Prior work has examined the self-consistency of LLM-generated explanations using surface-level perturbation methods. However, the strength of these perturbations is not explicitly measured and controlled. In this work, we propose an LLM-as-a-judge approach to measure perturbation strength in a unified manner across input and CoT perturbations. We then evaluate the self-consistency in explanations generated from various LLMs under controlled strength conditions, ensuring a fair comparison across perturbation types. Experiments show that our proposed LLM-based perturbation strength measure outperforms other embedding- and probability-based approaches and that input perturbations generally affect LLMs more strongly than CoT perturbations. Our work suggests that judgments about a model's self-consistency is fair only within the same perturbation type.
A robust world model must strike the balance between faithfully capturing environmental dynamics and abstracting away from irrelevant content. While reconstruction-based world models ensure faithful supervision, they misallocate representational capacity by pixel area rather than dynamics relevance for visual tasks, which can cause task-irrelevant content to dominate the learned representation. Alternatively, reconstruction-free methods avoid this bias but risk discarding possibly relevant information. We propose StarWM, which uses a cross-attention module trained on self-supervised dynamics to decide where reconstruction applies. A dual-stream decoder then restricts reconstruction to the attended regions, with stop-gradient barriers preventing interference between the two objectives. These components allows reconstruction to supervise the visual content of attended regions without contaminating the latent with non-predictive information. On DeepMind Control with dynamic video backgrounds, default (reward-free) StarWM achieves the strongest performance under random-frame distractors and substantially outperforms reconstruction-based baselines under sequential video. In addition, its reward-augmented variant matches or exceeds reconstruction-free methods on sequential video, achieving the highest overall return across all distractor regimes. Mechanistic probing confirms \starwm{} preserves state attributes with near-perfect fidelity through long-horizon imagination while systematically discarding distractors. Code is available on GitHub.
Discrete diffusion models have become leading architectures for inverse protein folding. Test-time alignment guides these models during inference to generate proteins with desirable attributes. Alignment methods for discrete diffusion models often treat inference as a unidirectional process, lacking effective mechanisms for revisiting undesirable token selections. We introduce Spectral Feedback, an algorithm that selects edit-positions in a feedback loop, allowing the model to iteratively correct its own generations. This approach leverages the masking structure of discrete diffusion models by re-masking and re-sampling tokens. While prior alignment methods focus on what token labels to assign to maximize a target reward, we instead treat which tokens to revisit as the central alignment problem. Selecting edit-positions is challenging because edit effects are interdependent: the impact of modifying one token depends on which others are edited simultaneously. We define an edit-set as a set of positions to edit by re-masking and re-sampling the corresponding tokens in a sequence. Motivated by prior work on sparse interactions in biological systems, we find empirically that edit-set value functions for protein inverse folding admit sparse Fourier representations. This structure enables Spectral Feedback to efficiently learn and optimize the value functions for edit-position selection. Spectral Feedback is model-agnostic and can be applied to pre-trained, test-time aligned, and fine-tuned diffusion models without modifying the underlying generative processes. Applied to inverse folding with a protein stability reward oracle, Spectral Feedback increases the proportion of stable proteins by 32.3 percentage points over a pre-trained model, 24.8 over Best-of-10, and 5.8 over a state-of-the-art RL fine-tuned diffusion model.
Distributed learning in embodied reinforcement-learning agents offers a degree of privacy by retaining raw sensor data on-device and transmitting only policy gradients to the server. Yet temporal structure can amplify this leakage beyond single-frame attacks. We introduce Temporal Reconstruction Attack on Consecutive Encodings (TRACE), an amortized temporal gradient-inversion attack that autoregressively reconstructs the sequence of private observation-action trajectories from per-step policy-learning gradients. The attack exploits two structural signals ignored by prior single-frame methods: (i) cross-time correlation between successive embodied gradients, which we formalize via a conditional mutual-information bound, and (ii) closed-form action recovery from policy-head gradient structure, which we prove exact when standard entropy regularization is sufficiently small. On held-out embodied scenes, TRACE reaches $18.8$ dB PSNR with near-perfect action recovery at $3$-$4.5$ ms per reconstructed frame, dominating the learning-based baseline across all reconstruction metrics and exceeding optimization attacks while running orders of magnitude faster. Further evaluation demonstrates TRACE's broader applicability across recurrent, residual, and compact transformer victim architectures, multi-modal inputs, and larger discrete action spaces. Defense experiments suggest that protecting temporal gradient streams may require sequence-aware privacy mechanisms.
Foundation models are post-trained with reinforcement learning (RL) to maximize specific rewards, such as helpfulness, correctness, or instruction following. This post-training process is computationally intensive, sometimes unstable, and has to be run from scratch every time the reward model changes or when we want to combine multiple rewards. We hence ask: given a new reward function, is it possible to predict the policy RL would produce without actually training on that reward? We answer this in the affirmative by introducing PoEM, a framework for predicting RL outcomes using models previously trained on other rewards. First, we build on the fact that if the new reward function can be written as a linear combination of existing ones, then the optimal policy can be expressed as a combination of the existing policies in log-probability space. Surprisingly, even in cases where the rewards are not linearly connected, we observe that policies learned through RL often span an approximately low-rank subspace across rewards. To our benefit, the weights for this combination can be estimated using only the reward or basis policy outputs on the samples. We turn these observations into an algorithm that takes post-trained models and a target reward function, and approximates the target RL policy without any additional RL training. We experimentally evaluate our approach across programmatic and learned rewards for open-ended generation, with additional experiments on math reasoning, tool calling, and image generation.
We characterize, up to logarithmic factors, the optimal higher-order oracle complexity of finding approximate solutions to composite monotone inclusion problems, formed by the sum of a smooth single-valued monotone operator and a maximal monotone set-valued operator. We introduce the Anchored Extra-Proximal (AEP) framework, which combines an anchored extrapolation step with an inexact anchored proximal update satisfying a relative-error condition. This framework recovers the composite Fast Extragradient method in the first-order setting and yields natural higher-order extensions through Taylor approximations of the smooth operator at the extrapolated point. For every $p\geq 2$, assuming that the $(p-1)$th derivative of the single-valued operator is $L_p$-Lipschitz continuous and the initial distance to a solution is at most $R$, we combine AEP with a bisection line search to obtain a $p$th-order method that attains a tangent residual of at most $\varepsilon$ in $\widetilde{O}((L_p R^{p}/\varepsilon)^{\frac{2}{3p-1}})$ oracle calls. We complement this result with a worst-case lower bound of $\Omega((L_p R^p/\varepsilon)^{\frac{2}{3p-1}})$ for every deterministic algorithm in the $p$th-order oracle model, without imposing any prescribed update structure. For the subclass of smooth nonexpansive fixed-point problems, we further establish a lower bound of $\Omega(\min\{R/\varepsilon, {(L_p R^{p}/\varepsilon)}^{\frac{2}{3p-1}}\})$. By choosing between AEP and Halpern iteration, we match this lower bound up to logarithmic factors, thereby also settling the deterministic oracle complexity of smooth nonexpansive fixed-point problems.
Long-running LLM applications and autonomous agents repeatedly send growing context, making prefix caching critical for reducing prefill cost. This raises a fundamental question: do decades of traditional cache-management techniques transfer directly to agentic and conversational LLM serving? We analyze production traces comprising over 20 billion processed tokens and evaluate 14 eviction algorithms across capacity-constrained HBM and large memory-pool settings. We find that simple Least Recently Used (LRU) is surprisingly robust, offering competitive performance against sophisticated traditional policies, though substantial headroom to Belady's optimum remains. The reason is structural: prefix reuse is dominated by the steady pacing of active sessions, making recency an unusually strong predictor for eviction. Nevertheless, prefix caching diverges from traditional workloads in ways that recency ignores, notably the prevalence of single-turn requests, extreme session footprint skew, and highly non-uniform miss costs because attention computation scales with token depth. To close the gap to optimality, we establish several design principles for prefix-cache management: keeping recency as the baseline while incorporating quick demotion to filter one-hit prompts, compute-aware eviction to prioritize expensive misses, and capacity-adaptive eviction granularity. Our traces and simulator are publicly available.
Coordinating a large population of flexible grid-edge devices can alleviate the need for time-consuming and capital-intensive network upgrades. Artificial intelligence (AI)-based control methods such as multi-agent reinforcement learning or imitation learning are promising to this coordination due to their real-time decision scalability. However, system operators still need an independent and rigorous way to decide whether a given AI system is safe enough for deployment. This paper develops a \emph{finite-sample} probabilistic safety certification framework for AI-based grid-edge coordination. The central idea is to reduce the complete input--AI--decision--evaluation workflow to a binary unsafe outcome under an operator-defined safety specification, and then use exact binomial inference to certify the corresponding unsafe operation probability. Given a set of held-out calibration scenarios, the framework returns the tightest one-sided upper certificate and an accept/reject deployment criterion that controls the probability of false safety certification. Because the certification is for the calibration distribution that may \emph{deviate} from the future operation, we further combine the nominal certificate with physically interpretable sample-space adversarial attacks, a concept widely used in AI models to investigate their fragility. Case studies on grid-edge flexibility coordination with 1{,}000-agent AI models (independent parameters) verify the safety guarantee and the value of integrating adversarial attacks into a rolling-window training-certification-deployment flow.
Pathology vision-language foundation models vary in performance across diseases and tasks, with no single model consistently performing best. The high cost of expert pathology annotation can also limit the labeled data available for task-specific adaptation. Combining complementary pretrained representations is a potential approach to these limitations, yet learning an effective fusion from few labeled examples remains challenging. We introduce Few-shot Fusion Foundation Models of Computational Pathology (FFM-CP), which is a framework that combines multiple pathology vision-language models in the few-shot learning setting. The framework first aligns heterogeneous representations using a closed-form Orthogonal Procrustes transformation estimated from corresponding support images. This alignment preserves within-model feature geometry without training an additional alignment network. Within the aligned space, a unified graph enables information exchange across backbones by jointly refining support-image features and visual and textual class prototypes. These refined representations support complementary text-prototype and case-retrieval branches that capture semantic class knowledge and within-class visual variation, respectively. Each branch learns to combine predictions from all ordered backbone pairs, allowing queries encoded by one model to draw on evidence represented by another. We evaluate three backbone combinations on six histopathology datasets at 4, 8, and 16 shots per class. FFM-CP achieves higher mean macro-F1 than the strongest individually adapted member of each fused set in 50 of 54 comparisons. These findings suggest that combining complementary pretrained representations can improve histopathological classification when annotations are limited.
Change Detection (CD) in remote sensing requires jointly modeling bitemporal context and fine-grained spatial details to accurately localize changed regions. While dense attention effectively captures bitemporal relationships, its computational cost limits scalability to high-resolution imagery. In contrast, conventional feature fusion and coarse decoding often struggle to distinguish true changes from appearance variations and preserve object boundaries. To overcome these limitations, we propose Bitemporal Mamba-Diffusion for Change Detection (BMD-CD), a framework that integrates temporally structured state-space modeling with logit-space diffusion refinement. BMD-CD first converts deep bitemporal features into region tokens and organizes them into explicit temporal partitions for bidirectional state-space propagation. It then employs a Learnable Trigonometric Fusion (LTF) module to integrate temporal information and a Conditional Diffusion Decoder that refines coarse predictions in logit space with a short DDIM schedule to generate fine-grained, high-resolution change maps. Extensive experiments on LEVIR-CD, WHU-CD, DSIFN-CD, CDD, and S2Looking demonstrate the effectiveness of the proposed approach. BMD-CD achieves IoU scores of 0.923, 0.957, 0.882, and 0.979 on WHU-CD, DSIFN-CD, LEVIR-CD, and CDD, respectively, and further exhibits strong zero-shot robustness on ValaisCD and B-FLAIR-test.
A fundamental question in physics is: When does classical behavior emerge from quantum systems? Bosonic Gaussian states provide a natural setting to explore this quantum-classical boundary, as they capture both the classical field behavior and the intrinsic quantum nature of light. Here, we address this problem from a learning-theoretic perspective by asking: When are bosonic Gaussian states classical to learn? That is, under what conditions (if any) can an n-mode bosonic Gaussian state be learned with as few samples, and with operations as simple, as are needed to learn a classical 2n-variate Gaussian distribution? We establish a smooth crossover in learnability governed by the state's thermal fluctuations: - Cold Gaussian states are non-classical to learn: When the covariance matrix satisfies $\Sigma\le(\frac12+O(\frac1n))I$, i.e. close to the vacuum covariance, tomography under single-copy (i.e., non-entangled) measurements fundamentally requires $\Omega(n^3)$ copies, strictly exceeding the sample complexity $\Theta(n^2)$ of learning classical Gaussian distributions. We show that this hardness persists even when few-copy entangled measurements are allowed. - Warm Gaussian states are classical to learn: When thermal fluctuations exceed the vacuum noise, parameterized by $\Sigma\ge(\frac12+\nu)I$ for any parameter $\nu>0$, we prove that single-copy tomography requires $N=\Theta\left(n^2\min(n,1+\nu^{-1})\right)$ copies. This bound is tight and is achieved by simple, non-adaptive, unentangled heterodyne measurements. Crucially, for $\nu=\Omega(1)$, the sample complexity drops to $\Theta(n^2)$, matching the classical case. Our results tightly characterize a quantum-to-classical crossover in the learnability of bosonic Gaussian states, reveal a novel connection between fundamental physics and statistical learning theory, and have implications for real-world sensing experiments.
Diffusion language models (dLLMs), such as LLaDA and Dream, have become competitive with autoregressive (AR) LLMs in generation quality while supporting native parallel decoding. A standard acceleration strategy is block-wise decoding, where each forward pass predicts a block of length B and commits high-confidence tokens. Confidence-threshold commitment adapts which tokens are accepted, but a fixed B still constrains the prediction scope. Existing horizon policies use manually chosen sizes, volatility statistics, or candidate-boundary scores. We instead use the model's own per-step confidence: in-window confidence typically follows a context-dependent cliff, providing a direct signal for adapting the look-ahead horizon. In this work, we propose PACE-dLLM, which refits this parametric cliff periodically in closed form, sets the next horizon at a shared confidence anchor, and retains confidence-threshold token commitment. Under a saturated-yield abstraction with iid, policy-independent monotone profiles, the oracle horizon is the smallest attaining maximal per-pass yield and improves the asymptotic forward-pass rate when fixed blocks undershoot it with positive probability. On four reasoning and code benchmarks across two open-source dLLM backbones, PACE-dLLM achieves the best average accuracy among the evaluated decoders on both backbones, with average throughput speedups of 5.23x on LLaDA and 3.06x on Dream (up to 8.52x on LLaDA MATH) over the unaccelerated semi-AR baseline, advancing the quality-throughput Pareto frontier.
Signed graphs arise in trust--distrust networks, financial correlation systems, biological interaction graphs, and many other domains in which edges can be positive or negative and may also be directed. While signed graph neural networks have improved task-specific learning, graph transfer learning on signed graphs remains underdeveloped. In this paper, we introduce TopoSIGN, a pioneer topology-guided graph pre-training and prompt learning framework for signed graphs. TopoSIGN combines a structural encoder built on the magnetic signed Laplacian with a novel persistent-homology branch that summarizes signed topology through Dowker-complex persistence images. The fused embeddings are then transferred to a prompt learning function. Experimental results on synthetic and real-world datasets demonstrate the efficacy of TopoSIGN in extracting useful structural information in signed graphs, as well as the adaptability and flexibility of the proposed general framework.
Unsupervised Continual chunking is a fundamental problem in machine learning and neuroscience, where the goal is to identify groups of states that frequently co-occur in temporal sequences. A key challenge is to form accurate chunks while maintaining their stability over time. In this work, we propose sheaf regularization to reduce local inconsistencies in Decentralized SyncMap, a self-organizing system, and thereby stabilize its chunking dynamics. We introduce a radial sheaf structure that penalizes distance-dependent radial motion between pairs of variables. Experimental results show that the proposed method achieves the highest normalized mutual information (NMI) among the evaluated SyncMap variants on 12 of 18 probabilistic Continual General Chunking Problem (CGCP) graphs with two-state memory and on 17 of 18 graphs with dynamic memory. In the sequential adaptation experiment, Sheaf SyncMap also achieves high NMI after shifts in the input distribution, indicating that it can adapt to new knowledge while avoiding the negative transfer commonly observed in modern machine learning systems such as neural networks.
Model merging combines separately fine-tuned checkpoints by arithmetic on their weights, and is now a routine way to assemble capabilities cheaply. We ask what it does to emergent capabilities: behaviors that arise from fine-tuning but were never an explicit training target. Studying two independent testbeds (activation oracles and emergent-misaligned models) across three model families, we find that the answer is threefold. First, merging preserves an emergent capability that both parents carry: merging two misaligned checkpoints retains most of their broad misalignment across the whole mixing range. Second, merging cannot create an emergent capability that is superadditive in its parents: no weighted merge of two single-task oracles reaches the jointly-trained oracle's auditing ability. Third, when only one parent carries the capability, merging dilutes it faster than the trained capability that accompanies it: the gap is significant in most settings.
The Platonic Representation Hypothesis (PRH) claims that independently trained models converge on a shared statistical model of reality, yet recent work finds that the evidence weakens under some conditions. In this paper, we measure what models share through the directions (displacement vectors) between samples. Under a single orthogonal alignment limited to just rotation and reflection, these displacement vectors are substantially preserved across 44 independently trained vision and language encoders spanning modalities and asymmetric capability pairs, consistent with the PRH. Pointwise similarity, however, is lower. Representations split into a shared semantic component that is linearly aligned across models, and a private component that is not, which we interpret as having model-specific capabilities. We trace this geometry to concept-level structure: within a model, parent concepts are orthogonal to their child variation vectors; across models, concept displacements are parallel. We predict (and experiments confirm) that fine-tuning preserves Pointwise similarity but reduces displacement similarity, and relational distillation does the opposite. Because semantics align linearly but capabilities do not, we propose importing capabilities from one model to another using a single cached forward pass through the source. We call this Shadow Casting. As a proof of concept, our ShadowCLIP instantiation outperforms strong fine-tuned baselines at far less compute. A cache can be released alongside open model weights, letting one model's capabilities be downloaded and imported into other models with frozen backbones.
Organizations in critical national infrastructure sectors must assess heterogeneous documents for sensitivity before routing or storage. Manual assessment is slow, inconsistent, and unscalable. Extending our prior leakage-controlled benchmark, BERT established the top single-encoder baseline (89.14% accuracy, 89.33% F1-score under 5-fold cross-validation on the Strategic 16K corpus). However, transformer baselines suffer from a structural limitation: fixed input length truncation discards evidence beyond the retained window-precisely where sensitive cables tend to be longest. We present Channel-Boosted MAS (CB-MAS) and instantiate it as IC-MAS (Iterative Consultation Multi-Agent System) to solve this without long-context computational costs. A Channel Critic Agent learns document-adaptive trust weights governing Gated Channel Boosting between two first-window encoders, while paired Consultation Agents iteratively exchange belief states to reconcile evidence from the beginning and end of long documents. IC-MAS holds computation constant regardless of document length by reconciling fixed windows in a compact representation space. Ablation studies show critic-controlled Channel Boosting provides the bulk of accuracy gains, while consultation recovers recall without precision collapse. Critic-Controlled Gated Channel Boosting with Max-Pool fusion and Blackboard Adaptive Consultation achieves 90.72% accuracy, 91.23% F1-score, 92.01% sensitive recall, and 90.46% sensitive precision, using about 54% less average computation than a fixed-round baseline. Gains over the single-encoder baseline are statistically significant (McNemar's test, p less than 0.000001; paired t-test). We include LIME/SHAP explainability, multi-agent evaluation, and an honest accounting of limitations.
Autoregressive (AR) models suffer from local greediness, while diffusion language models (DLMs) often lack the strict causal structure required for reasoning. To combine the advantages and overcome the drawbacks of the dual, we propose Causal Latent Revision (CaLR), a framework that reformulates reasoning as constrained latent optimization. By adopting a causal topology matrix (CTM) from an expert model and implicit differentiation, CaLR performs gradient-guided ``thought revision" to enforce logical consistency, enabling dynamic self-correction of intermediate steps during parallel generation. Empirically, CaLR achieves SOTA DLM performance on complex benchmarks, surpassing strong AR baselines and demonstrating superior robustness in constrained tasks like Sudoku.
We introduce a matrix generalization of the cone of Harvey and Lawson and prove that it is an exact special Lagrangian manifold. We further show that it belongs to a family of exact special Lagrangian manifolds that foliate the balanced manifold arising in deep learning.
This paper builds a theoretical bridge between the economic notion of marginal utility and two machine-learning constructs, matrix factorization and the Key--Value cache of transformer language models. The singular value spectrum of a rating matrix is shown to be a diminishing marginal utility schedule for latent factors, the eigenvalue spectrum of the projected covariance operator to be the marginal utility schedule of a model's learned representation, and cache eviction and low-rank cache compression to be instances of constrained utility maximization under a memory budget. The three collapse into a single allocation rule: retain the top dimensions whose eigenvalue exceeds the shadow price of the binding constraint. The framework is applied to the automated extraction of structured information from geo-mining documents, where it motivates a multi-pass inference protocol, a layer-wise TIES model merging procedure, and a selection policy combining extraction quality, localization drift and energy, scalarized with a Conditional Value-at-Risk term on drift. Two empirical contributions are reported. An 11.2-million-parameter hierarchical classifier, trained in about five minutes on a single GPU, reaches 90.0 per cent level-1 accuracy on a held-out test set from a 973-document uranium-exploration corpus, against 92.0 per cent for a proprietary model on a fifty-document human audit of the same corpus, at a latency of 2.62 ms per card against approximately 2,000 ms for the API and at negligible cost. A diagnostic of uniform-density TIES merging exposes a reproducible degenerate mode in which the merged model returns token-identical outputs across five geographically distinct districts while declaring high confidence; re-executing the merge under layer-wise calibrated densities removes that signature on the diagnostic sample. The full-scale extraction benchmark, including LoRA fine-tuning, is reported as projected rather than measured and remains an empirical extension of this work.