Long-Context Modeling

Capturing user interests across extensive watch histories is critical for short-form video recommendation, yet scaling sequence length is limited by two bottlenecks: the semantic sparsity of atomic Video IDs and the quadratic computational complexity of Transformers. Traditional orthogonal Video IDs fail to capture content relationships and demand large embedding tables, while the quadratic complexity of self-attention restricts the maximum sequence length under strict industrial latency and resource constraints. In this work, we present a production-deployed framework for modeling ultra-long user behavior sequences at a billion-user scale. We first address the representation bottleneck by adopting content-native Semantic IDs. By utilizing depth-truncated, coarse-grained Semantic IDs, we shrink the embedding table size from corpus cardinality. This compact representation naturally generalizes to cold-start content through shared semantic prefixes. Second, to overcome the sequence scaling barrier, we introduce a Global-Aware Compression Transformer that leverages non-parametric temporal folding and unified global query integration to effectively condense the sequence, alleviating both the memory and computational bottlenecks of standard self-attention. Offline profiling on our computing infrastructure demonstrates an order-of-magnitude reduction in peak memory footprint and a drastic decrease in computational overhead. This efficiency gain enables supporting longer sequence lengths at an affordable cost in production, yielding substantial online gains in satisfied user engagement and satisfied content consumption in large-scale online A/B tests.

Linear attention mechanisms have emerged as promising alternatives to softmax attention, offering linear-time complexity during inference. Recent advances such as Gated DeltaNet (GDN) and Kimi Delta Attention (KDA) have demonstrated that the delta rule, an online gradient descent update, enables superior associative recall compared to simple additive updates. While KDA refined the coarse head-wise decay gate into channel-wise decay, the learning rate $β_t$ in the delta update remains a scalar, limiting the model's capacity for dimension-specific adaptation. We introduce FG$^2$-GDN, which replaces the scalar $β_t$ with a channel-wise vector analogous to the transition from SGD to per-coordinate adaptive optimizers such as AdaGrad and Adam. We further propose FG$^2$-GDN+, which decouples the scaling for keys and values, enabling independent control of erasure strength and write strength. Experiments on synthetic and real-world benchmarks show that FG$^2$-GDN and its variant improve associative recall and long-context understanding over GDN and KDA, with comparable computational efficiency.

User preferences evolve across months of interaction, and tracking them requires inferring when a stated preference has been changed by a subsequent life event. We define this problem as long-horizon personalization and observe that progress on it is limited by data availability and measurement, with no existing resource providing both naturalistic long-horizon interactions and the ground-truth provenance needed to diagnose why models fail. We introduce a data generator that produces conversations from a structured mental state graph, yielding ground-truth provenance for every preference change across 6-month timelines, and from it construct HorizonBench, a benchmark of 4,245 items from 360 simulated users with 6-month conversation histories averaging ~4,300 turns and ~163K tokens. HorizonBench provides a testbed for long-context modeling, memory-augmented architectures, theory-of-mind reasoning, and user modeling. Across 25 frontier models, the best model reaches 52.8% and most score at or below the 20% chance baseline. When these models err on evolved preferences, over a third of the time they select the user's originally stated value without tracking the updated user state. This belief-update failure persists across context lengths and expression explicitness levels, identifying state-tracking capability as the primary bottleneck for long-horizon personalization.

Sparse attention has been proposed as a way to alleviate the quadratic cost of transformers, a central bottleneck in long-context training. A promising line of work is $α$-entmax attention, a differentiable sparse alternative to softmax that enables input-dependent sparsity yet has lagged behind softmax due to the computational overhead necessary to compute the normalizer $τ$. In this paper, we introduce AdaSplash-2, which addresses this limitation through a novel histogram-based initialization that reduces the number of iterations needed to compute $τ$ to typically 1--2. The key idea is to compute a coarse histogram of attention scores on the fly and store it in on-chip SRAM, yielding a more accurate initialization that enables fast forward and backward computation. Combined with a sparsity-aware GPU implementation that skips zero blocks with low overhead, AdaSplash-2 matches or improves per-step training time relative to FlashAttention-2 when block sparsity is moderate-to-high (e.g., $>$60%), which often occurs at long-context lengths. On downstream tasks, models trained with our efficient $α$-entmax attention match softmax baselines at short-context lengths and achieve substantial gains in long-context settings.

Reinforcement Learning (RL) has emerged as a critical driver for enhancing the reasoning capabilities of Large Language Models (LLMs). While recent advancements have focused on reward engineering or data synthesis, few studies exploit the model's intrinsic representation characteristics to guide the training process. In this paper, we first observe the presence of high-magnitude activations within the query and key vectors when processing long contexts. Drawing inspiration from model quantization -- which establishes the criticality of such high-magnitude activations -- and the insight that long-context reasoning inherently exhibits a sparse structure, we hypothesize that these weights serve as the pivotal drivers for effective model optimization. Based on this insight, we propose LongAct, a strategy that shifts from uniform to saliency-guided sparse updates. By selectively updating only the weights associated with these significant activations, LongAct achieves an approximate 8% improvement on LongBench v2 and enhances generalization on the RULER benchmark. Furthermore, our method exhibits remarkable universality, consistently boosting performance across diverse RL algorithms such as GRPO and DAPO. Extensive ablation studies suggest that focusing on these salient features is key to unlocking long-context potential.

Visual long-document understanding is critical for enterprise, legal, and scientific applications, yet the best performing open recipes have not explored reasoning, a capability which has driven leaps in math and code performance. We introduce a synthetic data pipeline for reasoning in long-document understanding that generates thinking traces by scoring each page for question relevance, extracting textual evidence and ordering it from most to least relevant. We apply SFT to the resulting traces within \texttt{<think>} tags, gated by a \texttt{<cot>} control token, and the resulting reasoning capability is internalized via low-strength model merging. We study Qwen3 VL 32B and Mistral Small 3.1 24B. With Qwen3 VL, we achieve 58.3 on MMLongBenchDoc, surpassing the 7$\times$ larger Qwen3 VL 235B A22B (57.0). With Mistral, we show that synthetic reasoning outperforms distillation from the Thinking version's traces by 3.8 points on MMLBD-C, and internalized reasoning exhibits 12.4$\times$ fewer mean output tokens compared to explicit reasoning. We release our pipeline for reproducibility and further exploration.

Despite the remarkable capabilities of Multimodal Large Language Models (MLLMs), they still suffer from visual fading in long-context scenarios. Specifically, the attention to visual tokens diminishes as the text sequence lengthens, leading to text generation detached from visual constraints. We attribute this degradation to the inherent inductive bias of Multimodal RoPE, which penalizes inter-modal attention as the distance between visual and text tokens increases. To address this, we propose inter-modal Distance Invariant Position Encoding (DIPE), a simple but effective mechanism that disentangles position encoding based on modality interactions. DIPE retains the natural relative positioning for intra-modal interactions to preserve local structure, while enforcing an anchored perceptual proximity for inter-modal interactions. This strategy effectively mitigates the inter-modal distance-based penalty, ensuring that visual signals remain perceptually consistent regardless of the context length. Experimental results demonstrate that by integrating DIPE with Multimodal RoPE, the model maintains stable visual grounding in long-context scenarios, significantly alleviating visual fading while preserving performance on standard short-context benchmarks. Code is available at https://github.com/lchen1019/DIPE.

Modern language models reason within bounded context, an inherent constraint that poses a fundamental barrier to long-horizon reasoning. We identify recursion as a core principle for overcoming this barrier, and propose recursive models as a minimal realization, where the model can recursively invoke itself to solve subtasks in isolated contexts. We prove that any computable problem admits a recursive decomposition of reasoning in which each subtask requires only exponentially smaller active context than standard autoregressive models; this strictly surpasses any context management approach confined to a single sequence, such as summarization. We further generalize our framework to modern agentic systems with arbitrary context processing and control flows, and prove that recursive models can achieve optimal power within this broader class. Experimentally, we test two settings: fine-tuning a pretrained base model for recursive SAT solving, and training a small model from scratch on Go traces generated by exact game-tree search. Both show improved long-horizon accuracy with small active contexts.

In real scenarios, videos can span several minutes or even hours. However, existing research on spatio-temporal video grounding (STVG), given a textual query, mainly focuses on localizing targets in short videos of tens of seconds, typically less than one minute, which limits real-world applications. In this paper, we explore Long-Form STVG (LF-STVG), which aims to locate targets in long-term videos. Compared with short videos, long-term videos contain much longer temporal spans and more irrelevant information, making it difficult for existing STVG methods that process all frames at once. To address this challenge, we propose an AutoRegressive Transformer architecture for LF-STVG, termed ART-STVG. Unlike conventional STVG methods that require the entire video sequence to make predictions at once, ART-STVG treats the video as streaming input and processes frames sequentially, enabling efficient handling of long videos. To model spatio-temporal context, we design spatial and temporal memory banks and apply them to the decoders. Since memories from different moments are not always relevant to the current frame, we introduce simple yet effective memory selection strategies to provide more relevant information to the decoders, significantly improving performance. Furthermore, instead of parallel spatial and temporal localization, we propose a cascaded spatio-temporal design that connects the spatial decoder to the temporal decoder, allowing fine-grained spatial cues to assist complex temporal localization in long videos. Experiments on newly extended LF-STVG datasets show that ART-STVG significantly outperforms state-of-the-art methods, while achieving competitive performance on conventional short-form STVG. Our code is at: https://github.com/HengLan/ART-STVG.

We present the first comprehensive, large-scale study of training long-context vision language models up to 344K context, targeting long-document visual question answering with measured transfer to long-context text. While several such strong are open-weight, namely Qwen3 VL and GLM 4.5/6V, their training recipes and data pipelines are not reproducible. We systematically study continued pretraining, supervised finetuning, and preference optimization for 24B and 32B parameter models, backed by extensive LC evaluations and ablations to bridge this gap, and achieve state-of-the-art performance on MMLongBenchDoc for both parameter scales. In addition to this, our key findings include: (i) training on context lengths that match evaluation context lengths outperforms training on longer contexts, (ii) training and evaluating with page indices provides a simple, high-impact boost to long-document performance, (iii) our synthetic data pipelines enable self-improvement via continued pretraining and supervised finetuning, and (iv) we extend the known text-to-visual long context transfer to the reverse, showing that visual long context training transfers to long-context text performance. We also release MMLBD-C, a manually corrected version of MMLongBenchDoc to reduce erroneous and low quality examples in the benchmark.

Long-context dialogue systems suffer from state inertia, where models over-attend to history and fail to adapt to evolving intents. We demonstrate that standard alignment methods like DPO and even recent long-context optimization techniques struggle to resolve this without incurring a severe contextual alignment tax--a substantial perplexity surge caused by disrupting pre-trained priors. To address this, we propose DZ-TiDPO, a minimally invasive framework that synergizes conflict-aware optimization (during training) with a structural temporal attention bias. This design effectively decouples state updating from general linguistic modeling. Experiments on Multi-Session Chat and our new Inertia Challenge (IC-Bench) show DZ-TiDPO preserves structural coherence while resolving inter-turn conflicts. Crucially, our framework supports dual inference strategies: a negligible-latency static mode for general robustness and a precision-focused dynamic mode for micro-semantic conflicts. Furthermore, our scaling analysis reveals a capacity-stability trade-off, confirming that highly capable mid-sized models (7B) can efficiently internalize temporal alignment. Code and data are available at: https://github.com/lyj20071013/DZ-TiDPO.

As model context lengths continue to grow, concerns about whether models effectively use the full context length have persisted. While several carefully designed long-context evaluations have recently been released, these evaluations tend to rely on retrieval from one or more sections of the context, which allows nearly all of the context tokens to be disregarded as noise. This represents only one type of task that might be performed with long context. We introduce Oolong, a benchmark of long-context reasoning tasks that require analyzing individual chunks of text on an atomic level, and then aggregating these analyses to answer distributional questions. Oolong is separated into two task sets: Oolong-synth, a set of naturalistic synthetic tasks, where we can easily ablate components of the reasoning problem; and Oolong-real, a downstream setting which requires reasoning over real-world conversational data. Oolong requires models to reason over large quantities of examples, to perform both classification and counting in-context, and to reason over temporal and user relations. Even frontier models struggle on Oolong, with GPT-5, Claude-Sonnet-4, and Gemini-2.5-Pro all achieving less than 50% accuracy on both splits at 128K. We release the data and evaluation harness for Oolong to enable further development of models that can reason over large quantities of text.

Long-context reasoning requires accurately identifying relevant information in extensive, noisy input contexts. In this work, we propose PERK (Parameter Efficient Reasoning over Knowledge), a scalable approach for learning to encode long contexts using gradient updates at test time. Specifically, PERK employs two nested optimization loops in a meta-training phase. The inner loop rapidly encodes contexts into a low-rank adapter (LoRA) that serves as a parameter-efficient memory module for the base model. Concurrently, the outer loop learns to use the updated adapter to accurately recall and reason over relevant information from the encoded long context. Our evaluations on several long-context reasoning tasks show that PERK significantly outperforms the standard long-context finetuning, achieving average absolute performance gains of up to 20% for Qwen-2.5 (0.5B & 7B) on synthetic and real-world long-context reasoning. PERK also maintains its advantages across model scales and families. Compared to specialized long-context LLMs, PERK matches or surpasses their performance. Finally, our analyses show PERK is more robust to reasoning complexity, length extrapolation, and the positions of relevant information in contexts. https://perk-long-context.web.app

Long-context understanding remains challenging for LLMs due to limited context windows. This paper introduces Long Input Fine-Tuning (LIFT), a framework that improves the long-context performance of arbitrary short-context LLMs by dynamically adapting their parameters to each long input. Instead of endlessly extending context windows to fit longer inputs in context, LIFT stores and absorbs the input in parameters. By fine-tuning long inputs into parameters, LIFT enables short-context LLMs to answer questions even when required information is absent from the inference context, avoiding the quadratic input-length complexity of standard long-context models. Rather than simple continued pretraining on new long contexts, LIFT uses carefully designed LLM-generated synthetic tasks to enhance comprehension beyond memorization. To offset fine-tuning overhead, we design a highly optimized pipeline that reduces Time to First Token (TTFT) to under 10 seconds for 8k context. We further analyze LIFT's strengths and limitations, discuss large-scale deployment feasibility, and highlight future research directions. Implementation is open-sourced at https://github.com/MuLabPKU/LIFT.

Linear recurrent neural networks, such as State Space Models (SSMs) and Linear Recurrent Units (LRUs), have recently shown state-of-the-art performance on long sequence modelling benchmarks. Despite their success, their empirical performance is not well understood and they come with a number of drawbacks, most notably their complex initialisation and normalisation schemes. In this work, we address some of these issues by proposing RotRNN -- a linear recurrent model which utilises the convenient properties of rotation matrices. We show that RotRNN provides a simple and efficient model with a robust normalisation procedure, and a practical implementation that remains faithful to its theoretical derivation. RotRNN also achieves competitive performance to state-of-the-art linear recurrent models on several long sequence modelling datasets.