Inner Thinking Transformer: Leveraging Dynamic Depth Scaling to Foster Adaptive Internal Thinking

Large language models (LLMs) face inherent performance bottlenecks under parameter constraints, particularly in processing critical tokens that demand complex reasoning. Empirical analysis reveals challenging tokens induce abrupt gradient spikes across layers, exposing architectural stress points in standard Transformers. Building on this insight, we propose Inner Thinking Transformer (ITT), which reimagines layer computations as implicit thinking steps. ITT dynamically allocates computation through Adaptive Token Routing, iteratively refines representations via Residual Thinking Connections, and distinguishes reasoning phases using Thinking Step Encoding. ITT enables deeper processing of critical tokens without parameter expansion. Evaluations across 162M-466M parameter models show ITT achieves 96.5\% performance of a 466M Transformer using only 162M parameters, reduces training data by 43.2\%, and outperforms Transformer/Loop variants in 11 benchmarks. By enabling elastic computation allocation during inference, ITT balances performance and efficiency through architecture-aware optimization of implicit thinking pathways.

FARM: Frequency-Aware Model for Cross-Domain Live-Streaming Recommendation

Live-streaming services have attracted widespread popularity due to their real-time interactivity and entertainment value. Users can engage with live-streaming authors by participating in live chats, posting likes, or sending virtual gifts to convey their preferences and support. However, the live-streaming services faces serious data-sparsity problem, which can be attributed to the following two points: (1) User's valuable behaviors are usually sparse, e.g., like, comment and gift, which are easily overlooked by the model, making it difficult to describe user's personalized preference. (2) The main exposure content on our platform is short-video, which is 9 times higher than the exposed live-streaming, leading to the inability of live-streaming content to fully model user preference. To this end, we propose a Frequency-Aware Model for Cross-Domain Live-Streaming Recommendation, termed as FARM. Specifically, we first present the intra-domain frequency aware module to enable our model to perceive user's sparse yet valuable behaviors, i.e., high-frequency information, supported by the Discrete Fourier Transform (DFT). To transfer user preference across the short-video and live-streaming domains, we propose a novel preference align before fuse strategy, which consists of two parts: the cross-domain preference align module to align user preference in both domains with contrastive learning, and the cross-domain preference fuse module to further fuse user preference in both domains using a serious of tailor-designed attention mechanisms. Extensive offline experiments and online A/B testing on Kuaishou live-streaming services demonstrate the effectiveness and superiority of FARM. Our FARM has been deployed in online live-streaming services and currently serves hundreds of millions of users on Kuaishou.

Exploring Preference-Guided Diffusion Model for Cross-Domain Recommendation

Cross-domain recommendation (CDR) has been proven as a promising way to alleviate the cold-start issue, in which the most critical problem is how to draw an informative user representation in the target domain via the transfer of user preference existing in the source domain. Prior efforts mostly follow the embedding-and-mapping paradigm, which first integrate the preference into user representation in the source domain, and then perform a mapping function on this representation to the target domain. However, they focus on mapping features across domains, neglecting to explicitly model the preference integration process, which may lead to learning coarse user representation. Diffusion models (DMs), which contribute to more accurate user/item representations due to their explicit information injection capability, have achieved promising performance in recommendation systems. Nevertheless, these DMs-based methods cannot directly account for valuable user preference in other domains, leading to challenges in adapting to the transfer of preference for cold-start users. Consequently, the feasibility of DMs for CDR remains underexplored. To this end, we explore to utilize the explicit information injection capability of DMs for user preference integration and propose a Preference-Guided Diffusion Model for CDR to cold-start users, termed as DMCDR. Specifically, we leverage a preference encoder to establish the preference guidance signal with the user's interaction history in the source domain. Then, we explicitly inject the preference guidance signal into the user representation step by step to guide the reverse process, and ultimately generate the personalized user representation in the target domain, thus achieving the transfer of user preference across domains. Furthermore, we comprehensively explore the impact of six DMs-based variants on CDR.

Mixture of Hidden-Dimensions Transformer

Transformer models encounter challenges in scaling hidden dimensions efficiently, as uniformly increasing them inflates computational and memory costs while failing to emphasize the most relevant features for each token. For further understanding, we study hidden dimension sparsity and observe that trained Transformers utilize only a small fraction of token dimensions, revealing an "activation flow" pattern. Notably, there are shared sub-dimensions with sustained activation across multiple consecutive tokens and specialized sub-dimensions uniquely activated for each token. To better model token-relevant sub-dimensions, we propose MoHD (Mixture of Hidden Dimensions), a sparse conditional activation architecture. Particularly, MoHD employs shared sub-dimensions for common token features and a routing mechanism to dynamically activate specialized sub-dimensions. To mitigate potential information loss from sparsity, we design activation scaling and group fusion mechanisms to preserve activation flow. In this way, MoHD expands hidden dimensions with negligible increases in computation or parameters, efficient training and inference while maintaining performance. Evaluations across 10 NLP tasks show that MoHD surpasses Vanilla Transformers in parameter efficiency and task performance. It achieves 1.7% higher performance with 50% fewer activation parameters and 3.7% higher performance with a 3x parameter expansion at constant activation cost. MOHD offers a new perspective for scaling the model, showcasing the potential of hidden dimension sparsity to boost efficiency

MoR: Mixture of Ranks for Low-Rank Adaptation Tuning

Low-Rank Adaptation (LoRA) drives research to align its performance with full fine-tuning. However, significant challenges remain: (1) Simply increasing the rank size of LoRA does not effectively capture high-rank information, which leads to a performance bottleneck.(2) MoE-style LoRA methods substantially increase parameters and inference latency, contradicting the goals of efficient fine-tuning and ease of application. To address these challenges, we introduce Mixture of Ranks (MoR), which learns rank-specific information for different tasks based on input and efficiently integrates multi-rank information. We firstly propose a new framework that equates the integration of multiple LoRAs to expanding the rank of LoRA. Moreover, we hypothesize that low-rank LoRA already captures sufficient intrinsic information, and MoR can derive high-rank information through mathematical transformations of the low-rank components. Thus, MoR can reduces the learning difficulty of LoRA and enhances its multi-task capabilities. MoR achieves impressive results, with MoR delivering a 1.31\% performance improvement while using only 93.93\% of the parameters compared to baseline methods.

Revealing the Challenge of Detecting Character Knowledge Errors in LLM Role-Playing

Large language model (LLM) role-playing has gained widespread attention, where the authentic character knowledge is crucial for constructing realistic LLM role-playing agents. However, existing works usually overlook the exploration of LLMs' ability to detect characters' known knowledge errors (KKE) and unknown knowledge errors (UKE) while playing roles, which would lead to low-quality automatic construction of character trainable corpus. In this paper, we propose a probing dataset to evaluate LLMs' ability to detect errors in KKE and UKE. The results indicate that even the latest LLMs struggle to effectively detect these two types of errors, especially when it comes to familiar knowledge. We experimented with various reasoning strategies and propose an agent-based reasoning method, Self-Recollection and Self-Doubt (S2RD), to further explore the potential for improving error detection capabilities. Experiments show that our method effectively improves the LLMs' ability to detect error character knowledge, but it remains an issue that requires ongoing attention.

NACL: A General and Effective KV Cache Eviction Framework for LLMs at Inference Time

Large Language Models (LLMs) have ignited an innovative surge of AI applications, marking a new era of exciting possibilities equipped with extended context windows. However, hosting these models is cost-prohibitive mainly due to the extensive memory consumption of KV Cache involving long-context modeling. Despite several works proposing to evict unnecessary tokens from the KV Cache, most of them rely on the biased local statistics of accumulated attention scores and report performance using unconvincing metric like perplexity on inadequate short-text evaluation. In this paper, we propose NACL, a general framework for long-context KV cache eviction that achieves more optimal and efficient eviction in a single operation during the encoding phase. Due to NACL's efficiency, we combine more accurate attention score statistics in PROXY TOKENS EVICTION with the diversified random eviction strategy of RANDOM EVICTION, aiming to alleviate the issue of attention bias and enhance the robustness in maintaining pivotal tokens for long-context modeling tasks. Notably, our method significantly improves the performance on short- and long-text tasks by 80% and 76% respectively, reducing KV Cache by up to 50% with over 95% performance maintenance. The code is available at https: //github.com/PaddlePaddle/Research/ tree/master/NLP/ACL2024-NACL.

LoginMEA: Local-to-Global Interaction Network for Multi-modal Entity Alignment

Multi-modal entity alignment (MMEA) aims to identify equivalent entities between two multi-modal knowledge graphs (MMKGs), whose entities can be associated with relational triples and related images. Most previous studies treat the graph structure as a special modality, and fuse different modality information with separate uni-modal encoders, neglecting valuable relational associations in modalities. Other studies refine each uni-modal information with graph structures, but may introduce unnecessary relations in specific modalities. To this end, we propose a novel local-to-global interaction network for MMEA, termed as LoginMEA. Particularly, we first fuse local multi-modal interactions to generate holistic entity semantics and then refine them with global relational interactions of entity neighbors. In this design, the uni-modal information is fused adaptively, and can be refined with relations accordingly. To enrich local interactions of multi-modal entity information, we device modality weights and low-rank interactive fusion, allowing diverse impacts and element-level interactions among modalities. To capture global interactions of graph structures, we adopt relation reflection graph attention networks, which fully capture relational associations between entities. Extensive experiments demonstrate superior results of our method over 5 cross-KG or bilingual benchmark datasets, indicating the effectiveness of capturing local and global interactions.

IBMEA: Exploring Variational Information Bottleneck for Multi-modal Entity Alignment

Multi-modal entity alignment (MMEA) aims to identify equivalent entities between multi-modal knowledge graphs (MMKGs), where the entities can be associated with related images. Most existing studies integrate multi-modal information heavily relying on the automatically-learned fusion module, rarely suppressing the redundant information for MMEA explicitly. To this end, we explore variational information bottleneck for multi-modal entity alignment (IBMEA), which emphasizes the alignment-relevant information and suppresses the alignment-irrelevant information in generating entity representations. Specifically, we devise multi-modal variational encoders to generate modal-specific entity representations as probability distributions. Then, we propose four modal-specific information bottleneck regularizers, limiting the misleading clues in refining modal-specific entity representations. Finally, we propose a modal-hybrid information contrastive regularizer to integrate all the refined modal-specific representations, enhancing the entity similarity between MMKGs to achieve MMEA. We conduct extensive experiments on two cross-KG and three bilingual MMEA datasets. Experimental results demonstrate that our model consistently outperforms previous state-of-the-art methods, and also shows promising and robust performance in low-resource and high-noise data scenarios.

Optimal Transport Guided Correlation Assignment for Multimodal Entity Linking

Multimodal Entity Linking (MEL) aims to link ambiguous mentions in multimodal contexts to entities in a multimodal knowledge graph. A pivotal challenge is to fully leverage multi-element correlations between mentions and entities to bridge modality gap and enable fine-grained semantic matching. Existing methods attempt several local correlative mechanisms, relying heavily on the automatically learned attention weights, which may over-concentrate on partial correlations. To mitigate this issue, we formulate the correlation assignment problem as an optimal transport (OT) problem, and propose a novel MEL framework, namely OT-MEL, with OT-guided correlation assignment. Thereby, we exploit the correlation between multimodal features to enhance multimodal fusion, and the correlation between mentions and entities to enhance fine-grained matching. To accelerate model prediction, we further leverage knowledge distillation to transfer OT assignment knowledge to attention mechanism. Experimental results show that our model significantly outperforms previous state-of-the-art baselines and confirm the effectiveness of the OT-guided correlation assignment.