Inference Engine

Fully homomorphic encryption (FHE) changes sequence-model design because rotations, encrypted products, ciphertext materialization, multiplicative depth, and bootstrapping pressure can dominate ordinary neural-network costs. This paper presents public-decay homomorphic state space models (HSSMs), recurrent/state-space blocks whose carried state is updated through ciphertext-plaintext public decay while ciphertext-ciphertext multiplication remains on a local write path. The design keeps a fixed encrypted state across the sequence. The evaluated workflow separates client-side tokenization, frozen fastText lookup, projection, clipping, encryption, decryption, and thresholding from server-side encrypted evaluation over bounded projected features. On full Rotten Tomatoes and SST-2 validation splits, the encrypted HSSM path exactly matches plaintext classifications and reaches 0.7505 and 0.7420 accuracy. Against HE-friendly polynomial attention on the same fastText workloads, HSSM matches or exceeds full-sequence task quality while running about 5x faster. Paired L40S operation-level rows show 1.34-1.62x lower latency than cached final-token polynomial attention, 30-258x lower latency than full-sequence polynomial attention, and lower logical encrypted-state footprint. A T = 16/32 comparator with encrypted public-linear input and Q/K/V projections shows projected HSSM succeeding under depth 8/ring 32768, while projected attention succeeds under depth 10/ring 65536. A matched T = 8 OpenFHE/FIDESlib trace finishes at final level 3 and noise-scale degree 2 on both backends. These results make public-decay carry a practical FHE co-design lever for encrypted sequence inference from bounded projected features.

Decoder-only transformers compute the conditional probability of the next token from a sequence of past observations. This paper derives, from first principles, inference architectures that solve the same prediction problem - and in doing so, recovers transformer-like layer operations as a consequence of optimal control theory. The framework is developed for two model classes: a nonlinear model of discrete-valued processes, directly motivated by the transformer, and a linear Gaussian model as a tractable baseline. For both model classes, the prediction objective is reformulated as an optimal control problem whose solution yields an explicit inference algorithm, the dual filter, with a layer structure that mirrors the layer structure of a decoder-only transformer. Numerical experiments provide a comparison of the optimal control to attention weights from a trained transformer. These experiments reveal that when the embedding dimension is insufficient, the transformer implicitly exploits non-Markovian structure.

Generative models such as diffusion and flow matching have become dominant paradigms for visuomotor policy learning, yet their reliance on iterative denoising incurs high inference latency incompatible with real-time robotic control. We present Fast Legendre-polynomial Action policy via Sparse History-anchored flow (FLASH Policy), which replaces discrete action-chunk generation with continuous Legendre polynomial trajectory representation. Specifically, by fitting expert demonstrations under sparse temporal sampling, FLASH enables a single inference to cover a significantly extended action horizon. To further accelerate generation, FLASH initiates the flow matching process from history polynomial coefficients rather than uninformative Gaussian noise, shortening the transport distance and enabling accurate single-step inference. Moreover, analytic polynomial differentiation directly provides desired velocity feed-forward signals to the torque controller without numerical approximation. Extensive experiments on five simulated and two real-world manipulation tasks demonstrate that FLASH achieves state-of-the-art success rates ($\ge 92\%$ across all tasks), a per-episode inference time of $31.40\,ms$ (up to $175\times$ faster than diffusion policies and $18\times$ faster than prior flow matching policies), up to $4\times$ faster training convergence than ACT, and $5\times$ to $7\times$ reduction in controller tracking error compared to discrete-action baselines.

Conventional transformer inference engines are request-driven, paying an O(n) prefill cost on every query. In streaming workloads, where data arrives continuously and queries probe an ever-growing context, this cost is prohibitive. We introduce a data-driven computational model centred on stateful sessions: a persistent KV cache advanced incrementally as new data arrives, so prefill is moved off the critical path and query latency becomes O(|q|), independent of accumulated context size. Building on this, Flash Queries reclaim idle GPU cycles between data arrivals to pre-evaluate registered questions and return cached answers before the user asks, a pattern that is structurally impossible in stateless engines because they discard intermediate state between requests. A multi-tenant continuous-batching scheduler with cell-budget admission and prefix-aware grouped prefill lets dozens of stateful sessions coexist on a single GPU while preserving full quadratic self-attention. On streaming market-data benchmarks the reference implementation achieves up to 5.9x speedup over conventional inference engines (vLLM, SGLang, TensorRT-LLM, llama.cpp), holding query latency constant as accumulated context grows.

Institutions may benefit from collaborative inference on time-series data. In settings where privacy is necessary, multi-party computation (MPC) is a straightforward approach to providing strong guarantees, yet it remains prohibitively expensive and scales poorly with modern transformer architectures. Vertical Federated Learning (VFL) offers efficiency but suffers from privacy leakage at the embedding level, and securing the entire VFL model head via MPC remains prohibitively slow and communication-heavy for larger models. To enable practical, secure inference at scale, we propose "Public/Private Hybrid Head-VFL" (PPHH-VFL). This hybrid architecture splits the model head into an efficient plaintext public head and a secure, lightweight MPC private head. By applying adversarial training to the public embeddings, we mitigate privacy leakage; concurrently, the small private head securely preserves the flow of sensitive information needed for high downstream utility. Empirical evaluations on models ranging up to 86 million parameters demonstrate that PPHH-VFL accelerates inference by up to six orders of magnitude compared to end-to-end MPC. Compared to a standard VFL+MPC baseline, our approach scales significantly better, achieving a speedup of up to 44.4x in WAN and a 91.2x reduction in communication costs (dropping from 1.7 GB to 19 MB per batch), while simultaneously improving downstream classification accuracy by 2.50% and regression RMSE by 40.7%.

Constructive neural routing solvers usually score the next action by matching a decoder context to candidate embeddings, hiding deterministic one-step consequences such as travel, waiting, slack, and capacity changes. We propose LINC (Local Inference via Normed Comparison), a decoder-side candidate decision architecture that computes these consequences explicitly. LINC uses them according to their decision role: centered relative consequences are compared by a shared linear local scorer, while feasible-set summaries modulate the decoder context. This preserves standard global matching and relieves the hidden state from rediscovering transition arithmetic. The Capacitated Vehicle Routing Problem with Time Windows (CVRPTW) serves as the main constrained-routing stress test; the same interface extends to the Capacitated Vehicle Routing Problem (CVRP) and Traveling Salesman Problem (TSP). In particular, for CVRPTW, LINC reduces PolyNet's Solomon/Homberger gaps from 13.83%/38.15% to 7.26%/14.71%; for TSP and CVRP, it also improves external-benchmark gaps.

For Transformer models, cryptographically secure inference ensures that the client learns only the final output, while the server learns nothing about the client's input. However, securely computing nonlinear layers remains a major efficiency bottleneck due to the substantial communication rounds and data transmission required. To address this issue, prior works reveal intermediate activations to the client, allowing nonlinear operations to be computed in plaintext. Although this approach significantly improves efficiency, exposing activations enables adversaries to extract model weights. To mitigate this risk, existing works employ a shuffling defense that reveals only randomly permuted activations to the client. In this work, we show that the shuffling defense is not as robust as previously claimed. We propose an attack that aligns differently shuffled activations to a common permutation and subsequently exploits them to extract model weights. Experiments on Pythia-70m and GPT-2 demonstrate that the proposed attack can align shuffled activations with mean squared errors ranging from $10^{-9}$ to $10^{-6}$. With a query cost of approximately $1, the adversary can recover model weights with L1-norm differences ranging from $10^{-4}$ to $10^{-2}$ compared to the oracle weights.

Large language models are increasingly used as surrogate models for low-data optimization, but their optimizer-facing prediction and its uncertainty remain poorly understood. We study the surrogate belief elicited from an LLM under sparse observations, showing that it depends strongly on prompt text and query protocol. We introduce an uncertainty-alignment criterion that measures whether model uncertainty tracks residual ambiguity among sample-consistent functions. Across controlled inference tasks and Bayesian optimization studies, we find that structural prompts act as effective priors, POINTWISE and JOINT querying induce different beliefs, and sequential evidence leads to non-monotonic, order-sensitive confidence updates. These effects change downstream acquisition decisions and regret, showing that elicitation protocol is part of the LLM surrogate specification, not a formatting detail.

Image editing instructions are heterogeneous: a color swap, an object insertion, and a physical-action edit all demand different spatial coverage and different reasoning depth, yet existing reasoning-based editors apply a single fixed inference recipe to every instruction. We argue that adaptivity along both the spatial and temporal axes is the missing degree of freedom, and we present PhysEdit, an editing framework built around this principle. PhysEdit introduces two inference-time modules that compose without retraining the backbone. At its core, (1) Complexity-Adaptive Reasoning Depth (CARD) predicts edit complexity directly from the instruction and reference image and allocates the reasoning step count N_r and reasoning-token length r per sample -- turning a previously fixed inference schedule into a conditional-computation problem. CARD is supported by (2) a Spatial Reasoning Mask (SRM) that extracts an instruction-conditioned spatial prior from cross-attention to confine reasoning to regions that semantically require it. On the full 737-case ImgEdit Basic-Edit Suite, PhysEdit delivers a 1.18x wall-clock speedup (64.3s vs. 76.1s per sample) over a strong reasoning baseline while slightly improving instruction adherence (CLIP-T 0.2283 vs. 0.2266, +0.7%) and matching identity preservation within noise (CLIP-I 0.8246 vs. 0.8280). The speedup is category-dependent and reaches 1.52x on appearance-level edits, validating CARD's adaptive allocation as the principal source of efficiency gain. A 30-sample pilot with full ablations isolates the contribution of each module.

Robotic ultrasound has advanced local image-driven control, contact regulation, and view optimization, yet current systems lack the anatomical understanding needed to determine what to scan, where to begin, and how to adapt to individual patient anatomy. These gaps make systems still reliant on expert intervention to initiate scanning. Here we present SAMe, a semantic anatomy mapping engine that provides robotic ultrasound with an explicit anatomical prior layer. SAMe addresses scan initiation as a target-to-anatomy-to-action process: it grounds under-specified clinical complaints into structured target organs, instantiates a patient-specific anatomical representation for the grounded targets from a single external body image, and translates this representation into control-facing 6-DoF probe initialization states without any additional registration using preoperative CT or MRI. The anatomical representation maintained by SAMe is explicit, lightweight (single-organ inference in 0.08s), and compatible with downstream control by design. Across semantic grounding, anatomical instantiation, and real-robot evaluation, SAMe shows strong performance across the full initialization pipeline. In real-robot experiments, centroid-based SAMe initialization outperformed the body-keypoint-based heuristic baseline under a budget-matched single-target setting for both liver (86.7% versus 46.7%) and kidney (80.0% versus 73.3%) initialization. Furthermore, The trial-level organ-hit rate reached 97.3% for liver and 83.3% for kidney when multiple candidate targets were available. These results establish an explicit anatomical prior layer that addresses scan initialization and is designed to support broader downstream autonomous scanning pipelines, providing the anatomical foundation for complaint-driven, anatomically informed robotic ultrasonography.

LLM-based multi-agent systems (MAS) show promise on complex tasks but remain prone to coordination failures such as goal drift, error cascades, and misaligned behaviors. We propose Explicit Trait Inference (ETI), a psychologically grounded method for improving coordination. ETI enables agents to infer and track partner characteristics along two established psychological dimensions--warmth (e.g., trust) and competence (e.g., skill)--from interaction histories to guide decisions. We evaluate ETI in controlled settings (economic games), where it reduces payoff loss by 45-77%, and in more realistic, complex multi-agent settings (MultiAgentBench), where it improves performance by 3-29% depending on the scenario and model, relative to a CoT baseline. Additional analysis shows that gains are closely linked to trait inference: ETI profiles predict agents' actions, and informative profiles drive improvements. These results highlight ETI as a lightweight and robust mechanism for improving coordination in diverse multi-agent settings, and provide the first systematic evidence that LLM agents can (i) reliably infer others' traits from interaction histories and (ii) leverage structured awareness of others' traits for coordination.

Deep learning-based object detectors have achieved remarkable success across numerous computer vision applications, yet they continue to struggle with small object detection in high-resolution aerial and satellite imagery, where dense object distributions, variable shooting angles, diminutive target sizes, and substantial inter-class variability pose formidable challenges. Existing slicing strategies that partition high-resolution images into manageable patches have demonstrated promising results for enlarging the effective receptive field of small targets; however, their reliance on fixed slice dimensions introduces significant redundant computation, inflating inference cost and undermining detection speed. In this paper, we propose \textbf{Adaptive Slicing-Assisted Hyper Inference (ASAHI)}, a novel slicing framework that shifts the paradigm from prescribing a fixed slice size to adaptively determining the optimal number of slices according to image resolution, thereby substantially mitigating redundant computation while preserving beneficial overlap between adjacent patches. ASAHI integrates three synergistic components: (1)an adaptive resolution-aware slicing algorithm that dynamically generates 6 or 12 overlapping patches based on a learned threshold, (2)a slicing-assisted fine-tuning (SAF) strategy that constructs augmented training data comprising both full-resolution and sliced image patches, and (3)a Cluster-DIoU-NMS (CDN) post-processing module that combines the geometric merging efficiency of Cluster-NMS with the center-distance-aware suppression of DIoU-NMS to achieve robust duplicate elimination in crowded scenes. Extensive experiments on VisDrone2019 and xView, demonstrate that ASAHI achieves state-of-the-art performance with 56.8% on VisDrone2019-DET-val and 22.7% on xView-test, while reducing inference time by 20-25% compared to the baseline SAHI method.

Deep learning surrogates for CFD flow-field prediction often rely on large, complex models, which can be slow and fragile when data are noisy or incomplete. We introduce FlowForge, a staged local rollout engine that predicts future flow fields by compiling a locality-preserving update schedule and executing it with a shared lightweight local predictor. Rather than producing the next frame in a single global pass, FlowForge rewrites spatial sites stage by stage so that each update conditions only on bounded local context exposed by earlier stages. This compile-execute design aligns inference with short-range physical dependence, keeps latency predictable, and limits error amplification from global mixing. Across PDEBench, CFDBench, and BubbleML, FlowForge matches or improves upon strong baselines in pointwise accuracy, delivers consistently better robustness to noise and missing observations, and maintains stable multi-step rollout behavior while reducing per-step latency.

Large language models (LLMs) offer strong capabilities but raise cost and privacy concerns, whereas small language models (SLMs) facilitate efficient and private local inference yet suffer from limited capacity. To synergize the complementary strengths, we introduce a dynamic collaboration framework, where an SLM learns to proactively decide how to request an LLM during multi-step reasoning, while the LLM provides adaptive feedback instead of acting as a passive tool. We further systematically investigate how collaboration strategies are shaped by SLM and LLM capabilities as well as efficiency and privacy constraints. Evaluation results reveal a distinct scaling effect: stronger SLMs become more self-reliant, while stronger LLMs enable fewer and more informative interactions. In addition, the learned dynamic collaboration strategies significantly outperform static pipelines and standalone inference, and transfer robustly to unseen LLMs.

We introduce EVIL (\textbf{EV}olving \textbf{I}nterpretable algorithms with \textbf{L}LMs), an approach that uses LLM-guided evolutionary search to discover simple, interpretable algorithms for dynamical systems inference. Rather than training neural networks on large datasets, EVIL evolves pure Python/NumPy programs that perform zero-shot, in-context inference across datasets. We apply EVIL to three distinct tasks: next-event prediction in temporal point processes, rate matrix estimation for Markov jump processes, and time series imputation. In each case, a single evolved algorithm generalizes across all evaluation datasets without per-dataset training (analogous to an amortized inference model). To the best of our knowledge, this is the first work to show that LLM-guided program evolution can discover a single compact inference function for these dynamical-systems problems. Across the three domains, the discovered algorithms are often competitive with, and even outperform, state-of-the-art deep learning models while being orders of magnitudes faster, and remaining fully interpretable.

Quantum kernel methods are among the leading candidates for achieving quantum advantage in supervised learning. A key bottleneck is the cost of inference: evaluating a trained model on new data requires estimating a weighted sum $\sum_{i=1}^N α_i k(x,x_i)$ of $N$ kernel values to additive precision $\varepsilon$, where $α$ is the vector of trained coefficients. The standard approach estimates each term independently via sampling, yielding a query complexity of $O(N\lVertα\rVert_2^2/\varepsilon^2)$. In this work, we identify two independent axes for improvement: (1) How individual kernel values are estimated (sampling versus quantum amplitude estimation), and (2) how the sum is approximated (term-by-term versus via a single observable), and systematically analyze all combinations thereof. The query-optimal combination, encoding the full inference sum as the expectation value of a single observable and applying quantum amplitude estimation, achieves a query complexity of $O(\lVertα\rVert_1/\varepsilon)$, removing the dependence on $N$ from the query count and yielding a quadratic improvement in both $\lVertα\rVert_1$ and $\varepsilon$. We prove a matching lower bound of $Ω(\lVertα\rVert_1/\varepsilon)$, establishing query-optimality of our approach up to logarithmic factors. Beyond query complexity, we also analyze how these improvements translate into gate costs and show that the query-optimal strategy is not always optimal in practice from the perspective of gate complexity. Our results provide both a query-optimal algorithm and a practically optimal choice of strategy depending on hardware capabilities, along with a complete landscape of intermediate methods to guide practitioners. All algorithms require only amplitude estimation as a subroutine and are thus natural candidates for early-fault-tolerant implementations.

High-resolution geometric prediction is essential for robust perception in autonomous driving, robotics, and AR/MR, but current foundation models are fundamentally limited by their scalability to real-world, high-resolution scenarios. Direct inference on 2K images with these models incurs prohibitive computational and memory demands, making practical deployment challenging. To tackle the issue, we present 2K Retrofit, a novel framework that enables efficient 2K-resolution inference for any geometric foundation model, without modifying or retraining the backbone. Our approach leverages fast coarse predictions and an entropy-based sparse refinement to selectively enhance high-uncertainty regions, achieving precise and high-fidelity 2K outputs with minimal overhead. Extensive experiments on widely used benchmark demonstrate that 2K Retrofit consistently achieves state-of-the-art accuracy and speed, bridging the gap between research advances and scalable deployment in high-resolution 3D vision applications. Code will be released upon acceptance.

Partial differential equation (PDE) models frequently contain unknown functional terms that cannot be measured directly, limiting their predictive utility. While data-driven methods for estimating scalar PDE parameters are well established, the recovery of unknown functions remains comparatively underexplored. Here, we show that standard parameter estimation workflows can be extended to infer functional components of PDEs directly from data. Our approach embeds neural networks within the PDE framework, allowing unknown functions to be learned during training with high accuracy. Using nonlocal aggregation-diffusion equations as a case study, we infer interaction kernels and external potentials from steady-state observations. We systematically examine how reconstruction accuracy depends on factors such as the number and diversity of available solutions, sampling density, and measurement noise. The resulting framework retains the advantages of conventional PDE calibration approaches while extending them to functional inference: once trained, the PDE model can be used in the standard way to analyse system behaviour and generate predictions.

Existing reinforcement learning methods for LLM reasoning implicitly assume that the policy generating training trajectories should coincide with the one producing inference responses. We argue that this is a misleading inductive bias: the optimization-optimal trajectory distribution favors informative gradients, whereas the inference-optimal response distribution emphasizes accuracy and consistency. Forcing both into a single policy entangles their gradients and suppresses exploration. We propose R$^2$PO (Residual Rollout Policy Optimization), which attaches a lightweight Residual Rollout-Head atop the policy to decouple training trajectories from inference responses, diversifying rollouts during training while keeping inference generation intact. Experiments show that R$^2$PO consistently outperforms baselines, with average accuracy gains of 3.4% on MATH-500 and 1.3% on APPS, alongside more diverse rollouts and reduced length bias. Our code is available at https://github.com/RRPO-ARR/Code.

Monocular 3D face reconstruction estimates a 3D morphable model (3DMM) representation from a single image, providing geometry-aware expression codes that are useful for facial expression analysis and affect understanding. Despite strong progress, most pipelines are trained with image-level self-supervision and evaluated primarily by geometric fidelity, which does not necessarily maximize the affective utility of the learned expression representation and may encourage intensity-amplifying shortcuts when affect supervision is naively coupled. We propose FIELDS (Face reconstruction with accurate Inference of Expression using Learning with Direct Supervision), a task-driven framework that learns FLAME expression codes for facial expression recognition (FER) under a geometric plausibility constraint. Using hybrid 2D/3D supervision, FIELDS improves affect prediction in both in-domain and external evaluations while maintaining competitive geometric fidelity on held-out and out-of-domain 3D benchmarks.

Large language model (LLM) steering has emerged as a promising paradigm for controlling model behavior at inference time through targeted manipulation of hidden states, offering a lightweight alternative to expensive retraining. However, existing steering frameworks suffer from critical limitations: computational inefficiency, limited extensibility, and restricted functionality that hinder both research progress and practical deployment. We present EasySteer, a unified framework for high-performance, extensible LLM steering built on vLLM. Our system features modular architecture with pluggable interfaces for both analysis-based and learning-based methods, fine-grained parameter control, pre-computed steering vectors for eight application domains, and an interactive demonstration system. Through deep integration with vLLM's optimized inference engine, EasySteer achieves 10.8-22.3$\times$ speedup over existing frameworks. Extensive experiments demonstrate its effectiveness in overthinking mitigation, hallucination reduction, and other key applications. EasySteer transforms steering from research technique to production-ready capability, establishing critical infrastructure for deployable, controllable language models.

Efficient scheduling of large language model (LLM) inference tasks is critical for achieving low latency and high throughput, a challenge that is becoming increasingly acute with the rise of reasoning-capable LLMs whose generation lengths are highly variable. Traditional strategies like First Come, First-Serve (FCFS) often suffer from Head-of-Line (HOL) blocking, where long-running tasks delay shorter ones queued behind them. In this paper, we introduce PARS, a prompt-aware LLM task scheduler that mitigates HOL blocking by approximating shortest-job-first (SJF) scheduling through pairwise ranking with a margin ranking loss. PARS effectively predicts response-length-based task ordering directly from prompts, thereby optimizing scheduling decisions with minimal overhead. In addition, it integrates seamlessly with vLLM, a state-of-the-art LLM serving system, for the research community. Extensive experiments across multiple LLM models and real-world inference use cases, including chat, math, and code generation, demonstrate that PARS significantly reduces latency by up to 15.7x compared to the vLLM default scheduler. Cross-model evaluations demonstrate that our design generalizes effectively, allowing effective scheduling across diverse LLMs without requiring model-specific retraining.

Efficiently scaling Large Language Models (LLMs) necessitates exploring alternatives to dominant autoregressive (AR) methods, with Masked Diffusion Models (MDMs) emerging as candidates. However, comparing AR (typically decoder-only) and MDM (often encoder-only) paradigms is confounded by differing architectures, obscuring true algorithmic and efficiency trade-offs. This research decouples these factors by evaluating MDMs within a decoder-only framework to: (1) Equitably compare MDM (as Any-Order AR) and standard AR paradigms through discrepancies on orders. (2) Investigate MDM architectural impacts on computational efficiency. We show decoder-only MDMs, despite a larger modeling space, can achieve significant inference speedups ($\sim25\times$) and comparable perplexity with techniques like temperature annealing, offering a path to reduced inference compute. This work provides insights for developing more computationally efficient foundation models by disentangling core modeling choices from architectural influences. Code is available at https://github.com/scxue/AO-GPT-MDM.

Modern language models (LMs) can learn to perform new tasks in different ways: in instruction following, the target task is described explicitly in natural language; in few-shot prompting, the task is specified implicitly with a small number of examples; in instruction inference, LMs are presented with in-context examples and are then prompted to generate a natural language task description before making predictions. Each of these procedures may be thought of as invoking a different form of reasoning: instruction following involves deductive reasoning, few-shot prompting involves inductive reasoning, and instruction inference involves abductive reasoning. How do these different capabilities relate? Across four LMs (from the gpt and llama families) and two learning problems (involving arithmetic functions and machine translation) we find a strong dissociation between the different types of reasoning: LMs can sometimes learn effectively from few-shot prompts even when they are unable to explain their own prediction rules; conversely, they sometimes infer useful task descriptions while completely failing to learn from human-generated descriptions of the same task. Our results highlight the non-systematic nature of reasoning even in some of today's largest LMs, and underscore the fact that very different learning mechanisms may be invoked by seemingly similar prompting procedures.

Longitudinal mammography screening has become an important source of information for improving future breast cancer risk prediction. However, the performance of current longitudinal mammography models degrades when prior examinations are unavailable at inference, creating a structured privileged-information setting in which temporal context is available during training but absent at deployment. We propose Single-Exam Mammography risk prediction with privileged History Distillation (SEM-HD), a framework that uses longitudinal history as privileged information available only during training to preserve the predictive benefits of longitudinal modeling while requiring only the current screening examination at deployment. During training, the student relies on the current examination to predict latent representations of prior visits, while horizon-specific teachers provide additional supervision from the observed longitudinal history. Together, latent history prediction and teacher distillation preserve the temporal modeling structure of longitudinal predictors under current-exam-only inference. We validate SEM-HD on three longitudinal mammography cohorts, the CSAW-CC, EMBED, and OMI-DB, using the transformer-based Longitudinal Mammography Risk (LoMaR) and recurrent Visual Memory Recurrent Attention (VMRA) backbones. Under current-exam-only inference, SEM-HD consistently improves long-horizon AUC and pAUC over longitudinal models evaluated without history, particularly in the clinically relevant low false-positive-rate region. It also recovers much of the performance gap with respect to full-history inference across datasets and backbones. Ablations further show that these gains are not reproduced by masking or heuristic history imputation. The strongest performance is achieved by combining patient-specific latent history prediction with distilled temporal risk supervision.