Transformers

Long-horizon and multi-turn agents typically generate short actions and process long observations from tools and environments. This growing context demands efficient prefill, compact KV-cache storage, and accurate long-context retrieval. To meet these demands, we introduce HySparse2, a hybrid sparse attention architecture with two-level KV sharing. At the outer level, KV Bridging adopts a YOCO-style self-decoder and cross-decoder structure, but bridges only full-attention layers. The self-decoder uses hybrid sliding-window attention (SWA), while the cross-decoder uses hybrid sparse attention. The KV caches for full-attention layers in the cross-decoder are generated from the hidden states of full-attention layers in the self-decoder. At the inner level, HySparse2 retains HySparse's core KV Reuse design with two refinements. First, it replaces block-level sparsity with token-level sparsity for finer long-context retrieval. Second, it removes the separate SWA branch from sparse layers and instead forces a sliding window of recent tokens into the sparse selection. This two-level KV sharing allows all cross-decoder KV caches to be constructed from self-decoder hidden states. Prefill can therefore exit after the self-decoder, skipping all cross-decoder layers. On an 80B-A3B MoE model, HySparse2 outperforms HySparse and Hybrid SWA on long-context retrieval and multi-turn agentic tasks, while substantially reducing prefill computation and KV-cache storage.

Despite their strong performance, large language models (LLMs) are bottlenecked by KV cache memory traffic during long-context inference. Sparse attention is widely used to accelerate LLM inference by computing exact attention over a selected subset of tokens. To recover the contribution of tokens excluded from exact attention, recent methods apply coarse-grained compensation to the omitted attention tail. However, existing methods typically select tokens based on attention mass and only then compensate for the unselected tokens. This decoupled design overlooks their interaction: selection should prioritize tokens that would leave the largest compensation error if omitted. To address this limitation, we introduce CompKV, the first compensation-aware sparse attention framework that divides tokens into blocks and explicitly optimizes selection for the downstream compensation mechanism. Our theoretical analysis shows that the residual left by block-level mean compensation is governed by both block attention mass and within-block logit variation. We approximate this residual using compact block-level statistics, yielding a deployable selection criterion. We further develop an efficient asynchronous implementation. Experiments on RULER and LongBench-Pro show that CompKV performs best among the evaluated sparse baselines while delivering up to a $6.85\times$ self-attention speedup over full attention.

Transformer architectures exhibit cross-layer redundancies, yet post-training compression pipelines typically optimize layers in isolation or rely on heuristic grouping strategies that disregard layer-specific activation geometries. We introduce a principled, training-free framework that sequentially optimizes cross-layer weight pairings and shared-dictionary factorizations. Rather than forcing weights of adjacent layers to share a basis or heuristically merging activation statistics, our approach identifies structurally compatible projections and learns a shared representation that better preserves each layer's distinct calibration geometry. Coupled with structured sparsity, this yields highly efficient weight decompositions without sacrificing functional fidelity. Across diverse architectures, scales, and modalities, our method achieves state-of-the-art results, consistently outperforming independent structured weight decompositions and alternative pairwise weight factorizations, which operate under heuristic grouping strategies. By replacing heuristic engineering strategies with a convergent, optimization-driven pipeline, we establish a theoretically grounded foundation for scalable, transformer compression across different modalities.

In this note, we show that the problem of checking, whether a sequence of bits is ordered, is not doable by 1-head 1-layer transformers but is doable by a 2-head 1-layer transformer. Unlike similar previous results, our results assume the model where transformers have an output MLP.

Linear RNNs based on the delta-rule enable efficient sequence modeling, but their linear updates with a low-rank correction constrain their expressivity. Prior work has shown that composing two delta-rule transitions in a single recurrent update can model a 2D rotation, but this increases the rank and the cost of the updates compared to a single transition. We show that Kimi Delta Attention (KDA) can realize 2D rotations by combining a single delta-rule transformation with a second reflection supplied by its channel-wise gate. This requires extending the parameter ranges of KDA by combining two existing range extensions: allowing gates in $[-1,1]$ and the delta-rule coefficient $β$ in $[0,2]$. We call the resulting model Complex KDA (CKDA). It preserves KDA's stability and efficiency, with transitions that remain diagonal-plus-rank-one and non-expansive, while reaching the state-tracking expressivity of DeltaProduct$_2$. We characterize the expressivity of CKDA and prove that every orthogonal diagonal-plus-rank-one matrix is exactly a CKDA transition matrix. A single CKDA layer can track every finite group isomorphic to a subgroup of $\mathrm{SO}(3)$, and many state-tracking results use one fewer layer for CKDA compared to other diagonal-plus-rank-one Linear RNNs. Empirically, combining both extensions yields the strongest length extrapolation among tested KDA range settings on $S_3$, $S_4$, and periodic audio continuation. In language modeling, CKDA outperforms Transformers and other linear RNNs, obtains similar results to a KDA baseline, and shows promising scaling behavior. Our code is open source at https://github.com/OpenEuroLLM/ComplexKDA and our models are available at https://huggingface.co/collections/openeurollm/complexkda.

Transformer models built on the attention mechanism have become a central building block in modern deep learning, yet softmax attention remains a major bottleneck for long-context workloads. While FlashAttention makes the forward and first backward passes I/O-efficient, it does not support backward-over-backward (BoB), which enables exact differentiation through the backward pass for applications such as second-order optimization, test-time training, gradient-based memory, and meta-learning. Existing BoB implementations either materialize large intermediate tensors or exhaust GPU memory at long sequence lengths. We present FlashBoB, an exact, I/O-efficient algorithm for BoB in softmax attention that keeps computation within on-chip tiles and avoids all $N \times N$ intermediate tensors, where $N$ is the sequence length. The key insight is a hierarchical affine structure in the softmax double backward: two row-wise scalars determine all outputs through affine transformations. This yields a two-pass schedule with bounded on-chip static random-access memory (SRAM) usage and minimal off-chip high-bandwidth memory (HBM) traffic. FlashBoB achieves $Θ(N^2 d^2/M)$ HBM traffic ($d$ is the head dimension and $M$ is the memory size) and, within the standard FlashAttention-style score-recomputation model, matches the inherited large-cache lower bound for exact forward attention. Empirically, it scales exact attention BoB to $N=262\text{K}$ on a single A100 80GB GPU, where prior PyTorch exact baselines fail by $N=16\text{K}$, and is up to $6.3\times$ faster than FlashBack. These results make exact second-order attention practical at long-context sequence lengths where prior implementations cannot run efficiently.

Activation steering modifies LLM behavior at inference time, but identifying where and how strongly to steer remains manual. We introduce Deep Noir, a framework that uses Logit Lens convergence and causal head-level attribution to autonomously discover optimal steering parameters. Across three scales (1B x 3, 2-3B x 2, and 7-9B x 4), our engine achieves 16.7 percentage-point improvement on spam at 1B (standard deviation 4.7; 39 runs), with gains increasing to 21 to 42 percentage points at 7-9B across four architectures. On SST-2 sentiment, it achieves a 13.1 percentage-point improvement with zero code changes. Mechanistic grounding enables automated discovery of intervention points that generalize across tasks and architectures. On sentiment, RepE without head masking fails to improve over baseline, while Deep Noir improves all models (p less than 0.01). We further show that steering creates a predictable prompt-injection attack surface whose vulnerability increases monotonically with steering magnitude. This finding is relevant to agent systems deploying steered classifiers.

Deep learning predictive maintenance models suffer from poor transferability across machines and operating conditions, especially when labelled data are scarce and signals span five orders of magnitude in sampling frequency (1 Hz to ~100 kHz). We propose FreqCondNorm, a Transformer-based architecture that introduces a FiLM-style frequency-conditioned normalization layer to unify heterogeneous time-series within a single model. The architecture is pretrained on five public predictive maintenance datasets (CWRU, MFPT, UOC18, PRONOSTIA, CMAPSS) using masked auto-encoding and contrastive learning with balanced domain sampling. On fault diagnosis, the model achieves 99.2% accuracy on CWRU (+6.4 pp over CNN) and 82.1% zero-shot accuracy on MFPT, demonstrating strong transfer across sampling frequencies. However, the approach does not improve remaining useful life prediction, suggesting a mismatch between pretraining and RUL objectives that warrants future investigation.

We present Relational BabyLM, a system submission to the BabyLM 2026 challenge that combines two cognitively motivated inductive biases in a single decoder-only Transformer. Architecturally, we replace standard self-attention with a Dual Attention Transformer (DAT), which separates the routing of object-level ("sensory") lexical features from structural/relational information (Altabaa and Lafferty, 2025; Altabaa et al., 2024; Webb et al., 2024; Kerg et al., 2022; Webb et al., 2021). Relational attention (RA) disentangled from self-attention greatly increases data efficiency and out-of-training-sample generalization on purely relational tasks, but language modeling requires object-level and relational information to be integrated as well as disentangled, and RA-based LMs have remained largely unexplored. BabyLM's data-constrained training and comprehensive evaluation is an ideal testing ground for whether that data efficiency transfers. As a training intervention, we add a Next-Latent Prediction (NextLat; Teoh et al. 2026) objective that encourages hidden states to compress history incrementally into a dense belief state. Architecture is the dominant factor for structural linguistic generalization; the objective is secondary but still significant. DAT's three relational attention types (full RA vs. the simpler RCA and DisRCA variants) are largely interchangeable at 10M words; full RA pulls ahead at 100M. We also introduce a novel symbol-retrieval mechanism (RoPE-based, as opposed to learned, relative symbols) that matches learned symbol libraries while adding no parameters. On the strict (100M-word) track, our best model ranks 6th of 55 overall and 3rd of 55 on the leaderboard's NLP-task subset at the time of writing; our two strongest models outperform the GPT-2 baseline on most benchmarks, with one attaining the highest EWoK score among strict-track entries.

Recent findings indicate that Vision Transformers settle into locally similar computational phases, implying a level of depthwise computational redundancy. However, existing methods to exploit this redundancy either fail to reduce inference compute or severely degrade model expressivity. In this work, we formalise a unified view of block redundancy that decouples the geometry from specific surrogate interventions. We then introduce Transformer-Within-Transformer (TWT), a post-hoc method that fuses contiguous groups of redundant layers into a single learned surrogate layer. TWT reduces parameter count and inference compute while remaining competitive with original models using half the depth on natural images, and in several downstream histopathology settings, TWT matches or even improves on the original baseline.

Long-term multivariate time series plays a significant role in many application areas such as power systems, trading, etc. However, their accurate prediction is quite difficult for conventional forecasting methods as they often exhibit high dimensionality and complex relationships. Recent works show that transformer-based approaches are quite effective for long-term forecasting thanks to their attention mechanism. However, in the presence of complex high-dimensional inputs, they show evidence of oversmoothing, limited capacity, and opacity. To this end, this paper introduces SETTer, a transformer-based model that addresses these challenges by incorporating novel techniques for decoupled self-attention and hybrid masking. The proposed techniques enable SETTer to effectively capture the dominant short- and long-term patterns across the temporal and channel dimensions. In addition, we enrich the model layers with simple explainable structures that indicate the discriminative pattern of SETTer. We show that with a single-layer transformer architecture, SETTer can effectively model long-term dependencies in the presence of varying data complexities. Extensive experiments on real-word benchmark datasets for long-term multivariate time series forecasting demonstrate that SETTer outperforms state-of-the-art models in 88% of the scenarios.

The widespread adoption of long-horizon agents has made model workloads increasingly input-heavy. Although prior work has substantially reduced the cost of long-context computation, prefill remains computationally expensive, and large KV caches continue to strain HBM and SSD capacity and data-transfer bandwidth. Together, these compute, storage, and bandwidth demands constitute the primary bottleneck to further lowering deployment costs. To address this challenge, we introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. With its Causal Encoder-Decoder (CED) architecture, the model activates 16B parameters per token during decode but only 8B parameters during prefill, substantially improving cost efficiency for agentic workloads. To push the limits of KV cache compression, DeepSeek-V4.1-Flash combines cross-layer KV cache reuse in Compressed Sparse Attention 2 (CSA2) with FP4 KV caching. These designs reduce its global KV cache footprint (always in HBM) to 890 bytes per token, roughly 1/4 of the corresponding footprint of DeepSeek-V4-Flash. Further, through a dedicated deployment optimization known as SWA Bounded Replay, DeepSeek-V4.1-Flash reduces its persistent KV cache footprint (always on SSD or in host memory) to roughly 1/8 of that of DeepSeek-V4-Flash. Despite its much smaller KV cache footprint, the model delivers substantially better performance than the baseline. In addition, we streamline the DeepSeek-V4 architecture and introduce several efficient architectural extensions. We pretrain DeepSeek-V4.1-Flash on a multimodal corpus comprising 45T tokens and conduct comprehensive post-training, yielding strong performance across diverse text-based and multimodal agentic scenarios. Model checkpoints are available at https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash.

Depth-recurrent language models iteratively apply a small layer stack, decoupling per-token compute from distinct parameter count. To determine whether such a model genuinely utilizes its depth, both recurrence and layer-pruning literatures rely on a shared evaluation: truncating depth at inference time, plotting quality against retained depth fraction, and reading off the slope. While cheap and training-free, this metric suffers from an unexamined flaw: it extracts a single scalar from an intervention that alters multiple model properties simultaneously. Depth truncation concurrently reduces the number of block applications, decreases the volume of distinct computation performed, and pushes the readout head onto an out-of-distribution residual stream. The observed slope conflates all three factors, yet is conventionally interpreted as reflecting solely the second. We propose the Depth Control Protocol (DCP), a diagnostic suite that disentangles these three quantities. DCP comprises three positive controls that isolate each factor while varying the others, a negative control applying the identical interventions to dense transformers to ensure the effect is not an artifact of the measurement protocol, and a controlled training intervention to verify causality. The linchpin control, running the full budget of block applications while executing only a single distinct iteration, is strictly realizable only in depth-wise weight-sharing architectures, since in a dense network repeating a layer yields an entirely different model rather than the same model in an alternative configuration.

Autoregressive image generation has emerged as a paradigm for multimodal AI systems due to its compatibility with transformer-based LLM serving infrastructures. However, generating thousands of visual tokens per request makes decoding increasingly bottlenecked by KV cache accesses during attention computation. Sparse attention is particularly attractive for this workload because many visual generation applications tolerate moderate quality degradation in exchange for improved performance and efficiency. While sparse attention has been extensively explored for text-based LLM inference, it remains unclear whether its sparsity assumptions generalize effectively to autoregressive image generation. We present the first systematic characterization of attention sparsity in autoregressive image generation across diverse workloads and representative open-source models. Our analysis reveals several distinguishing properties, including a pronounced prefill-decode asymmetry, strong attention concentration on prompt and local tokens, and a unique diagonal attention sparsity pattern arising from the spatial locality of visual tokens. Motivated by these observations, we propose a diagonal-aware sparse attention mechanism that selectively skips KV entries along the diagonal attention direction within a recent window. Implemented on top of a GPU-based serving system using FlexGen, FlashAttention-2, and custom kernels, our approach achieves up to 3.1x throughput and 1.19x latency improvements with less than 2% quality degradation compared to dense inference.

Recurrent-depth Transformers apply shared computation repeatedly, but differ in how they retain information across steps. We compare a Block Universal Transformer (BUT), which carries only its current hidden state; CoTFormer, which also retains an expanding attention cache; and a new LSTM Universal Transformer (LSTM-UT) with bounded gated memory. On Rule 30 cellular automata, BUT extrapolates to unseen recurrent depths more reliably than CoTFormer, although its accuracy eventually degrades. State and cache interventions show that CoTFormer's failure depends on their interaction: correcting the current state can temporarily restore accuracy, while retained history can undermine that correction. In a delayed-recall task, BUT also outperforms CoTFormer despite lacking direct access to past states; CoTFormer does not reliably select the requested cached representation. LSTM-UT improves both depth extrapolation and delayed recall over these baselines. The results support bounded gated memory as an effective inductive bias for repeated computation and later retrieval in these tasks.

Scaling laws predict how loss decreases with increases in computation. We show, contrary to conventional wisdom, that architectural interventions can modify scaling exponents in pre-training, leading to power-law improvements in performance as computation increases. As an anchoring point, we consider the architectural formulation of looped transformers. Although not typically used in this way, looping, also known as recursive depth, provides a mechanism for model growth, by increasing the number of loops during training. Model growth, with and without shared weights, provides the biggest changes to the scaling exponents. In particular, a 7.4B model growth architecture matches GPT-3 13B on CORE with roughly $20\times$ less compute, and has compute efficiency gains that increase with scale. Moreover, simply using a boundary operator in a vanilla transformer, which normalizes and injects an earlier block, also provides an exponent increase, although to a lesser extent. In the data-constrained, multi-epoch setting, standard looping has a useful regularizing effect, where we find it is compute-optimal to increase the number of loops with scale. These results can be understood through the lens of computational depth: for a given computational budget, we wish to increase the usable depth of the transformer, which can lead to efficiency gains that increase with scale.

We analyze the learned input-output behavior of GLU-based neurons in large language models (LLMs). We propose a simple analysis method: For each neuron, we compute the cosine similarities between its input (reading) and output (writing) weight vectors. In this scheme, a strong negative cosine similarity indicates the neuron weakens the direction it detects in the residual stream, so we call this a weakening neuron. This allows us to gain a number of novel insights. First, we show that nine different LLMs have similar patterns: weakening neurons appear mostly in late layers whereas their counterparts, (conditional) strengthening neurons, are frequent in early-middle layers. Second, we find that weakening neurons display surprising behavior: even though there are few, they activate often and have a large influence on model behavior. Third, weakening neurons have a strong effect on model output when gate values are negative -- which is surprising since negative gate values are not expected to encode functionality.

We propose a compact NISQ-compatible quantum transformer architecture for synthetic QNLP sequence modelling. The model preserves the autoregressive next-token interface of a classical transformer, but replaces attention and feed-forward sublayers with variational quantum encoder blocks, connector circuits, decoder blocks and a direct two-qubit measurement readout. Token contexts are angle-encoded into small quantum registers, processed by parallel variational heads and encoder integration circuits and conditioned through decoder ancillae to produce a distribution over a four-token vocabulary. We evaluate several architecture variants on deterministic and lexicographic grammar-generation tasks against a compact classical transformer baseline. The quantum models are trainable end-to-end and learn nontrivial grammar structure, including perfect deterministic generation in individual runs and high lexicographic validity in the strongest variant. The classical baseline remains more accurate and stable and the quantum models are sensitive to initialization. The contribution is therefore not a claim of quantum advantage, but a concrete architecture and evaluation of transformer-inspired QNLP sequence modelling under near-term quantum constraints.

Autoregressive transformer decoding is constrained by irregular key-value (KV) cache movement on tiled accelerators. Prior compression and DRAM-placement systems still concentrate traffic on centralized memory paths that bottleneck long-context serving. We present MeshKV, a KV cache fabric that moves blocks as packetized flows over a lightweight NoC. It co-designs (i) TaKV affine striping to spread homes and cut hotspot load, (ii) Mare multicast with verified duplicate suppression, and (iii) Pad, which overlaps prefetch, tile multiply, and streaming softmax behind credit-aligned FIFOs. Together they convert bisection back-pressure into useful KV transfer. On our 8x8 FPGA implementation with LLaMA-2-7B and Mistral-7B at 8K-32K, MeshKV reduces interconnect traffic by up to 58%, improves KV bandwidth utilization by 2.1x, and delivers up to 1.9x multi-stream throughput.

Large Language Models (LLMs) frequently exhibit hallucinations, presenting a major barrier to reliability in complex reasoning tasks. While traditional detection methods rely on output-based confidence metrics, these logits are often miscalibrated by modern alignment techniques. In this paper, we investigate the temporal volatility of internal attention mechanisms as an alternative diagnostic signal for hallucination that does not depend on output calibration. By introducing an unsupervised metric for attention dispersion, we show that epistemic uncertainty leaves a measurable trace within intermediate layers, where spikes in attention entropy are associated with reasoning breakdowns. We evaluate our approach on mathematical reasoning benchmarks (GSM8K and MATH-500) using the Qwen2.5 model family (1.5B and 3B parameters), finding statistically significant AUC improvements of up to +0.076 over output-based baselines across all tested conditions. These findings suggest that attention dispersion is a promising complement to traditional hallucination detection methods, requiring further investigation across broader model families and task domains.

Mixture-of-Experts (MoE) architectures decouple model capacity from computational cost, yet incur high memory footprints as parameters grow linearly with the number of experts. Recurrent Transformers achieve parameter efficiency by reusing layer weights, but typically lack the capacity for competitive language modeling. We propose Mixture of Reused Experts (MoRE), a hybrid that shares expert pools across groups of adjacent layers. Each layer retains its own router but selects from a larger shared pool, expanding the diversity of routing combinations without additional parameters. To enable shared experts to distinguish between layers, we introduce lightweight learnable depth embeddings that condition each layer's input before routing. Experiments across three model scales (114M-1.15B parameters) show that MoRE consistently achieves lower perplexity and stronger downstream performance than standard MoEs and state-of-the-art weight-sharing architectures at matched compute and parameter budgets, with only minimal modifications to existing MoE implementations.

We study the problem of demonstration selection, which involves selecting a subset of examples for prepending to a query to a language model. This problem is closely related to in-context learning and language model inference. Since the inference cost of a transformer model scales quadratically with sequence length, the selection problem becomes especially challenging in a long-context scenario. In this paper, we tackle this problem by building on state space models (SSMs), which require only linear inference time given the input. Our approach involves two algorithms. The first learns a small set of SSMs through distillation of a (trained) transformer model. We partition all the layers into consecutive groups. Then for each group, we estimate a separate state space model to replicate the input-output behavior within the adjacent layers. Second, we map the distilled model outputs to a small set of tokens, and apply these embeddings for demonstration selection in downstream applications. We perform extensive experiments in both synthetic and real-world datasets to validate our approach. We demonstrate that the distilled SSMs only incur an approximation error of less than $0.7\%$ relative to the true output. In downstream evaluation, we show that on several text classification and reasoning tasks, our approach reduces FLOPs by $14.2\times$ and improves accuracy by $6.48\%$ relative to baseline demonstration selection methods.

Quantum machine learning offers a compelling representational perspective: angle-encoded states inhabit Hilbert spaces in which periodic similarities and interactions can be expressed naturally. Realizing this perspective for visual recognition remains difficult, however, because present quantum neural networks are constrained by limited qubit counts, costly circuit simulation and measurement, noise, and unstable optimization on noisy intermediate-scale quantum devices. We investigate whether useful structural ideas from quantum models can instead be realized as scalable classical Transformer operations. We introduce QiT, a Quantum-inspired Transformer for vision tasks with three components: (i) angle-inspired encoding that maps image tokens to learned trigonometric Hilbert-space features analogous to quantum rotation-based state encoding; (ii) self-attention over these periodic features, inducing a classical cosine kernel approximated to quantum fidelity kernels; and (iii) gated multiplicative emulation, a trainable classical surrogate for interaction terms found in variational circuits. All components are differentiable tensor operations, so QiT claims neither quantum computation nor quantum speedup and retains the $\mathcal{O}(N^2D)$ attention complexity of a standard Vision Transformer. Across image-classification benchmarks, QiT is competitive with a matched classical Transformer while avoiding the severe runtime cost observed for a small simulated quantum Transformer. QiT-B reaches 78.3% ImageNet-1K top-1 accuracy with 45.7M parameters and 11.5 GFLOPs. These results position QiT as a scalable baseline for isolating and evaluating quantum-motivated inductive biases in visual recognition.

Reconstructing the mass of a heavy resonance from its decay products with missing energy is one of the central tasks that directly determine the sensitivity in new physics searches at collider experiments. Supervised learning approaches to this problem often struggle to generalize well due to the presence of various systematic uncertainties and distribution shifts. Exhausting all possible variations in the labeled data can be very compute-intensive, while a failure of the model to generalize can corrupt the reconstructed resonance widths that are critical in peak-hunting analyses. In this work, following the foundation model paradigm, we use a self-supervised approach to pre-train a transformer encoder with VICReg to learn an embedding invariant to various corruptions, then fine-tune it for mass regression on a heavy resonance with masses ranging from 2.5 to 6.5 TeV and a SUSY-like cascade decay into an eleven-body final state. We show that the pre-trained model reconstructs sharper resonance peaks and has a more stable performance under various realistic corruptions, compared to a supervised model of the same architecture trained on the same augmented data from scratch.

We introduce a simple architectural modification to decoder-only transformers: a persistent recurrent state that observes hidden representations via cross-attention, updates itself through a GRU, and modulates subsequent processing via gated addition. Inserted between the lower and upper halves of a 6-layer transformer, this module adds only 3.7% additional parameters while reducing evaluation loss from $2.438 \pm 0.004$ to $1.743 \pm 0.018$, corresponding to a 28.5% reduction on held-out language modeling data. The improvement is statistically significant across 5 random seeds ($p < 0.01$) and corresponds to reduced overfitting (generalization gap 0.12 vs 0.26). Through controlled ablations, we demonstrate that the improvement stems entirely from the persistent memory topology, not from auxiliary self-prediction objectives. A model with identical topology but no auxiliary loss performs equivalently, while a random auxiliary loss provides no benefit. Representation probing reveals that the persistent state encodes narrative position (52% vs 33% chance level)---information that standard attention maintains less efficiently. Our results suggest that bridging transformer layers with a lightweight recurrent memory is a simple, effective approach to improving generalization in small-scale language models.

A common small-model deployment runs one shared backbone with several LoRA specialists that answer over the same context. Serving them naively re-prefills that shared context once per specialist. We study a narrow, practical question: for already-trained standard LoRA adapters -- not adapters retrained for cache compatibility -- how much task quality is preserved if the backbone's prefill KV cache is computed once and reused across specialists, and what does that buy in serving cost? On a Qwen3-1.7B backbone with two adapters (extractive QA on HotpotQA, arithmetic reasoning on GSM8K), we sweep the boundary at which the specialist takes over from the reused base cache and measure paired quality differences and serving cost. Full-prefix reuse had the lowest prefill cost and a small quality difference on held-out GSM8K (Delta = -4.6 EM at a 160-token budget; -3.0 at 320 tokens; -0.8 under a second training seed -- all favoring native, only the first excluding zero, and the magnitude not consistent). Partial recomputation provided no demonstrated advantage. Neither quality equivalence nor a general boundary-selection rule is established. We also report a closed-form ridge KV translator that did not beat direct reuse, and specialist-dependence contrasts whose intervals all include zero. The measured serving benefit is warm-cache time-to-first-token, which grows with context (~16x at 8K); two-branch peak memory was only 12% lower and, on inspection, the prefix was never physically shared across branches -- this implementation reuses KV values but copies their storage, so shared-cache memory savings are not achieved.

Query-based segmentation methods have shown promising potential for surgical instrument segmentation and recognition, which is essential for scene understanding and downstream tasks in computer assisted surgery. However, most existing approaches predominantly rely on per-frame predictions and overlook cross-frame temporal priors as well as temporal-consistency constraints. This limitation often leads to unstable query representations and suboptimal category recognition. In this paper, we propose TEDi, a Temporal memory-Enhanced and Denoising transformer for surgical instrument segmentation that addresses these is sues through Memory Search Enhancement and Temporal Consistency Denoising. The former introduces a query-level memory bank and a memory search enhancement encoder to retrieve discriminative representations from historical frames, enriching current-frame features. The latter constructs a temporally consistent reference as a cross-frame semantic anchor to suppress temporally unstable predictions and promote semantic coherence across frames. Extensive experiments on two benchmark datasets, EndoVis 2017 and EndoVis 2018, demonstrate that TEDi consistently outperforms state-of-the-art methods, highlighting its potential to further advance computer-assisted surgery. Our code is available at github.com/argon-xixi/TEDi.

Action Chunking Transformers (ACT) are widely used to learn robot manipulation from demonstrations. Their conditional variational autoencoder includes an encoder meant to capture differences between demonstrations during training. The original ACT paper reported that encoder removal dropped the mean success rate from 35% to 2% on two simulated tasks with human demonstrations. We re-ran this ablation in the original code and checked whether the findings depend on the implementation or training data. The published drop does not reappear in our tests, although smaller gains or losses in success rate remain uncertain. To investigate the discrepancy, we varied training length and how checkpoints are selected for evaluation. Both can reverse which policy scores higher, but the published drop's cause remains unknown. Success rates alone leave open whether the encoder provides information that helps the policy reconstruct demonstrated actions. On the tested ACT benchmark, the sampled latent provides little reconstruction benefit at every tested nonzero weight of the penalty on latent information. At inference, ACT leaves this latent unused and sets it to zero. Skipping the encoder increases training throughput in both implementations we timed. We release code, evaluation tools and results so others can repeat the comparisons and test the encoder on other tasks.

Answerability on Visual Question Answering is a novel and attractive task to predict answerable scores between images and questions in multi-modal data. Existing works often utilize a binary mapping from visual question answering systems into Answerability. It does not reflect the essence of this problem. Together with our consideration of Answerability in a regression task, we propose VT-Transformer, which exploits visual and textual features through Transformer architecture. Experimental results on VizWiz 2020 dataset show the effectiveness and robustness of VT-Transformer for Answerability on Visual Question Answering when comparing with competitive baselines.

Transformers and state-space models (SSMs) are the two dominant families of sequence models, and a central open question is how far the analytical knowledge built for transformers transfers to SSMs. We address this through the lens of layer importance which underpins compression, selective fine-tuning, and interpretability across both families. We decompose layer importance into two distinct notions. \emph{Necessity} captures how much the pretrained model depends on a layer's existing contribution, measured by the loss increase from bypassing it. \emph{Plasticity} captures where the model absorbs new information during fine-tuning, measured by the magnitude of task-specific weight updates. Our analysis reveals that the two families behave fundamentally differently: in every evaluated residual transformer up to $14$B parameters, Necessity and Plasticity anti-align across depth, whereas in the evaluated Mamba-style SSMs they point to overlapping regions. The sign of this alignment also predicts downstream adaptation behavior. In the evaluated transformers, concentrating updates in the most plastic layers increases catastrophic forgetting, while this tier-dependent effect disappears in the evaluated Mamba-style SSMs.

Emergency vehicle detection in autonomous driving is a safety-critical perception task that demands robustness under diverse and adverse real-world conditions. Existing approaches rely on a single modality, either audio or video, which leads to systematic failure when that modality is degraded: microphone-based systems fail in noisy urban environments, and camera-based systems fail at night or under occlusion. This report presents AVNet, a multimodal audio-visual transformer that classifies emergency vehicles (ambulance, fire engine, police car) and road background using both audio and video, while gracefully handling the absence of either modality at inference time. AVNet introduces three key contributions: (1) a temporally aligned cross-modal fusion module that performs second-level cross-attention between audio spectrogram tokens and video frame tokens, exploiting their exact temporal correspondence without any learned alignment mechanism; (2) learned null embeddings that substitute for missing modality tokens, enabling a single unified model to operate in audio-only, video-only, or joint audio-visual mode without retraining; and (3) a knowledge distillation training strategy in which specialist unimodal teacher models transfer inter-class dark knowledge into the multimodal student fusion branch via soft probability targets. Evaluated on 281 clips from the Google AudioSet dataset, AVNet achieves 66.6% overall accuracy in audio-visual mode, outperforming the audio-only branch by +10.4% and the video-only branch by +15.0%. The largest per-class gain is observed for the hardest class, Ambulance, where fusion achieves +29.5% over either unimodal branch alone, demonstrating that the two modalities provide complementary information that the aligned cross attention mechanism successfully exploits.

Transformer representations evolve through learned additive transformations that either preserve their current direction or redirect it. We study this evolution as a functional geometry, decomposing learned updates into parallel and perpendicular components. Across pretrained models, we find substantial parallel components beyond the residual identity path. We then apply the decomposition in two spaces: to attention and MLP updates relative to the hidden state, and to attention value aggregation relative to the current token's value. Targeted edits reveal a strongly space-dependent asymmetry: exclude-self value-space parallel manipulation is markedly more robust than residual-space and perpendicular counterparts, preserving the direct self message while scaling only the non-self aggregate. The same decomposition gives a component-resolved description of compression-induced update error: perpendicular error separates compression methods more clearly than parallel error. Extensive experiments further demonstrate that full-aggregate parallel suppression during from-scratch pretraining lowers validation-loss trajectories and improves downstream averages, with the value-space variant strongest. Together, these results connect representation geometry to editing robustness, compression diagnosis, and training-time intervention. Code is available in the \href{https://github.com/Shwai-He/Transformer-Geometry}{project repository}.

Diffusion Transformers deliver state-of-the-art video generation, but their long spatiotemporal sequences make attention the dominant deployment cost, and a deployable low-bit kernel must be accurate and fast. Accuracy is limited by outliers: a block's quantization scale is set by its largest entries, leaving typical entries confined to a narrow range of representable values. Prior work smooths queries and keys, but value outliers follow no fixed channel or spatiotemporal structure and remain the dominant source of output error. Speed is limited by softmax: low-bit Tensor Cores accelerate only the two matrix multiplications, so the high-precision exponential between them becomes the longest pipeline stage on datacenter GPUs. We propose VC-Attention, a training-free low-bit attention framework that addresses both by pairing Value smoothing with a fused probability Cast. V-Smooth reorders value tokens by lightweight online clustering, so the tokens in a hardware block quantize well together. It quantizes only the residual after subtracting the block mean, and restores that mean from the row sum the online softmax already maintains. ExpCast-FP8 maps log-domain scores directly to E4M3 probability codes with one fused multiply-add, eliminating the FP32 exponential and the format conversion. We implement VC-Attention for B200, B300, H200, RTX PRO 6000, and RTX 5090. Across Wan2.2, LongCat-Video, HunyuanVideo-1.5, and MiniMax-H3, VC-Attention improves fidelity over low-bit baselines, speeds up the attention kernel over BF16 FlashAttention-4 by 1.46-1.59x on datacenter Blackwell and Hopper and by 2.3-3.6x on workstation cards, and generates a clip 1.13-1.19x and 1.36-1.70x faster end to end.

Recurrent Transformers reusing their weights rather than stacking $L$ distinct layers are becoming widely adopted due to their parameter efficiency [1,2,3]. However, the exact representational and dynamical differences between looped and stacked architectures remain uncharacterized. This paper presents a controlled study on the example of bViT model [1] applying one weight-tied block $L$ times. We train two models: bViT and standard ViT [4] on 12-lead electrocardiogram (ECG) classification tasks from the PTB-XL dataset under identical training protocols. Despite an $8.9\times$ parameter reduction, bViT achieves accuracy parity with ViT. Geometric similarity metrics demonstrate that both architectures construct comparable latent representations in an equivalent canonical order. Crucially, their dynamics differ: bViT exhibits smaller step sizes and inter-patient sensitivity, as well as near-neutral behavior away from the data manifold, whereas ViT exhibits collapsing dimensionality of representations and out-of-distribution feature expansion.

Tokenization determines both the physiological content presented to an ECG Transformer and the sequence over which attention operates. We compare eight tokenization strategies across Transformer, Informer, Reformer, and FEDformer on the nine-label CPSC2018 classification task. The input projection and principal backbone capacity are controlled to isolate the effect of token construction. Median-beat and HeartLang tokenization achieve mean macro-AUCs of 0.893 and 0.889 across the four backbones, compared with 0.822 and 0.824 for point-wise and patch-wise tokenization. Pooling the two physiology-aware representations yields an 8.2% relative improvement in macro-AUC. They also reduce mean sequence length from 1,250 to 158 tokens and mean peak training memory from 5.21 to 0.27 GB. The results show that aligning tokens with ECG morphology can improve both predictive performance and memory efficiency without increasing backbone capacity. The source code is available on https://github.com/LeeJarvis996/ecg_tokenizer.

We propose a pre-fine-tuning probing method for Parameter-Efficient Fine-Tuning (PEFT) layer selection, aiming to obtain more stable and higher gains with fewer trainable parameters when adapting large vision--language models (VLMs). Unlike the common practice of applying LoRA and other adapters to all layers at once---where layer selection often relies on heuristic rules---we focus on the vision encoder and directly evaluate the "adaptability'' of each Transformer layer. Specifically, we characterize each layer from two perspectives: (i) the statistical properties of its Q/K/V projection weights (e.g., norms and condition numbers); (ii) robustness under controlled parameter perturbations. We then systematically compare these indicators with the downstream performance gains brought by applying PEFT to a single layer. Across experiments covering seven benchmarks and five PEFT variants, we observe a consistent correlation: layers (or matrices) with larger weight norms and higher condition numbers are usually more robust to perturbations and are more likely to yield larger fine-tuning gains. These results show that distribution-statistics analysis and perturbation tests before fine-tuning can provide practical signals for adaptation-layer selection, thereby maintaining or improving performance while reducing trainable parameters.

Looped Transformers have recently demonstrated strong performance in both reasoning and language tasks by reusing a shared set of parameters across multiple iterations, achieving parameter efficiency without sacrificing representational power. Besides, looped Transformers perform inference directly in the latent space (latent reasoning) to reduce the number of tokens consumed during inference, thereby achieving improved sample efficiency. However, these models typically apply a fixed recursion depth uniformly to every token, leading to suboptimal compute allocation and leaving significant efficiency gains on the table. In this work, we propose \textbf{dynamic token-choice routing} for looped transformers, enabling each token to adaptively determine its own number of loop iterations based on its hidden state. We use a dynamic router to decide whether a token should continue recursing or exit early, allowing simple tokens to bypass unnecessary computation while hard tokens receive deeper processing. To ensure that this adaptive mechanism does not compromise decoding efficiency, we further introduce recursion-wise KV caching, which maintains an independent key-value cache for each recursion loop. This design ensures that tokens at different depths only attend to their corresponding cached states, effectively eliminating redundant computations for exited tokens and enabling fast autoregressive decoding. Extensive experiments show that T-LoopFormer reaches the sota performance under the same parameters on PPL and 10 zero-shot reasoning tasks, even surpassing the base model at 24x FLOPs and our model could reach the lowest inference latency, which validate the effectiveness of token-choice router and recursion-wise KV cache. Code: https://github.com/YuMingQian1234/T-LoopFormer.

Diffusion Language Models (DLMs) generate text by iteratively denoising a masked sequence, independently predicting multiple tokens at each step. This conditional independence discards inter-token dependencies and degrades coherence-an issue that parallels the multi-modality problem in Non-Autoregressive Translation (NAT). Drawing on the Directed Acyclic Transformer (DAT), which tackles this problem in NAT via a Directed Acyclic Graph (DAG), we propose DA-DLM, a model that adapts DAG-based dependency modeling to DLMs' iterative setting through a position-oriented DAG design. The position-oriented DAG binds node groups to fixed output positions so that tokens fixed in earlier steps anchor neighboring predictions via learned transitions, and evolves with denoising to focus on remaining uncertainty as anchors accumulate. On language modeling, open-ended generation, and summarization, DA-DLM consistently outperforms Block Diffusion, especially under fewer denoising steps, and matches autoregressive models while preserving the parallel generation advantage. Our code is publicly available at https://github.com/jipy0222/DA-DLM.

Low-rank gradient projection reduces the optimizer-state memory cost of large language model (LLM) pretraining, but the steps and wall-clock time needed to reach a target quality remain a meaningful axis for improvement. We attribute this to two design choices in existing methods: the projection-rank budget is allocated uniformly across Transformer modules with heterogeneous projection sensitivity, and projecting a raw gradient attenuates its magnitude and direction jointly. We propose MoARa, which combines a static profiling-based module-aware projection-rank allocation with a block-wise magnitude-direction decomposition; the default block size is set in the neighborhood of the attention head dimension. Across five Transformer architectures spanning Llama, Qwen, and DeepSeek at 300M to 7B scales, GaLore with MoARa reaches standard GaLore's final perplexity in 37% fewer steps and 34% less wall-clock time on Llama 2 7B, with only 0.2% peak reserved memory overhead under standard graph compilation. Across the six low-rank pretraining methods we evaluate, module-aware rank allocation alone delivers directionally consistent step reductions on all six. On compatible hosts, the two-component design reaches up to 41.7% step reduction and 37.1% wall-clock reduction.

Small models are made more capable through distillation from a larger one that shares their tokenization scheme. However, do distilled byte and token models behave similarly in terms of scaling trends as compute and data increases? To enable this comparison, we introduce two variants to efficiently convert token logits to Byte Logits: 1) approximate: Marginalize-It, and 2) exact: End-Of-Token. We then present the first large scale study of overtraining decoder-only dense transformer models varying two dimensions simultaneously: the tokenization scheme (Tokens, Bytes, Bytes w/ eot) and the training objective (Distillation vs. Cross-Entropy), sweeping layer-parameter-matched models with roughly 1 billion parameters up to 1 trillion bytes of data. Across eight benchmarks spanning three categories: Multiple Choice QA, Language Generation, and Machine Translation, we find that Token-1B models outperform byte models (End-Of-Token-1B and Bytes-1B) in the low-FLOP regime but eventually plateau; byte models start worse yet surpass Token-1B models with more compute, reaching a higher downstream task performance ceiling. Extrapolating the average top-1 error vs. validation BPB scaling laws predicts that, asymptotically, distilled End-Of-Token-1B outperforms distilled Token-1B by up to 4%. They are also far more data efficient, matching the performance of distilled Token-1B using only one-sixth of the training data. Moreover, by operating over a small vocabulary of 256 bytes instead of on the order of 100K tokens, they circumvent the need for top-k truncation during logit dumping, while also reducing logit storage costs to roughly one-fifth. Finally, our downstream performance scaling laws predict that our distilled End-Of-Token-1B models asymptotically surpass the Llama 3.2-1B, Gemma-3-1B-pt, and Gemma 2B models on averaged downstream tasks by up to 6.5%, 8.1%, and 2.1%, respectively.

With the recent rise and adoption of tabular foundation models, optimizing their inference performance becomes an emerging field for efficiency research. While the models are architecturally similar to transformer-based large language models (LLMs), the size and serving patterns differ significantly. We show that the focus should be on the attention calculation and less on weight or KV cache quantization, which are more popular in LLMs. We develop a quantization strategy for queries, keys, and values to FP8 and use explicit FP8 matrix multiplication instructions to speed up the attention calculation. We find that it is crucial to align the quantization error in the test rows with the quantization error in the training rows, as otherwise the accuracy drops drastically. Our Triton kernel achieves a speedup up to 1.7x over regular 16-bit kernels, and we show that on TabPFN-v3 and TabICLv2 there is no relevant accuracy loss across TabArena and BeyondArena.

Learning user-defined concepts as constraint networks has been extensively studied in the constraint acquisition (CA) literature. However, existing approaches typically rely on intensive interactions with a human oracle, making the learning process costly in terms of time and number of queries. In this paper, we propose a neuro-symbolic framework for automatic CA that significantly reduces user involvement by introducing neural Oracle Transformer models which learn to emulate user responses and to generalize conceptual knowledge. Trained on previously available examples, the learned oracle interacts with a dedicated CA engine, FastCA, which systematically refines the oracle's responses into a sound, consistent, and interpretable constraint network. This neuro-symbolic interaction enables the recovery of structured symbolic models from data without prior domain knowledge. Our results demonstrate that this neuro-symbolic interplay effectively aligns data-driven pattern recognition with symbolic reasoning, offering a robust approach to automating model construction in combinatorial domains.

Deep learning has been widely applied across many areas of experimental high-energy physics, yet existing models address only event-level classification or object tagging and therefore still require reconstruction algorithms tailored to each decay channel. We present the first application of segmentation to resonance tagging at BESIII and introduce ResoSeg, a deep learning model that jointly performs particle-level segmentation and event-level classification, enabling a one-pass analysis of resonance to anything decays while precisely reconstructing the relevant resonance properties. We demonstrate the reconstruction of $\eta_c$ with $e^+e^-\to\pi^+\pi^-h_c$, $h_c\to\gamma\eta_c$, $\eta_c\to\text{anything}$. The model is trained on BESIII-$\eta_c$ dataset, which is constructed with per-track true labels obtained via a Truth-Matching Algorithm. Experimental results show that the average combined efficiency of ResoSeg is more than double that of the conventional 16-channel approach across energy points from 4.19 to 4.60,GeV. The model generalizes to unseen energy points, adapts to other $\eta_c$ production modes through transfer learning, and remains robust against variations in the $\eta_c$ mass, width, and branching fractions, providing a general, resonance-aware model applicable beyond $\eta_c$ and BESIII. The source code is available at https://github.com/oashen/ResoSeg.

Post-training attention sparsification reduces the quadratic cumulative attention cost of pretrained Transformers by selecting a small set of context units (tokens or blocks) for each query. Existing trainable methods usually use a lightweight selector to score context units, followed by hard Top-K selection that blocks gradients from the language modeling loss. Consequently, these methods commonly distill layer-wise dense attention distributions. Although this encourages the selector to rank context units by dense attention weights in the original model, the ranking is not directly aligned with their impact on predictions under a fixed attention budget (i.e., the number of attended context units per query), potentially wasting the limited budget on less useful units. To address this misalignment, we propose Simple Attention Sparsification (SAS), a gated sparse attention mechanism that optimizes context ranking end-to-end with the language modeling loss. The key idea is to inject the selector's continuous scores into attention logits during training, allowing the loss to update the selector through standard backpropagation. We identify several choices crucial for this simple design to work well in practice: placing the gate inside the attention softmax in log form, using normalized softmax gates to calibrate historical context against the always-retained current block, and preserving continuous selector scores so the model learns relative priorities rather than only hard selections. To support long-sequence training, we implement a memory-efficient Triton kernel that integrates SAS into FlashAttention-style computation. Across reasoning, long-context understanding, and agentic tasks, SAS consistently outperforms trainable sparse attention baselines across attention budgets, with especially large gains under tight budgets, demonstrating more effective context ranking for downstream tasks.

Compositional generalisation has been divided into lexical and structural generalisation. Previous work has found that structural generalisation is harder than lexical for Transformers. We propose that this difference is not inherent to Transformers, but due to the high diversity of lexical types and low diversity of structural types in the specific datasets of these previous works. By type diversity we mean the number of different constructors of that type, instead of, for example, the specific word combinations that might populate the structure. To test this, we vary the amounts of type diversity of lexical and structural types in previously published datasets. We create linguistically diverse variants of the COGS and SLOG datasets using Grammatical Framework. We find that type diversity correlates with compositional generalisation equally in lexical and structural test cases, supporting our hypothesis. We note a contradiction with the proposition in previous work that compound divergence explains the difficulty in compositional generalisation tasks. We further investigate the effects of other dataset properties on compositional generalisation, such as the diversity of types other than the novel test structure, and surface properties of the logical semantics format.

Generative recommendation (GR) relies on large-beam decoding to generate hundreds of candidate items, creating a new scaling challenge for recurrent linear attention. Existing linear attention serving systems either materialize a full recurrent state for every beam or repeatedly replay shared history, incurring substantial memory and traffic overhead. To address this, we present OneLA, a linear-attention decoding framework that exploits the shared prompt and short divergent suffixes of GR workloads. Specifically, OneLA represents all beam states using a single shared prompt-derived state and compact, append-only records of their divergent transitions. Using this representation, OneLA computes only the state information required at each decoding step, without reconstructing a full recurrent state for every beam. Furthermore, OneLA uses a lightweight ancestry index to track the transition records that make up each beam's history, allowing beams to be updated without moving or copying existing records. A fused GPU kernel further reuses the shared state across beams. Our analysis shows that OneLA achieves 1.54-2.46x end-to-end decode speedups while substantially reducing recurrent-state memory use and data movement.

Linear attention and state-space models provide linear-time sequence modeling, but their recurrent memory remains a second-order tensor (a matrix), limiting the order of interactions that can be represented in the state. We introduce the RunningTensor, which generalizes this memory to an order-$o$ tensor, updated by a rank-1 outer product and read by contracting against $o-1$ vector queries. Order $2$ recovers linear attention; we study order $3$ as a proof of concept, retaining both recurrent and parallel forms while remaining linear in sequence length $T$ and improving working memory capacity from $\mathcal{O}(W^2)$ to $\mathcal{O}(W^o)$. On synthetic multi-query associative recall, RunningTensor outperforms linear-attention and SSM baselines. After pretraining, it also improves performance on language-understanding and non-synthetic retrieval tasks, suggesting that higher-order recurrent state can provide useful additional memory capacity beyond matrix-valued state.

Frontier language models are more aggressively using subquadratic attention to reduce the memory footprint and compute requirements during inference while still delivering frontier accuracy. While existing systems make dense attention-centric disaggregated serving decisions, we show that disaggregating inference around the unique arithmetic intensity and memory footprint of subquadratic attention LLMs can achieve significant throughput and energy efficiency gains on emerging DRAM-based and SRAM-only heterogeneous systems. We introduce SQD (SubQuadratic Disaggregation), a fine-grained heterogeneous disaggregation scheme that splits decode by quadratic and subquadratic attention rather than by operator type, and that applies across subquadratic attention variants. For sparse attention LLMs, we disaggregate decode into top-k selection, which must index through the full KV, and top-k attention plus FFN, which have static memory footprints. For linear and sliding-window attention LLMs, we disaggregate decode into dense attention layers and subquadratic attention layers plus FFN. In an adjusted 8xB200 heterogeneous system proxy, we observe average tokens/J improvements of 53% on GLM 5.2, 31% on Nemotron 3 Ultra, and 56% on Gemma 4 31B over the strongest GPU-only baselines. In an analytical model of a Rubin plus LPX system with fixed power budgets, we observe 1.2x to 1.5x tighter achievable latencies and up to 3.6x higher throughput over the best baseline of attention-FFN disaggregation. Our experiments also reveal architectural insights on chip and interconnect provisioning for next-generation heterogeneous systems serving subquadratic attention.

Text-to-image diffusion transformers (DiTs) are powerful generators, yet direct prompting provides limited control interface for style intensity and can fail to suppress unwanted concepts. To enable these controls, we introduce AcFlow, an inference-time controller that transports intermediate layer image-token activations through a learned concept-conditioned velocity field while keeping the base DiT frozen. A textual concept description specifies the desired intervention, while the integration horizon provides a continuous control parameter. The field produces token-varying, activation-dependent updates. With parameters shared across concepts within each task family, the field supports fine-grained descriptions and generalizes to concepts unseen during training without per-concept fitting. On style control, AcFlow achieves the best style--content trade-off among the evaluated baselines in the high-style-alignment regime. At a fixed operating point, AcFlow attains style--content alignment of 0.5365/0.2860, compared with 0.4397/0.2684 for the baseline with the highest style alignment. Qualitative results demonstrate suppression of diverse concepts, including cases where direct prompting fails. Our analyses support the learned velocity field as an adaptive control mechanism, with update directions varying across tokens and depend on their activation states. Our code is available at https://github.com/Nove1yst/AcFlow.

Wireless Capsule Endoscopy (WCE) enables non-invasive visualization of the gastrointestinal tract, but its miniaturized optics, sensor limitations, and wireless transmission constraints result in low-resolution images with reduced visibility of diagnostically important structures. This paper proposes CEM-TUDASR, a computationally efficient unsupervised Transformer-based super-resolution framework for WCE image enhancement without paired low-resolution (LR) and high-resolution (HR) training data. A domain-adaptive degradation network synthesizes realistic WCE-like LR images from HR conventional endoscopy images, reducing the domain gap and enabling effective unpaired learning. The SR generator integrates Deep Attention Blocks (DABs) and a Fusion Attention Block (FAB) to capture long-range contextual dependencies and fine local structures while preserving perceptual and structural fidelity. The model is trained on a curated dataset derived from Kvasir Capsule and evaluated on KID and GIANA for cross-dataset generalization. No-reference quality metrics, including BRISQUE, PIQE, NIQE, and the domain-specific EndoQM, show that CEM-TUDASR consistently outperforms existing unsupervised SR methods. Qualitative results further demonstrate improved restoration of mucosal textures, vascular patterns, and clinically relevant anatomical details. Cross-domain experiments on retinal images additionally demonstrate the adaptability of the framework. With only 2.67 million parameters and 169.94 GFLOPs, CEM-TUDASR achieves high-quality reconstruction while maintaining computational efficiency, making it suitable for resource-constrained clinical and embedded endoscopic applications.

Out-of-distribution length generalization, namely to extrapolate a task from short to longer context, has been studied intensively for transformers. Here we focus on distance generalization, which probes performance when inter-token distances are changed between training and inference, while keeping a fixed context length. We construct two synthetic delay copy tasks, both involving finite distances between source and recall, where tokens are copied either fully or selectively, and test models on delays unseen during training. We address three questions: (A) Do positional encoding schemes such as RoPE and ALiBi improve distance resolution relative to no positional encoding (NoPE)? (B) How does data diversity, the number of inter-token distances seen in training, affect performance? (C) When is distance transfer learning positive or negative? We present a thorough investigation, finding that it is paramount to improve our understanding of the underlying mechanisms.

Transformer-based foundation models are powerful but opaque, motivating Mechanistic Interpretation methods to uncover the black-box by identifying small computation subgraphs responsible for a task. DiscoGP is a joint weight-and-edge circuit discovery framework originally developed for text decoders. We extend DiscoGP to speech encoders and present, to our knowledge, the first circuit discovery study for modern speech foundation models. Across HuBERT and Wav2Vec 2.0 on several speech classification tasks, we find that the discovered circuits are extremely compact, yet often match or even exceed the performance of the full pretrained encoder with the same downstream head. Through ablations, we show that these circuits reflect pretrained computation rather than random structure or task-head artifacts. We also introduce a memory-efficient DiscoGP variant that reduces the GPU memory cost of edge-circuit discovery at runtime from quartic to cubic. Overall, our results broaden Mechanistic Interpretation beyond text decoders and show that circuit-level analysis can reveal both explanatory structure and unexpected functional behavior in speech encoders.

Large language models (LLMs) have demonstrated impressive performance in various natural language processing tasks, yet their ability to perform multi-step logical reasoning remains an open challenge. Although Chain-of-Thought prompting has improved logical reasoning by enabling models to generate intermediate steps, it lacks mechanisms to assess the coherence of these logical transitions. In this paper, we propose a novel, lightweight evaluation strategy for logical reasoning that uses query-key alignments inside transformer attention heads. By computing a single forward pass and extracting a "QK-score" from carefully chosen heads, our method reveals latent representations that reliably separate valid from invalid inferences, offering a scalable alternative to traditional ablation-based techniques. We also provide an empirical validation on multiple logical reasoning benchmarks, demonstrating improved robustness of our evaluation method against distractors and increased reasoning depth. The experiments were conducted on a diverse set of models, ranging from 1.5B to 70B parameters.

Accurate plant growth estimation is essential for greenhouse monitoring, yet obtaining labeled data remains costly and time-consuming. To address this, we propose a few-shot regression framework that combines Vision Transformer (ViT) feature embeddings, clustering-based task construction, and gradient-based meta-learning, and show that task construction in embedding space is a primary driver of performance. The approach leverages an unlabeled image pool to organize data into structured tasks using fuzzy c-means clustering, enabling efficient learning from a small number of labeled samples. We systematically evaluate meta-learning methods and show that second-order methods (e.g., Model-Agnostic Meta-Learning variants such as MAML++) outperform classical baselines in the few-shot regime. Furthermore, intra-cluster support selection has a limited and dataset-dependent impact. Experiments on two plant datasets show that structured task design combined with meta-learning enables reliable plant growth estimation under severe label scarcity.

Feature transformation improves predictive performance on tabular data by constructing informative abstractions from raw features. Recent generative approaches encode transformation knowledge into continuous embedding spaces for efficient exploration of candidate strategies, but face three key limitations: (1) overlooking hierarchical relationships between low-level features, operations, and high-level abstractions; (2) enforcing order-sensitive embeddings on inherently permutation-invariant transformation sequences, thereby introducing systematic bias; and (3) relying on gradient-based search, which is ill-suited to non-convex transformation spaces. We propose a framework with two complementary components. First, a permutation-invariant hierarchical module captures interactions across features, operations, and abstraction levels, with a self-attention pooling mechanism that maps semantically equivalent structures to consistent embeddings aligned with downstream performance. Second, a policy-guided multi-objective reinforcement learning strategy initializes the search from empirically strong seeds and jointly optimizes predictive accuracy and transformation efficiency. Extensive experiments on diverse tabular benchmarks demonstrate the effectiveness and robustness of our framework against strong baselines. Our code and data are publicly available at: https://github.com/RayLiu1103/PHER.

How many of a transformer's components decide a token? Counted by the absolute value of each unit's and channel's contribution to the logit, one prediction rests on thousands to hundreds of thousands of them. But contributions are signed, and across eighteen models the mass pushing away from the predicted token is a median of seven times the mass carrying it. Divide by the net and the count is dozens: on the baseline, 53 components carry ninety percent of a prediction, 13 it cannot survive losing, and 8 suffice to produce it alone. Across twelve models trained elsewhere, 124M to 7B parameters, the sufficient set runs from two components to sixteen, and what a prediction draws on, followed all the way back, is one to three percent of the model, a share that does not grow with size. Three quarters of a layer's update is a fixed linear map of the state it received. Everything is read from the model's own parameters and activations, with nothing trained or fitted, and it names a component on both sides: what it writes, from the predictions it drives, reaching close to half of every model; what it reads, from its weights in the frame of its own layer, at 58.9 percent above chance over its eight strongest inputs. Sorting the remainder by upstream source yields grammatical categories the embedding cannot see. A name can be acted on. An association the model does not hold installs into one spare unit, key and value read from the weights, for a quarter of a percent of held-out loss, a fortieth of what a rank-one update costs. An installed attention head and a unit two layers above it make an edit fire only where a token occurred earlier in the context, and a unit the model trained for itself is driven from two layers upstream, 86 percent of the effect passing through it. An order-preserving activation puts a unit's inputs at the instrument's ceiling, at the price of a two-part install.

Driver gaze provides information regarding driver visual attention and situational awareness to the surrounding traffic. Existing driver gaze estimation studies represent gaze in terms of gaze zone or gaze vector/point-of-gaze (PoG). However, object-level gaze information provides a more semantically meaningful representation of visual attention by identifying attended objects, such as vehicles, pedestrians, or traffic signals. In this study, we propose an end-to-end driver gaze object prediction framework, TransGaze-Object, Transformer-based Gaze Object prediction model. The proposed framework first extracts facial features, including face and iris-weighted eye features, along with trafficobject spatial features. A transformer based cross-attention mechanism is then used to compute similarity scores and attention weights for predicting the drivers gaze object. To train this model, we propose a benchmark driver gaze dataset, Urban Driving-Face Scene Gaze (UD-FSG), comprising synchronized driver-face and traffic-scene images, scene objects bounding boxes, and gaze labels in terms of 2D gaze coordinate and gaze object. The TransGaze-Object model achieves an overall accuracy of 60% for gaze-object prediction, compared to 51% accuracy obtained from associating the estimated Point-of-Gaze to traffic objects. The error analysis reveals that TransGaze-Object reduces confusion between traffic objects (predicted) and the background (ground-truth), achieving an error rate of 11.68%, a 49.7% relative reduction compared with 23.21% error obtained from PoG-based gaze-object association. Overall, the results demonstrate the effectiveness of directly predicting gaze objects from driver-face and traffic-scene information, rather than estimating an intermediate Point-of-Gaze and subsequently associating it with traffic objects.

Large language models (LLMs) have achieved strong performance across a broad range of classification settings, yet the reliability of their predictions remains a major obstacle to deployment in high-stakes scenarios. Although confidence estimation for LLMs has been widely studied, confidence calibration for LLM-based classification remains underexplored. We introduce ProbPlug, a lightweight confidence estimation framework for LLM-based binary classification, which predicts whether an output is correct using internal token features extracted from a frozen LLM. ProbPlug employs a self-attention module to aggregate hidden representations and can be integrated into the original inference pipeline without modifying the base model. Experiments across multiple tasks involving both text-based and multimodal large models show that ProbPlug provides more reliable confidence estimates, improves classification performance with negligible additional overhead, and exhibits strong generalization across tasks. These results indicate that ProbPlug serves as a practical solution for confidence estimation in LLM-based classification. Our code is publicly available at Github.

Reading a transformer's internal states in token space is easy to do and hard to trust: a logit lens on a single hidden state is dominated, at intermediate layers, by the generic tokens the model would predict for almost any input. We read the difference instead. Subtracting two closely matched prompts' hidden states and projecting through the unembedding cancels the shared component and surfaces what separates them, an operation equivalent to reading a RepE/ActAdd steering vector through a logit lens. Built into a training-free tracer that reads at every position, sub-layer, and head and averages over designed baselines, it traces a compound- noun MLP->attention chain in Phi-2, confirmed there by activation patching, with the same distinction recovered across three architectures by readout and probe rather than by patching; it reads what retrieval surfaces for real versus fictional entities, and reads metaphor as a set of domain-to-domain mappings rather than a single figurativity feature. A cross-seed control marks the boundary: across five networks differing only in initialization, the same distinction surfaces as almost entirely different tokens (top-10 overlap 0.08). What a computation looks like in token space is network-specific; the distinction it draws is not

This paper studies unmanned aerial vehicle (UAV)-mouted reconfigurable intelligent surface (RIS)-assisted device-to-device (D2D) communication with stochastic link activation. It models UAV motion and attitude, time-varying Rician angles, and angle-dependent RIS reflection. A joint optimization of UAV trajectory, attitude, and RIS phases is formulated to maximize average sum rate under mobility, energy, and hardware constraints. The problem is addressed using deep reinforcement learning and a Decision Transformer trained on expert trajectories from multiple scenarios. Results demonstrate effective cross-scenario generalization, with zero-shot transfer outperforming direct DRL transfer and online fine-tuning achieving competitive performance with fewer interactions.