBB often ignore the rewards of the non-executed actions, leading to underutilization of feedback information. In this paper, we propose an efficient approach called Sketched Policy Updating with Imputed Rewards (SPUIR) that completes the unobserved rewards using sketching, which approximates the full-information feedbacks. We formulate reward imputation as an imputation regularized ridge regression problem that captures the feedback mechanisms of both executed and non-executed actions. To reduce time complexity, we solve the regression problem using randomized sketching. We prove that our approach achieves an instantaneous regret with controllable bias and smaller variance than approaches without reward imputation. Furthermore, our approach enjoys a sublinear regret bound against the optimal policy. We also present two extensions, a rate-scheduled version and a version for nonlinear rewards, making our approach more practical. Experimental results show that SPUIR outperforms state-of-the-art baselines on synthetic, public benchmark, and real-world datasets.

The debut of ChatGPT has recently attracted significant attention from the natural language processing (NLP) community and beyond. Existing studies have demonstrated that ChatGPT shows significant improvement in a range of downstream NLP tasks, but the capabilities and limitations of ChatGPT in terms of recommendations remain unclear. In this study, we aim to enhance ChatGPT’s recommendation capabilities by aligning it with traditional information retrieval (IR) ranking capabilities, including point-wise, pair-wise, and list-wise ranking. To achieve this goal, we re-formulate the aforementioned three recommendation policies into prompt formats tailored specifically to the domain at hand. Through extensive experiments on four datasets from different domains, we analyze the distinctions among the three recommendation policies. Our findings indicate that ChatGPT achieves an optimal balance between cost and performance when equipped with list-wise ranking. This research sheds light on a promising direction for aligning ChatGPT with recommendation tasks. To facilitate further explorations in this area, the full code and detailed original results are open-sourced at https://github.com/rainym00d/LLM4RS.