Byte-level language models such as the Byte Latent Transformer (BLT) group bytes into patches and run their large global model once per patch. BLT starts a patch where a small model's next-byte entropy is high. We show that under tight patch budgets this rule has a blind spot: positions whose type the preceding text fixes but whose value must be computed, such as the number after "= " in a worked math solution. A nearly certain digit keeps the next byte's entropy near or below log2(10) = 3.32 bits, about BLT-1B's threshold for a 10% budget. In Meta's BLT-1B with patch starts on 10% of bytes, the entropy rule puts a patch start at 16% of the computed results in GSM8K solutions and gets 19.0% of them exactly right given the true preceding text; a boundary after each "=" at the same patch count gets 51.8%, and entropy combined with a label-free boundary-dependence signal gets 67.1% (default layout at 26% of bytes: 76.8%). The gap survives adapting BLT-1B to the budget with low-rank adapters (32.9% vs 72.7%, three runs per rule). In byte models trained from scratch at a 10% budget on a 1.9 GB math corpus, it is not detectable at 1M parameters, appears on final answers at 12M and on both targets at 50M, as training data grow too: at 50M boundary dependence gets 35.9% of computed results against 13.9% for entropy and 78.9% of final answers against 59.0% (3 seeds each), although final answers mostly copy the solution's last number and entropy starts a patch at every one. In BLT-1B the blind spot narrows at larger budgets (entropy covers about 90% of the values right after "= " at 20% of bytes) and in the same solutions rewritten as prose, where a patch start at each value still pays. Boundary dependence, the change in the model's own loss when a patch start is removed or added, measured per two-byte context on text in the same format, finds these positions without labels.
Training-free sparse attention offers a practical acceleration solution to Diffusion Transformers (DiTs) via reducing computations without fine-tuning. It typically involves estimating the importance of query-key regions and deriving sparse masks to compute only the important candidates, which inevitably introduces approximation errors that may degrade generation quality. To better balance the efficiency-quality trade-off, we propose iCATS, integrating improved importance estimation and sparse mask construction with an efficient hardware execution strategy. Specifically, for importance estimation, unlike previous works that perform independent clustering over query and key tokens based on feature similarity to estimate attention scores, iCATS demonstrates that clustering based on query-key dot-product interactions is more accurate and further reformulates this objective as a simple quadratic form for low-cost computation. For sparse mask construction, instead of using a fixed top-p rule, we observe that tolerance to sparse approximation errors varies across denoising timesteps and therefore introduce an SNR-guided sparsity schedule to adjust sparsity dynamically, leading to higher accuracy. Finally, for hardware execution, we devise a tail-merging strategy to reduce padding overhead caused by irregular cluster sizes, improving GPU kernel utilization. Extensive experiments show that iCATS achieves $2.03\times$ acceleration with 31.017 dB PSNR on HunyuanVideo-T2V-13B and $1.55\times$ acceleration with 29.301 dB PSNR on Wan2.1-T2V-14B, delivering a state-of-the-art efficiency-quality trade-off.
Joint spatial-geometric and analytic function reasoning requires translating a perceived spatial configuration into a symbolic function whose executed curve satisfies geometric constraints. We present GAGR-Lab, a framework for measuring this capability through Cartesian game scenes, explicit function semantics, and authoritative Rust trajectory execution. It distinguishes spatial perception, metric grounding, geometric relations, function interpretation, function construction, and constrained synthesis. We specify four configurable scene-difficulty presets and a prospective 24-cell diagnostic design, while reporting only the subset actually evaluated. A bounded pilot of one hosted model (Llama 3.2 11B Vision Instruct) using two API credentials as execution replicas yields 72 balanced games with 432 attempts, 429 valid provider responses, and no target hits; exploratory ordinary-function prompt variants also fail to hit, while the structured localization interface yields no scoreable outputs. A privileged analytic search control independently succeeds on 600 directional cases from 300 generated scenes, with exact repeatability and 1,200 successful vertical-reflection or translation checks. The framework separates serving reliability, symbolic compliance, and geometric success, and preserves exact model-visible inputs and realized paths. A staged protocol outlines diagnostic calibration, held-out replication, multi-model comparison, and paired robustness tests. The contribution is an operational research framework with an executed pilot and a clearly identified prospective study plan; the full difficulty matrix and comparative model results remain untested.
Hessian spectra at trained models in deep learning exhibit a persistent pattern: eigenvalues organize into distinct clusters, including a large bulk near zero and a few isolated outliers. This paper shows that a natural account of these spectral phenomena emerges when the original setting is understood as a departure from a nearby, otherwise hidden, highly symmetric reference. Modifications, including changes to the architecture, data distribution, or parameter metric, expose a nearby reference configuration whose Hessian exhibits rich invariances-ones not accounted for by weight symmetries. There, symmetry enables a precise description of the spectra, forcing high-dimensional kernels and eigenvalues of large multiplicity. Returning to the original configuration breaks the Hessian symmetry and thereby produces the observed hierarchy of clusters and outliers. The framework is developed in some generality, with a detailed analysis of three-layer ReLU networks and applications to convolutional, graph, and transformer models, as well as to the NTK. The same mechanism is further shown to yield analogous spectral structures in layerwise Hessians and the Gauss-Newton matrix.
Text-to-image diffusion models fail predictably on compositional prompts: attributes bind to the wrong objects, spatial relations invert, and multi-object scenes lose count. Recent architectures already augment CLIP with a T5 encoder precisely because CLIP's contrastive embedding loses compositional structure, yet these failures persist. We argue the binding problem is therefore not one of missing information but of misaligned information: a text encoder preserves compositional structure, but in a representation space shaped by language modelling rather than vision, and the denoising objective does not directly reward aligning the two. We show this correspondence can be supplied as an explicit training signal, that the relevant cross-modal information is concentrated in a low-rank subspace of self-supervised visual features, and that supplying it can be folded into diffusion training as a single auxiliary loss. Our method, ORCA (Orthogonal Residual Compositional Alignment), aligns the latent of a diffusion transformer with a low-rank target derived from a frozen visual encoder, through a predictor whose orthogonal basis is parameterised by a learned residual between T5 and CLIP embeddings, which provides a prompt-dependent signal for selecting the visual readout subspace. We prove that the cross-modal information recoverable at a given rank is bounded by the spectral mass of the visual encoder's covariance in the top components. Across three diffusion-transformer backbones (DiT-B/2, DiT-L/2, U-ViT-L), ORCA improves FID and GenEval over both vanilla and REPA baselines at zero inference-time cost; on DiT-L/2 it reaches FID 16.65 and GenEval 0.291 at 200K steps, exceeding the strongest 400K baseline at half the training cost, with the largest gains concentrated on attribute binding, spatial relations, and multi-object prompts.
Many organizations fine-tune publicly available pretrained Automatic Speech Recognition (ASR) models and deploy them in black-box settings, assuming limited access provides protection. We show this assumption is fragile: adversarial perturbations crafted on the public base model transfer effectively to fine-tuned target models, severely degrading performance and posing concerns for safety-critical applications. We propose TransferBreaker, a unified fine-tuning framework that suppresses adversarial transfer by integrating Base Adversarial Fine-Tuning, which restricts adversarial training to base-effective perturbations; Latent Jacobian Regularization, which enforces latent-space invariance by suppressing adversarially sensitive directions; and HybridGrad-AFT, which improves robustness against adaptive attacks by interpolating transferable perturbations from base and target gradients. We theoretically justify all components and evaluate TransferBreaker across three languages and four large ASR models, reducing adversarial WER from 92.6 to 27.8. Our code is publicly available at https://github.com/rohban-lab/TransferBreaker.
Modern cache replacement designs saturate because they operate within fixed representational structures, hand-crafted and heuristic based feature-engineered predictors, or offline imitation models that cannot generate new decision logic on their own. At the same time, replacement is shaped by the causal interaction of prefetching, thrashing, spatial locality, and access-type behavior, producing an enormous design space that is difficult to traverse manually. Prior approaches typically rely on heuristics, parameter tuning, or imitation of an offline optimal policy, capturing correlations rather than synthesizing new mechanisms. As a result, their performance gains often plateau and they overfit under dynamic workload conditions. CACHEFORGE is the first framework to evolve cache-replacement policies end-to-end by embedding a large language model inside a governed hardware-aware loop. In each iteration, the LLM proposes new C++ replacement logic, the policy is evaluated under a trace-based CRC-2 ChampSim simulator, and the framework enforces feasibility through reward shaping, structural checks, dynamic mutation, temperature scheduling, and cross-policy crossover. This closed-loop generation-evolution loop specifically designed for cache replacement policy enables the discovery of compact policies that satisfy hardware constraints while exploring algorithmic transformations beyond fixed predictor structures. Across SPEC CPU2006, CACHEFORGE outperforms all CRC-2 baselines. It improves the total hit rate by 27.36%, 19.69%, 13.72%, 13.15%, 11.83%, and 5.73% over MPPPB, ReD, Hawk-eye, SHiP++, LIME, and LRU, respectively. On memory-intensive workloads, it increases IPC by 10.15%, 7.89%, 6.34%, 3.64%, 3.12%, and 2.71% over LRU, MPPPB, LIME, ReD, SHiP++, and Hawkeye.
Large language models can reason over scientific literature, but textual plans do not reliably produce biological sequences with the intended properties. Sequence generators capture sequence patterns, yet they do not interpret evidence, choose interventions, or reflect on outcomes. This separation creates an evidence-to-execution gap between scientific reasoning and sequence-level control. We introduce EASER, a reflective agent that connects the two through a learned property interface. The interface consists of property-specific low-rank directions trained offline and fixed during search; an LLM controller (the controller) composes them to steer a frozen diffusion generator. From retrieved evidence, sequence context, and prior outcomes, it proposes intervention hypotheses that specify anchors, editable positions, and coefficients. Probe-and-Steer tests these hypotheses and allocates further samples from their joint predicted responses; reflection updates the agent's memory. On computational antimicrobial peptide design, EASER balances activity, non-hemolysis, and non-toxicity. Explicit hypothesis formulation yields larger multi-objective gains than direct action generation, and ablations support evidence retrieval, episodic history, reflection, and Probe-and-Steer. EASER achieves the highest mean hypervolume and lowest mean IGD+ among the evaluated search policies and generators on screened candidates. These results show that a learned property interface can make literature-grounded reasoning executable at the sequence level.
Clinical LLMs must decide not only what diagnosis to produce, but also whether the available evidence is sufficient for autonomous decision making. Binary DECIDE/ABSTAIN formulations merge distinct non decision states and do not explicitly evaluate information acquisition. We introduce a DECIDE/ASK/DEFER formulation together with a blinded protocol that prevents models from using evidence completeness metadata. We evaluate Qwen, Gemini, and GPT on 200 matched clinical evidence states constructed from DDXPlus. The models show substantial differences in action selection under identical evidence, with disagreement in 137 of 200 states. For Qwen, a matched targeted versus random analysis shows that selected information changes the likelihood of a subsequent autonomous decision more clearly than diagnostic correctness. Its matched DECIDE/ABSTAIN baseline further reveals a safety autonomy tradeoff: the three action policy rescues some erroneous autonomous decisions but also removes some correct autonomous deci sions. These results show that separating information acquisition from clinician deferral exposes behavior that binary abstention hides, without yielding a uniformly improved decision policy.
Modern large language models (LLMs) are trained on massive, largely unfiltered datasets, including content scraped from nearly every accessible website and user inputs. As a result, LLMs often memorize and reproduce personally sensitive information (PSI) such as birth dates, phone numbers, and home addresses. This leads to significant privacy risks, particularly for high-profile individuals such as executives, politicians, and judges. Existing mitigations largely rely on machine unlearning. However, these methods often remove more information than needed, degrade model utility and safety, and are highly vulnerable to attacks. This paper presents Whiteout, a practical tool that, upon requests by individuals, prevents LLMs from regurgitating their genuine PSIs, by overwriting them using precise and carefully designed obfuscation samples. We evaluate Whiteout on modern LLMs of varying sizes and makers, including a widely-used OpenAI model. Results show that Whiteout effectively prevents disclosure of the targeted PSIs, has negligible impact on model utility and safety, and outperforms existing alternatives. We also test Whiteout against a wide range of countermeasures, from black-box attacks like jailbreaking to white-box adaptive attacks like relearning and quantization. Finally, we conclude with a discussion on the security and ethical implications of Whiteout.
Low-Rank Adaptive Residual Connections (LARC) give a frozen model a compact numerical state that can learn from feedback. The map $h+BAh$ changes the hidden input to a pretrained mixer. An episode copies a learned initialization $\\rho$ into a private state $\\Phi$, updates its factors when feedback arrives, and resets to the bound $\\rho$. We analyze the factor-space updates and trace this input-side policy carrier into Orbit, the post-training realization of Memory-Mediated Learning Architecture studied here. In Orbit's released default, the learned residual shapes predictions while memory adaptation changes the retained evidence.\n\nThe controlled studies use a rank-4 residual with 12,288 factor entries on frozen MiniCPM5-1B-SFT. Two feedback updates reduce expected query execution error by 24.65 and 36.65 percentage points relative to the respective trained reset states in a four-candidate development task with 16 parameter groups and three paired seeds. A direct support-loss rule reaches 0.78125\\% error. In repository-balanced chronological replay of public continuous-integration jobs, retaining updates raises half-Brier loss from 0.1274 to 0.1808. A follow-up intervention observes same-batch non-descent and inconsistent future benefit from shrinking steps. These comparisons measure when numerical state helps a later read and when the trained reference is more effective.
Recent LLMs increasingly adopt hybrid designs that replace standard attention with linear attention, such as Gated DeltaNet (GDN) and Kimi Delta Attention (KDA). Although they compress the context into a fixed-size recurrent state and substantially reduce the cost of long-context processing, repeatedly reading and updating that state remains a major inference bottleneck. Quantization offers a natural way to reduce this cost, but can significantly degrade model quality, due to the accumulation of rounding errors and the presence of outlier rows and columns in the state. To address these challenges, we propose LeapQuant, a training-free method that achieves near-lossless performance under 8-bit recurrent-state quantization. First, to mitigate error accumulation, we propose per-window quantization, which leaps over a window of tokens and quantizes the state only once at its end. Within a window, outputs are computed from the fixed low-bit state together with high-precision buffered updates. Second, to reduce the error introduced by each quantization, LeapQuant retains the state's largest outliers as a few high-precision Compensator Tokens, which share the update path of real tokens. We then smooth the remaining residual before quantization to further reduce the error. Comprehensive experiments across the Qwen, Kimi, and GLM model families show that LeapQuant substantially reduces memory and compute costs during inference. With accuracy comparable to the FP32 baseline, it achieves average speedups of 2.05--3.70$\times$ at the kernel level and 1.47$\times$ for end-to-end inference on NVIDIA B200, RTX PRO 6000, and RTX 5090 GPUs.
Knowledge distillation provides a teacher-student framework for transferring capabilities at reduced cost and has been widely studied for large language models (LLMs). As LLM-based agents move from response generation to sustained, closed-loop task solving, distillation expands beyond model parameters to knowledge retained in reusable resources and execution workflows. This shift from model-level to agent-level knowledge transfer supports efficient deployment, broader capability access, and continual self-evolution. However, existing surveys largely examine model distillation and agent system design separately, limiting comparison of transfer strategies across agent components. To bridge this gap, we formalize Agent Distillation as persistent knowledge transfer between teacher and student agents and present its first unified survey. Distinguishing experience used for transfer from knowledge retained for reuse, we develop a knowledge-substrate taxonomy based on where knowledge persists: models, artifacts, or harnesses. Building on this taxonomy, we analyze cross-substrate transfer paths and hybrid strategies. To assess the resulting capabilities, we review benchmarks and evaluation protocols for retention, causal attribution, and deployed utility without further teacher support. This survey provides a unified basis for studying agent knowledge transfer and developing efficient, adaptable, and self-evolving agents.
Reinforcement learning (RL) has become an effective post-training paradigm for long-horizon large language model (LLM) agents. However, we find that the resulting policies can be sensitive to various policy perturbations, such as hidden-state noise, pruning, and quantization. In this work, we study how to improve perturbation robustness during policy optimization. We first introduce the notion of a perturbation robust policy and analyze conditions under which perturbed policy updates preserve stable monotonic improvement. Based on this analysis, we introduce Stable Perturbation-Robust Policy Optimization (SPrPO), which applies adaptive and sensitivity-aware perturbations during RL training. We evaluate SPrPO on ALFWorld and WebShop and conduct systematic experiments across multiple perturbation types and scales, showing improved perturbation robustness while maintaining stable policy optimization.
Derivative-informed training improves neural operators by directly supervising their input-output sensitivities, which is crucial when neural operators are used as differentiable surrogates for inverse problems, PDE-constrained optimization, design, and control. However, existing methods rely on offline-generated derivative labels, making data generation slow, storage-intensive, and difficult to adapt across datasets, resolutions, or perturbation bases. We propose sketched tangent consistency loss (sTCL), an on-the-fly derivative-informed training objective that, for PDEs with a known and differentiable residual, enforces sensitivity consistency directly from the governing equation without offline tangent labels or neural-operator architecture changes. sTCL uses randomly sketched input perturbations to provide a lightweight derivative-level physics constraint during training. However, raw forward-sensitivity residual penalties can fail for stiff, ill-conditioned, indefinite, or coupled saddle-point tangent operators. To address this, we introduce lightweight operator-aware loss-conditioning mechanisms selected by a simple tangent-operator decision rule. Across Helmholtz, nonlinear diffusion-reaction, Burgers, Allen-Cahn, and Navier-Stokes, with the same neural-operator backbone for all methods, the PDE-specific sTCL losses achieve solution and Jacobian accuracy comparable to offline derivative-informed training (DIFNO) while eliminating the offline derivative-data generation and storage pipeline. These results show that on-the-fly derivative-informed training need not merely amortize offline tangent-solve cost into training; with appropriate sketching and loss design, sTCL provides an effective drop-in path to derivative-informed neural operators. Code is available at https://github.com/Mou-group-UMN/Derivative-informed-training-on-the-fly.
Artificial neural networks (ANNs) usually represent neural transmission with fixed trainable weights during a training batch, which omits short-term changes in synaptic efficacy. In addition, the discrete time-step simulation requires additional temporal processing that many conventional ANN architectures do not use. To overcome these challenges, we propose Purin, a biology-inspired and ANN-compatible mechanism, that introduces synaptic efficacy modulation into conventional convolutional neural networks. Purin uses a time-interval-based abstraction for neural activities, which allows Purin to introduce short- and long-term synaptic efficacy changes without using discrete time-steps. Purin introduces a bounded factor to represent temporary synaptic efficacy changes, together with two weight matrices that represent input-side and output-side efficacy. The weight matrices are updated by backpropagation and interpreted as the long-term synaptic efficacy changes. Experimental results show that after removing the confounding factors in the AlexNet, VGG11, and GoogLeNet architectures, Purin improves the classification accuracies in all three models across the evaluated datasets.
Extreme events in chaotic systems are difficult to learn from short trajectories because they are controlled by finite-time instabilities rather than by frequently observed bulk dynamics. We propose a mechanism-aware conditioning plug-in framework that turns a nudged ensemble of an imperfect model into a non-intrusive sensor of the local instability geometry. In the small-noise regime, the ensemble covariance has the same finite-time deformation kernels that govern local instability, providing a Jacobian-free proxy for local amplifications around a trajectory. A small Feature-Wise Linear Modulation (FiLM) network injects the ensemble geometry into an otherwise unchanged backbone without modifying the low-fidelity model simulator. We demonstrate this interface in two settings: a Transformer-style residual-attention corrector for a controlled low-dimensional chaotic system and a probabilistic recurrent STORN corrector for topographic two-layer quasi-geostrophic (QG) flow. In the low-dimensional benchmark, leading ensemble-covariance directions identify transient instabilities that are also detected by Optimally Time-Dependent (OTD) modes. FiLM conditioning improves 99th-percentile exceedance-frequency errors over an otherwise identical unconditioned Transformer. In QG, an ensemble-conditioned FiLM-STORN model substantially improves long-horizon rare-event statistics in the data-limited regime, including density-tail errors, exceedance frequencies, and spatial exceedance-area distributions relative to an unconditioned STORN trained on the same data. On averaged high-threshold exceedance diagnostics, it also outperforms the unconditioned baseline STORN trained with 20 times more expensive trajectory data. These results show that local instability geometry is not merely interpretable post hoc, but an actionable conditioning signal for data-efficient rare-event emulation.
When the weights and biases of a neural network range over a fixed finite alphabet, training the network to fit a data set exactly becomes a finite combinatorial problem, and parameterized complexity asks which structural quantities of an instance make this problem tractable. Ganian, Sommer, and Sorge (ICLR 2026) showed that training quantized networks is fixed-parameter tractable when parameterized jointly by the input dimension $\alpha$, the output dimension $\omega$, and the treewidth of the architecture, and they asked whether $\alpha+\omega$ alone suffices. We show that it does not: training binary quantized networks is W[1]-hard parameterized by $\alpha+\omega$, so it admits no algorithm with running time $f(\alpha+\omega)\cdot|I|^{O(1)}$ for any computable $f$ unless FPT = W[1]. Hardness holds already for exact fitting of $k+1$ examples whose inputs equal their targets and form a coordinatewise prefix chain, and also when all biases are fixed to zero. The proof rests on a one-flip routing equivalence. On such data, nonnegative binary weights force every neuron to switch on at most once, and a network fits the data exactly if and only if the architecture contains vertex-disjoint paths that link each changing input to the corresponding output. Every neuron therefore has one of only $k+1$ activation profiles, yet deciding whether these profiles can be realized simultaneously is as hard as disjoint-path routing. Small input and output dimensions thus limit what each neuron can do locally, but not how the architecture forces neurons to share resources globally.
Electricity forecasting often involves spatially related signals observed over regions, substations, and feeders, and Graph Neural Networks (GNNs) represent these relations through a graph. Building a complete GNN forecasting experiment is nonetheless laborious, because graph construction, model selection, training, aggregation, and interpretation often require separate tools. We present GraphToolbox, an open-source Python framework that unifies these stages in one configuration-driven pipeline built on PyTorch Geometric. It offers data-driven graph construction, a single interface for most graph convolution operators of PyTorch Geometric and for recurrent graph cells, online expert aggregation, forecasting interpretability, and significance testing on forecasts. We use GraphToolbox to run a controlled benchmark: we compare architectural and modeling choices for French electricity load and net-load forecasting at regional and national levels under a common experimental protocol. On load, most evaluated graph convolutions outperform additive, recurrent, and foundation-model baselines. Pooled XGBoost, trained on the same inputs, has a lower RMSE than the best graph models, although the difference is not significant. Online expert aggregation and daily model updates improve accuracy further. On net-load, graph models outperform additive, recurrent, and XGBoost baselines, with further gains from forecasting the physical components separately. Statistical tests do not identify a single best convolution for load forecasting.
Speculative decoding accelerates large language model (LLM) inference by using a lightweight draft model to generate multiple candidate tokens that are verified by the target model in a single forward pass. Its speedup is largely determined by the acceptance length, yet existing draft-model training methods mainly optimize cross-entropy or Kullback-Leibler (KL) divergence as proxies. These objectives encourage distribution matching but do not directly optimize acceptance length, and the acceptance mechanism also differs between greedy and sampling-based decoding. In this work, we propose acceptance-length-aware training losses that directly optimize the expected number of accepted tokens within a speculative window. For greedy verification, we derive an expected accepted length (EAL) loss that explicitly maximizes expected acceptance length. For sampling-based decoding, we introduce a window total variation (WTV) loss that optimizes the overlap between temperature-scaled draft and target distributions while accounting for sequential acceptance dependencies. Both objectives can be further combined with a group-relative reinforcement learning stage (GRPO) using simulated acceptance length as the reward. Experiments across different target and draft models, tasks, and decoding settings show that our losses consistently improve acceptance length over KL-based training. WTV provides particularly strong gains under sampling-based decoding, while EAL better matches greedy verification. These results show that directly optimizing the acceptance objective, with losses tailored to the decoding mode, is more effective than conventional distribution-matching objectives.