The problem. A long-lived KV cache must be compressed before the queries that will read it exist; selection by observed attention (H2O, SnapKV) collapses there (0.00-0.33 needle retrieval on a NoPE MLA model), because a token's importance has not yet been observed. The method. On Kimi Linear, VestigeKV evicts by a query-independent signal the cache already carries: the 64-dimensional decoupled branch, a vestige of RoPE that NoPE training repurposes into a salience channel. Reading 11% of each row, it partitions the cache: the top-m rows stay in the attended tier; every other row moves -- exactly, never deleted -- to a GPU-resident archive reachable per step by a certified trigger. No training, no quantization, no weight or kernel change. Cost. Nothing measurable: retrieval holds at 1.00 under 8x and 0.92 under 32x from 8k to 65k context, zero gap to full-row selection. The attended tier is 0.25 KB of Kimi Linear's 8.1 KB per-token cache at 32x; the archive stays bit-exact and GPU-resident, with host offload as the VRAM-reclaiming variant. The recall tier -- the standard configuration -- holds 128x at 1.00. Kimi K3 is reported to use a NoPE Gated-MLA variant; if its cache layout matches, the method plausibly extends there -- we make no claim beyond the measured model. NoPE exclusivity. The identical operator on a RoPE MLA collapses to 0.08 (plain eviction: 0.42); query-independent salience itself exists only without rotation (top-1 targets span 2.3-6.7% of tokens vs. 10.2-46.8%), and query-universal exact merging is provably impossible under RoPE. All thresholds were frozen before data; 20 archived verdicts and 8 closed routes accompany the paper.
Cheolseung Baek, Dhammiko Arya, Eunki Kim +40cs.AI cs.CL
We introduce A.X K2, a 688B-parameter Mixture-of-Experts (MoE) language model trained from scratch as a high-performance foundation for \emph{agentic} applications. Trained on approximately 8.5T tokens---fewer than its predecessor, A.X K1---on a smaller but higher-quality mixture with substantially expanded agentic and software-engineering data, it nonetheless improves over A.X K1 across the board, by over 30 percentage points on some benchmarks, reflecting large gains in token efficiency. To support long contexts efficiently, we introduce Sparse Gated Attention (SGA), which combines sparse attention with gated attention, and adopt Gated Norm (GN) to stabilize large-scale training. SGA is trained natively at 128K through a \emph{sparse} indexer warmup that optimizes the indexer against its own sparse top-$k$ selection rather than the dense attention distribution, making adaptation markedly cheaper: each query reads only 2,048 positions, yet long-context quality is unchanged and A.X K2 scores 94.6 on RULER out to 256K. The outlier suppression of GN in turn keeps 4-bit NVFP4 serving within one point of FP8 accuracy. A simple yet effective Think-Fusion recipe further lets users switch between thinking and non-thinking modes within a single unified model. Extensive evaluations show that A.X K2 performs competitively against strong open-weight baselines, matching or exceeding them on math and Korean-language benchmarks.
Transformers rely on position embedding mechanisms in long context modeling in most cases. Rotary Position Embedding (RoPE) embeds positional information with independent 2D rotations, forming relative position terms in self-attention. However, its pairwise, block-based, and decoupled structure limits deep mixing and robustness across channels. We propose HD-RoPE, which extends RoPE from independent 2D rotations to higher-dimensional rotations and introduces a Paley-I orthogonal basis to obtain balanced, isotropic, and dense phase mixing within each rotation subspace. This significantly enhances channel coupling and rotational degrees of freedom while maintaining orthogonal stability and the relative position closure property. Furthermore, HD-RoPE is easily optimized for engineering efficiency without introducing additional trainable parameters. We have conducted extensive evaluation results demonstrating that HD-RoPE achieves significant performance improvements over standard RoPE across various popular benchmarks and in both long and short contexts.
Zhexi Feng, Ruiyi Zhang, Yongbo Yang +1cs.CL cs.AI
Conversations with chat assistants increasingly span many topics in a single long-running thread, challenging memory systems. Existing long-context and memory benchmarks often expose session or topic boundaries, or probe direct personal-memory questions. These settings understate a harder assistant-memory regime: a flat mixed-topic thread where the system must infer which earlier episode makes a later task decision valid. We introduce SCALE-QA, a constraint-grounded task QA benchmark for flat unsegmented threads targeting episode integrity failure. The dataset contains 3,000 audited questions across 10 domains, uses deterministic four-way multiple-choice grading, and includes a deterministic runtime builder; experiments use all 3,000 questions through 128k and a stratified 400-question diagnostic at 1M. SCALE-QA questions are ordinary task-oriented requests whose correct answer depends on causally related evidence introduced earlier in the conversation. We also propose Temporal-Semantic Interleaved Memory Reconstruction (TSIM), which segments the turn stream into coherent episodes and indexes them through a hierarchical multi-view memory stack with deterministic episode-level summary and cluster-routing views. Experiments show that SCALE-QA challenges strong RAG baselines and long-context LLMs alike; across three open-source and proprietary LLM backends, TSIM achieves the highest accuracy in every backend setting, gaining 5.6-17.6 accuracy points over the strongest corresponding baseline.
The quadratic growth of attention computation and key-value (KV) cache with respect to sequence length is a central bottleneck for ultra-long-context language models and high-resolution generative models. We propose ProxyFormer, a general dual-stream architecture built upon proxy tokens. In each layer, fine-grained local features are compressed bottom-up into a small set of proxy states; expensive global interactions are performed only in the compressed proxy space; the globally contextualized proxies are then decompressed and injected top-down back into the local stream. Because the local stream persists across layers, fine-grained information that is not captured by one compression step remains accessible for later refinement, alleviating the irreversible information loss of conventional one-shot compression. We further introduce factorized multi-level compression/decompression, layer-wise dynamic compression ratios, asymmetric dual embeddings, and a proxy-only KV-cache inference scheme. On a 16GB GPU with batch size 1, a standard decoder-only model can train sequences of only about 20K tokens, whereas ProxyFormer with a compression ratio of 64 extends the trainable sequence length to about 0.7M. A model trained with a 64K window retains 92%-95% retrieval accuracy on a multi-needle retrieval task with 1,048,576 tokens, and a model trained with an 8K window exceeds 94% accuracy when extrapolated to 256K tokens. Preliminary image-generation experiments demonstrate the feasibility of ProxyFormer for both pixel-space and latent-space flow matching.
Modern language models and agents increasingly require persistent memory for complete codebases, long interaction histories, and heterogeneous enterprise records. The key challenge is to keep hundreds of millions of tokens searchable while passing only bounded source evidence to the answer model. We introduce MegaMem, a source-resolved dual-view retrieval system that separates semantic access from generation evidence. Distilled records and detailed evidence are searched with original and transformed queries; every distilled hit resolves to an immutable source ID before reciprocal-rank fusion, deduplication, and cross-encoder reranking; and only the highest-ranked detailed evidence within a fixed budget supports generation. Post-answer attribution then identifies which loaded sources support the fixed answer. We evaluate MegaMem on EnterpriseRAG-Bench, which contains more than 500,000 heterogeneous enterprise documents and approximately 650M tokens. MegaMem improves Overall from 68.22 to 82.26 and reaches 86.50 Correctness. These results show that MegaMem supports ultra-large persistent memory while preserving strong answer accuracy under a bounded generation context. By separating searchable memory scale from answer-context size, MegaMem provides a practical path toward accurate retrieval over memories ranging from hundreds of millions to one billion tokens. Our code is available at https://github.com/ xfab-xinyuansong/MegaMem.git.
Dense causal attention remains expensive at long context even when implemented with highly optimized exact kernels. We study BF1, a deterministic block-aligned dyadic sparse-attention route that combines a small exact local neighborhood, a global first block, and logarithmically spaced historical blocks. The route is related to prior log-sparse and dilated attention patterns; our contribution is a correctness-gated pretrained-model retrofit, a matched topology-control study, and a systems characterization that connects per-layer sparsity to whole-model latency. For fixed block width, every converted layer uses O(n log n) selected token interactions and has O(log n) graph communication depth. On an NVIDIA RTX PRO 6000 Blackwell GPU, an optimized BF16 implementation crosses dense attention between 2K and 4K tokens and reaches a 10.91x per-layer prefill speedup at 32K. Retrofitting eight of 28 Qwen3-0.6B attention layers lowers warm whole-model time to first token by 7.7%, 11.3%, and 15.3% at 8K, 16K, and 32K, respectively, while the remaining dense layers keep the complete model asymptotically quadratic. Under a matched 1,000-step, 16.384M-token adaptation protocol, BF1 ranks first across three training seeds: mean report perplexity is 1.68639 versus 1.69154 for a matched static-random nonlocal graph, 1.69258 for dense continued training, and 1.81505 for equal-budget local sliding. At seed 1234, the packed-report paired interval places Dense-CT 0.3169-0.4055% above BF1 and static-random graph 17 0.2441-0.3642% above BF1. These results establish BF1 as a reproducible sparse operator and selective retrofit primitive with real long-context systems value. This paper evaluates numerical correctness, selected-interaction scaling, kernel performance, partial-model inference, and matched next-token language modeling.
Vincenzo Dentamaro, Pancrazio Auteri, Giuseppe Pirlocs.AR cs.CL
Current assessment of KV-cache compression performance confuses resident bits with read bandwidth and is affected by the artifacts of chunked teacher-forcing. We present Geodesia-KV, a family of training-free KV cache policies based on monotonic block-wise precision allocation, exact rate-distortion residuals, and query-sparse reading, enabling proper hardware-ready compression. With proper separation of resident and read bits and causal evaluation, we show that Geodesia-KV significantly outperforms other approaches. Specifically, on WikiText-2 with 16k context, the 5-bit operating point of Geodesia-KV results in lower perplexity at lower bitrate than KIVI-4 on Qwen. In addition, our compressed-Quest version delivers improved perplexity and reduces resident (9.83 vs 16.25 bits/value) and read rates (1.95 vs 2.32 bits/value) over baseline sparse methods on PG-19. As Geodesia-KV is implemented as native GeodesiaKVCacheManager plug-in of vLLM, Geodesia-KV fully removes the need for dense cache residency via monotonic bit demotion. With the full consumer hardware evaluation, Geodesia-KV leads to 1M-token context generation on a single 16 GiB GPU with up to 71.7% peak VRAM savings on all leading architectures (Qwen, Llama, DeepSeek).
We present MoNe, a lightweight modular neural memory that attaches to any frozen pretrained Transformer to enable long-context inference without retraining. MoNe reads context in fixed-size segments via test-time learning of fast-weight neural memory networks with layer-localized gradient updates; at inference, the memory generates keys and values from the query tokens alone, with no context tokens re-read. This two-phase design decouples inference cost from context length, achieving $O(N)$ preprocessing and $O(1)$ query cost with peak GPU memory that does not grow with $N$. At 128K tokens, MoNe reduces both compute and peak GPU memory by approximately 80% compared to ICL with only 6.4% parameter overhead. MoNe generalizes to context lengths far beyond the backbone's native window, achieving strong performance on needle-in-a-haystack and word extraction benchmarks from RULER, where ICL degrades sharply.
Linear attention keeps a state of fixed size. At long context, many stored items share this state, and interference between them degrades retrieval. Gated DeltaNet-2 (GDN-2), like every delta-rule model before it, derives its erase vector from the key of the current token. However, the interference in its reads is measured through the query, and the erase step cannot reach it. We introduce the Query-derived Erase Direction (QED). QED adds a second erase direction derived from the query and orthogonal to the key. In the fast-weight view, a key-directed delta edit cannot change the key-orthogonal part of a read. It uses the editable part to cancel old-state content measured along the query. It also improves retrieval at every length past the training window, and it about doubles the usable context length on S-NIAH-1.
Amanda Bertsch, Luca Soldaini, Matthew R. Gormley +4cs.CL
One might imagine that architectural variations within the dense transformer paradigm have a limited effect on accuracy. However, we demonstrate that this is not the case in the long context setting. Specifically, we show that a set of four minor architectural decisions --- all made by at least one of the Olmo, Llama, and Qwen dense model families --- have a compoundingly negative effect on long context extensibility. Any one of these choices alone has a minor impact on long context performance, but combining three or more can drop the performance downstream by up to 47%. Furthermore, these differences are not detectable from short-context loss or validation datasets. We show that much of the variation in long context ability across model families is driven by these architectural features and detectable from applying context extension early in pretraining. We demonstrate this with controlled ablations that hold data, tokenizer, and extension recipe fixed while varying normalization, GQA, pretraining context length, and sliding window attention. After over 170,000 GPU hours of training, we release the resulting set of models as OlmPool, a set of 26 comparable 7B models with checkpoints before and after long-context extension. This pool includes several architectures that outperform the Llama 3 architecture on long context extensibility. In an analysis of our ablation models, we identify patterns in attention sink behavior and attention distributions across context that are attributable to specific architectural differences.
Junghwan Lim, Joon Son Chung, Sungmin Lee +24cs.AI
We introduce Motif 3, a decoder-only Mixture-of-Experts language model with 314 billion total parameters and 13.2 billion activated per token. Each sparse MoE layer contains 384 routed experts, with eight selected per token. This fine-grained sparsity provides substantial expert capacity while limiting computation. Motif 3 is built around Grouped Differential Latent Attention (GDLA), which integrates grouped differential attention with the compressed key-value representation of Multi-head Latent Attention. The architecture further incorporates modified manifold-constrained hyper-connections, Expert Specific PolyNorm activations, and multi-token prediction to improve optimization stability, expert specialization, and inference efficiency. We pretrain Motif 3 on approximately 12.5 trillion tokens spanning web documents, STEM, code, mathematics, multilingual content, and domain-specialized corpora. Expert-balancing and numerical-stabilization techniques support stable training at scale, while selective MXFP8 computation and communication, memory-efficient fused kernels, and window-aware context parallelism enable training with context lengths up to 256K tokens. Our post-training pipeline combines general supervised fine-tuning, six specialist teachers trained with reinforcement learning, a software-engineering teacher trained with supervised fine-tuning, and Multi-teacher On-Policy Distillation. The resulting unified model consolidates complementary capabilities in reasoning, coding, tool use, professional work, long-context understanding, calibrated abstention, and instruction following. Across a broad evaluation suite, Motif 3 demonstrates competitive performance against leading open weight models, including strong results on long-horizon agentic tasks, mathematical reasoning, scientific knowledge, and hallucination-sensitive evaluation.
Large language models (LLMs) increasingly read long inputs in the agentic era, from whole documents and codebases to conversations across many turns. Their inference memory is then dominated by the key-value (KV) cache, the stored attention keys and values of every token the model has read and generated. Because the cache grows with context length and is re-read in full at every generated token, a longer context means more GPU memory. To reduce this cost, most existing methods compress the KV cache by lowering every stored value to the same low precision, a technique known as quantization. They can push this to nearly two bits per value, but rarely further, because quality drops sharply at this 2-bit cliff: four levels are too few for the cache's outlier-heavy values, where a few large entries consume the levels and collapse the rest into noise. A natural remedy is to spend more bits on the channels (feature dimensions) that matter and fewer on the rest, but the raw cache offers no handle: its channels are strongly correlated, so none stands out as more important. Our analysis shows that this handle appears once the cache is rotated into a coordinate system computed from its own statistics, removing these correlations. There, a small fraction of channels carries almost all the information, and spending the budget on those few is far more accurate than spreading it evenly. Guided by this analysis, we develop SPECTRA, a training-free, drop-in codec that re-encodes the cache into this coordinate system and concentrates the bit budget on the channels that carry the signal. On Llama-3.1-8B and Qwen2.5-7B over long-context benchmarks, SPECTRA is near-lossless at 4x compression, competitive at 8x where uniform quantization has collapsed, and reaches up to 12x, pushing usable compression past the 2-bit cliff so the same GPU holds longer contexts and larger batches.
As large language models (LLMs) process increasingly long contexts, KV cache storage and repeated access have become a major bottleneck. Existing KV cache compression methods rely on predefined, fixed compression rules and are typically developed around either token eviction or merging. As a result, cache resources can neither flow freely across layers, heads, and context slots, nor be jointly allocated to balance local resolution and information coverage. Therefore, we propose GraceKV, a global approach for the allocation of resolution and coverage in KV cache compression, and formulate the compression process as a global resource allocation problem under a fixed cache budget. GraceKV treats each layer-KV head-slot combination as an atomic unit and builds a prototype tree. Leaf nodes correspond to token-level KV entries, while each internal node uses a single prototype to compress the KV space covered by its children. A set of non-overlapping nodes in the tree forms the representation of an atomic unit. Adding the root of a new tree expands information coverage, whereas splitting a selected node improves local resolution. All candidate actions compete globally for a shared cache budget. Finally, the nodes retained across all trees form the compressed KV cache. This process adaptively determines the allocation of cache resources among atomic units globally and the balance between resolution and coverage. GraceKV requires no additional training, and the entire compression and inference process is performed on the GPU. Systematic experiments across diverse long-context tasks and compression ratios show that GraceKV ranks first in 24 of 32 settings and remains robust up to 128-fold compression. These results validate the effectiveness of global budget allocation in coordinating information coverage and local resolution.
Nicolas Zumarraga, Lorenzo Steno, Ning Wang +9cs.LG
Precise anomaly localization over long-context time series is a crucial task in monitoring applications across clinical care, industrial operations, financial services, and logistics, where brief evidence may hide inside long spans of high-frequency data. Time-Series Language Models (TSLMs) are able to ingest time series data and verbalize findings on anomalies in natural language; however, recent benchmarks report a decrease in retrieval performance at long contexts, mirroring failure modes in text, vision, and audio. In the text domain, Recursive Language Models (RLMs) can recover much of this lost performance by keeping context external to the large language model (LLM), allowing the model to query it through code. We present TimeRLM, an RLM formulation for time-series that sequentially manipulates the signal using code and vision capabilities. We further introduce AnomalyXL, a synthetic long-context anomaly localization benchmark with programmatically injected anomalies that require precise retrieval. We implement five different task categories and two variants: AnomalyXL-MCQ and AnomalyXL-Localize. TimeRLM outperforms every evaluated TSLM and single-pass baseline on four of the five AnomalyXL-Localize tasks, reaching 0.682 IoU on localization and 0.745 on classify-with-evidence, versus at most 0.329 and 0.072 across all baselines. We post-train TimeRLM using reinforcement learning. The resulting model further improves performance and requires approximately one-third as many agent interaction turns as its untrained base model to produce a final answer. On unseen real-world ECG, sleep and software observability recordings, the post-trained TimeRLM retains or improves performance, surpassing TSLMs despite being trained exclusively on synthetic data. Our findings suggest recursive interaction with time-series is an effective approach for long-horizon retrieval.
The attention score with rotary position embeddings (RoPE) decomposes exactly into a sum over its 2D-rotation frequency pairs, and each pair's wavelength limits how far it can discriminate position. Aligned with this structure, we propose the per-RoPE-wavelength distance window: it prunes the query--key inner-product terms beyond a wavelength-proportional distance. Unlike a sliding window, every key remains reachable, at least through the low-frequency pairs. The reduction rate is input-independent, with a closed form logarithmic in the sequence length $N$, in contrast to dynamic-sparse methods like MInference. Such token-level selection is orthogonal to our frequency-level pruning. The window can therefore be applied on top of those methods. On Qwen2.5-0.5B and Llama-3.2-3B, the window prunes 37--48\% of the query--key inner-product terms within each model's native context length. Relative to full attention, the top-1 match rate stays at 96--98\% and the mean output-distribution KL at the $10^{-3}$-nat level on LongBench-v2 contexts. We examine absolute scores on long-context benchmarks such as RULER, OpenAI-MRCR, LongCodeQA, and $\infty$Bench: they are broadly preserved. We implement the window as a slice of the query--key contraction axis, leaving the online-softmax recurrences untouched, and port it with minimal diffs into the released FlashAttention-4 prefill and FlashInfer decode. On RTX PRO 6000 with Llama, both ports outpace stock with gains growing with context length, up to $1.29\times$ at 128K. End to end on Qwen2.5-7B-1M, with 57\% of the inner-product terms pruned, the speedup reaches $1.31\times$ at a 1M-token context.
The growth of context window lengths in Large Language Models (LLMs) significantly enhances their long-context capabilities but incurs prohibitive memory costs due to the Key-Value (KV) cache. Although low-rank compression of KV cache is a promising remedy, existing methods face a dilemma: offline approaches depend on external calibration data, whereas online approaches incur substantial compute for full-prompt decomposition and reconstruction. In this paper, we propose S$^4$R, which builds low-rank subspaces from selectively sampled tokens and computes attention over a sparsely reconstructed KV representation. S$^4$R uses prompt-aware initialization to build initial key/value bases from a representative prompt subset, trading off calibration-data dependence against prefilling cost. Because fully reconstructing the cache at every decoding step is prohibitively expensive and hurts throughput, we further adopt sparse reconstruction to retain only informative positions during decoding. Extensive experiments on LongBench and RULER with Llama and Qwen model families show that S$^4$R achieves up to 5$\times$ KV compression with near full-cache accuracy, combining the efficiency of fixed compression with the adaptability of prompt-dependent methods.
Yao Xiao, Reuben Tan, Zhen Zhu +3cs.CV cs.AI cs.LG
Long visual context poses a challenge for vision-language models: performance degrades as the number of distractors grows, and processing all tokens at once is computationally infeasible under GPU memory constraints. We present ReToken, a single learnable embedding trained as an explicit retrieval target that selects a sparse set of query-relevant visual tokens from a pre-filled visual KV cache. Trained on only a small image-QA dataset, ReToken yields consistent gains across image and video benchmarks: on Visual Haystacks it improves Qwen3VL-8B by 13.4 points and InternVL3.5 by 12.4 points (>20% relative), and on LVBench it transfers zero-shot to long video for an 8.0-point gain with Qwen3VL-8B. Thanks to its lightweight design, both training and long-video inference fit on a single H100. Code is available at: https://github.com/avaxiao/ReToken
Long-horizon LLM agents accumulate reasoning traces, actions, and tool observations that can eventually exceed a model's fixed context window. Existing compaction methods address this limitation by discarding, summarizing, or retrieving earlier information, but they may remove task-critical details or fail to recover them reliably. We propose ARC (Addressable Recall Compaction), a context-management framework that separates archival storage from active-context presentation. ARC stores tool observations in an append-only, ID-addressable log and replaces older observations with compact citations when compaction is required. The agent can subsequently use these identifiers to request stored content without re-executing the corresponding tools or depending solely on similarity-based retrieval. We evaluate ARC using Qwen3-8B with a 16k context window and Qwen3-32B with a 32k context window. On the Needle-in-a-Haystack evaluation, ARC achieves an average exact-answer accuracy of 99.40%, compared with 88.12% for the best-performing baseline in our evaluation. ARC also reduces estimated serving time and HBM traffic under our hardware-cost model. On the LongBench-v2 Hard subset, ARC obtains an average accuracy of 29.97%, compared with 28.25% for the best-performing baseline. These results indicate that explicit, address-based recall can improve information retention and serving efficiency relative to the evaluated context-management baselines under the tested settings.
Kimi Team, Tongtong Bai, Yifan Bai +399cs.CL cs.LG
We introduce Kimi K3, a 2.8T parameter Mixture-of-Experts model with 104 billion activated parameters, native vision capabilities, and a 1-million-token context window. Kimi K3 is built on Kimi Delta Attention and Attention Residuals, which improve information flow across sequence length and model depth. Together with Stable LatentMoE, which effectively activates 16 of 896 routed experts per token, and refined training and data recipes, these advances yield an approximately 2.5x improvement in overall scaling efficiency over Kimi K2. Post-training highlights reinforcement learning across general, agentic, and coding domains and multiple reasoning-effort levels, enabling compositional generalization and robust long-horizon execution. At 2.8T scale, Kimi K3 is supported by infrastructure advances in multiple areas: algorithm-system co-design for KDA, perfectly balanced expert-parallel training with efficient memory management, million-token agentic RL with persistent rollout and sandbox states, and deployment innovations. Extensive evaluations show that Kimi K3 achieves frontier-level performance across long-horizon coding, agentic, knowledge, reasoning, and vision tasks. While its overall performance still trails the most powerful proprietary models, namely Claude Fable 5 and GPT-5.6 Sol, Kimi K3 consistently outperforms other open and proprietary models evaluated in our suite. We release the full Kimi K3 model weights to facilitate future research and accelerate the broader deployment and adoption of frontier intelligence.
Token-level sparse attention, as implemented by DeepSeek Sparse Attention (DSA) in production systems, makes the downstream attention efficient but shifts the bottleneck to the indexer that feeds it. To select the top-k tokens for each query, the indexer must still score every preceding token, incurring a cost of O(L^2) per layer for a sequence of length L. We observe that this per-query scan is largely redundant: nearby queries select highly overlapping top-k tokens, and the indexer scores are long-tailed along the key axis. We exploit these properties in PIVOT, Proxy Indexing Via One full-prefix Traversal, a training-free, drop-in replacement for the DSA indexer that shares one prefix scan across a group of nearby queries. PIVOT aggregates a group into a single proxy query, performs one shared full-prefix scan to obtain a candidate set, and then selects a top-k for each query from that set. Two variants trade speed for fidelity: PIVOT-Reuse shares the proxy top-k across the group for maximum speed, whereas PIVOT-Refine re-scores the candidate set with the indexer of each query and then selects an individual top-k, matching the dense indexer at a small additional cost. A single algorithm covers both inference phases, differing only in how groups are formed: fixed-size groups of consecutive queries in prefill, and the queries decoded together in one multi-token prediction (MTP) step in decode. On DeepSeek-V3.2 and GLM-5.1 across LongBench and RULER, PIVOT matches the accuracy of the dense DSA indexer while accelerating it by up to 4x and reducing end-to-end latency by up to 1.6x at long context.
Speculative decoding accelerates autoregressive generation by having a cheap draft propose tokens that a target verifies in parallel. Frontier models increasingly ship a built-in Multi-Token-Prediction (MTP/NEXTN) draft head under the assumption that the draft is negligibly cheap. At million-token context this breaks: an MTP draft head typically runs full attention over the entire KV cache at every draft step, so its read grows linearly with context and comes to dominate the draft cost -- precisely where speculation is most valuable. The effect compounds with draft length (a deep native draft can turn net-negative, slower than no speculation) and sharpens under hybrid/linear-attention targets, where cheaper verification leaves the draft's full-attention read exposed. We apply a StreamingLLM-style sliding window plus attention sink to the draft's attention only (Windowed-MTP), leaving full-attention verification intact. It is training-free, drop-in, and lossless by construction: the full-attention target still decides every accepted token, so windowing changes only which tokens are proposed, never which are accepted. It bounds the draft's KV working set to a constant, dropping ~99% of KV entries at 1M. Across three architecture families (Qwen GDN-MoE 35B/122B and a Mamba2-hybrid NoPE 120B) at 1M context on a single GPU in SGLang, windowing cuts the per-decode-step cost over the shipping native MTP draft by +28% to +44%, an input-invariant margin that widens with context. Since per-token latency is this cost divided by acceptance length, at matched acceptance end-to-end decode latency improves by the same amount, and more where windowing also lifts acceptance, while preserving the target's verified output distribution. Finally, the unread draft KV -- 7.7-11% of total KV at 1M -- is reclaimed via a compact ring buffer at no acceptance or quality cost.
Pruning a long context means committing to the blocks a model will keep, and the usual selector is distilled from a dense teacher's attention. That assumes attention shows which context the answer depends on. We test the assumption on retrieval tasks where the evidence is known exactly, by masking context and measuring whether the answer changes. Attention and causal dependence disagree. Teachers attend to outdated facts that the answer does not depend on, and they attend differently across training runs that use the same evidence. Selectors trained on that attention copy both failures. On a multi-hop retrieval task, a selector distilled from attention routes at 36% to 98% depending on the training run. The same selector trained on causal evidence sets reaches 99% or better on every run. Dense accuracy does not tell the teachers apart. Masking the frozen teacher recovers the causal sets of these tasks without annotations. Frozen pretrained models show the same conflict, and selectors supervised with known evidence labels beat attention-based eviction through 32B when context must be pruned before the question arrives.
Sungrae Park, Sanghoon Kim, Gyoungjin Gim +50cs.CL
We present Solar Open 2, a 250B-A15B Mixture-of-Experts language model built for long-horizon agentic tasks, scaled up from Solar Open 1 (Solar Open 100B). To hold entire agent trajectories in a single context, Solar Open 2 reaches a 1M-token window through a hybrid attention stack that interleaves one softmax layer among every three linear-attention layers, using no positional encoding and a gated delta rule extended to negative eigenvalues. To train at this scale under a fixed compute budget, we make training efficient in two ways: a stronger starting point, and higher-value data. For the starting point, we initialize Solar Open 2 from Solar Open 1, transferring the 5.69B-parameter shared skeleton that survives the architectural change and learning everything else through full pre-training. For the data, we curate for value per token: quality- and rarity-aware data curation and mixture-ratio optimization refine a 20T pool into a 10T mixture that, at equal token budget, outperforms the Solar Open 1 recipe. To build its agent skills, we train twelve domain specialists across purpose-built scenarios, then consolidate them into a single model by Multi-teacher On-Policy Distillation (MOPD). Against comparably sized open-weight models on English benchmarks, Solar Open 2 leads on MMLU-Pro, LiveCodeBench, and the APEX-Agents agentic suite, and stays competitive with the strongest (DeepSeek-V4-Flash and MiMo-V2.5) elsewhere. On Korean benchmarks, Solar Open 2 records the highest average of any model compared, including fast-tier closed APIs, and on Ko-GDPval, an in-house Korean officework-agent benchmark, it is competitive with DeepSeek-V4-Pro (1.6T) at less than a sixth of its size.
Large language models and LLM-based agents are widely used as personal chat assistants, enterprise copilots, and autonomous workflow agents. In all these applications, memory (the ability to retain, access, and reason over information accumulated over long contexts and multiple interactions) plays a crucial role in determining the reliability of any agent. We introduce RECON (Reasoning over Extended Contexts with Obfuscated Narratives), a benchmark for evaluating compositional reasoning over long contexts. RECON spans 24 case files across three domains (criminal, medical, and financial), each ranging from 50k to 100k tokens, and tests agents on six memory intensive tasks: reconstructing multi-hop evidence chains, propagating cascading invalidations, resolving source conflicts, counterfactual reasoning, satisfying temporal constraints, and temporal fact retrieval. Recent memory benchmarks evaluate whether agents can retrieve scattered facts or detect if a fact has changed whereas RECON evaluates what happens after the change, whether agents can trace which downstream conclusions are affected, which survive through independent support, and how alternative timelines would have unfolded. Our evaluation reveals substantial limitations across current architectures: even the strongest non-Oracle system reaches only 22.4% Accuracy, with retrieval and reasoning each surfacing as challenges.
Gleb Kuzmin, Ivan Rodkin, Aydar Bulatov +8cs.CL cs.AI
Extending the context length of large language models (LLMs) is critical for many real-world applications, yet standard transformers remain constrained by quadratic compute and linear memory scaling. In this work, we investigate the Associative Recurrent Memory Transformer (ARMT) as a practical approach for enabling long-context processing in LLMs, constant memory scaling, and better efficiency. We make three main contributions. First, we construct two domain-specific long-context datasets designed to evaluate realistic workloads, focusing on narrow-domain fine-tuning scenarios. Second, we propose a comprehensive training recipe for ARMT-based context extension, combining continued pre-training, synthetic long-context data generation, curriculum learning, and selective integration of associative memory into chosen model layers. Third, we present an extensive experimental study demonstrating that ARMT-augmented models: (i) process inputs well beyond their original context limits without degrading performance relative to in-limit baselines; (ii) generalize more effectively to out-of-distribution context lengths; and (iii) need 30% less FLOPs while preserving baseline performance within the original context window.
Context engineering decides what information a model carries forward, and current designs meter it in tokens: compressing the past into a bounded recurrent state, keeping a key-value entry for every token, or imposing a fixed budget through a window or eviction rule. All three make the token the unit of memory even when the stream is redundant and the task depends on the distinct information it carries. Building on a companion mechanism paper that opens a cache slot only when an incoming key is novel, so memory scales with the number of distinct items rather than tokens, we develop that allocate-on-novelty cache as a working-memory component and organize context by how a task depends on the past: recall-carried information belongs in a content-addressed novelty cache, summary-carried information in a recurrent state, and locality-carried information in a recency window. The claim is empirical and bounded. On a matched character-level control, novelty-gated attention reaches full-attention performance while attending to about half the tokens, and coupling the cache with a state-space summary matches full-attention coupling at that reduced cost; the advantage grows as context lengthens, while a sliding window is preferable on short, locality-dominated spans. On next-code prediction over synthetic Medicare claims the coupled component leads full attention and every fixed-budget eviction policy at a thousand-event horizon, whereas cost forecasting over the same stream is summary-carried and the cache is neutral. The retained memory is an inspectable table of templates, codes, drugs, or places rather than an opaque state. The experiments are small-scale and use only public data; they establish the primitive that context can scale with distinct information rather than tokens, in a working memory that is content-addressable and auditable.
Large Language Models are increasingly deployed for sophisticated data engineering tasks such as generating structured queries from natural language, Text-to-SQL, and automating complex spreadsheet operations. However, maximizing their utility demands both higher finetuning-free accuracy and solutions to the computational bottleneck imposed by the Transformer architectures inherent quadratic (On2) time complexity. This paper introduces a novel drop-in neurosymbolic layer designed to seamlessly integrate into existing LLM backbones enhancing logical reasoning and mitigating long-context resource consumption. On the reasoning front, the layer immediately and significantly improves performance yielding an average accuracy increase of 85% across rigorous benchmarks including BIRD-CRITIC and LiveSQLBench, critically achieving these gains without any task specific finetuning or RLHF. Concurrently, we repurpose this approach to address the severe computational strain of long context inference. By leveraging symbolic processing to prioritize and compress relevant contextual information the layer reduces the effective token usage by over 50% and brings the effective time complexity down from O(n2) to approximately O(n) on certain long context tasks. This dual impact approach not only makes LLMs substantially more reliable for data engineering but also drastically reduces the computational pressure on inference chips, making long context tasks more manageable and cost effective.
Shiva Kaul, Min-Gyu Kim, Anjum Khurshid +1cs.CL cs.AI
The verbose context problem occurs when structured concepts have token-inefficient textual representations. This bottleneck is acute in population health: cohort-level analysis of longitudinal patient records requires reasoning over thousands of medically-coded events, often exceeding 400K tokens in total. We present PopMedQA, a benchmark isolating this problem through computational tasks on groups of longitudinal patient records. We construct the benchmark using neopatient, a new library for language-controlled generation of artificial patient records. Through extensive ablations -- including prompting strategies, prompt compression, and agentic decomposition -- we find that domain-independent methods fail to alleviate the verbose context problem. There remains significant opportunity to exploit domain-specific structure in language model inputs for population-scale reasoning.
Proposal. Long context can replay history, but it does not decide which completed observations deserve authority. MMLA formalizes a bounded resident memory between transient context and slow weight updates. A completed local segment is eventized; for each event, a target-conditioned constructor proposes semantic content and a trusted assembler produces a complete versioned row; deployment either commits that row atomically or returns NULL. Realized futures may price actions during training, while deployment remains causal and future-blind. Validated components. Controlled studies establish narrower ingredients. Lifecycle execution is exact on 300/300 held-out records for each of three seeds. Calibrated selection with full-archive fallback improves over a weak budget-matched dense baseline by 5.5--16.6 F1 and over BM25 by 4.0--6.2 F1 on held-out multi-hop QA; the original Llama budget execution is retained as failed, while the corrected Qwen packer satisfies the stated per-record caps. Typed anchor--filler transport reaches 240/240 held-out exactness per seed while three same-checkpoint controls obtain 0/240 whole-record successes. Current blocker. The integration loop is not complete. Dense-row, structured-span, and checkpoint-native readers trained from V28, the frozen 352.3M-parameter model-only checkpoint produced by a 50.0M-token native-scaffold pilot, all fail semantic qualification across three seeds. Candidate exactness is 0--45/23,040, query exactness is 10--1,536/9,216, and record-macro Brier remains near the 1,025-class uniform reference. Structural mapping passes, but none of the nine jobs qualifies. Predictive overwrite is therefore closed by gate.