Papers

Dynamic graph learning models how entities and their relations evolve from historical observations to predict future graph states. Existing methods do not fully address dynamic graph world modelling, which requires learning transitions from temporal context to future latent graph states. This is difficult for three reasons: nodes and relations rely on different historical evidence, structural evolution is path-dependent, and future snapshots provide sparse and realization-specific targets. We propose Dynamic Graph World Modelling (DyG-WM), the first framework to jointly predict future node and relation representations from temporal context by answering three questions: what has evolved, how it evolved, and what will evolve next. To capture what has evolved, a historical node-relation context encoder summarizes node-neighborhood trajectories and pair-specific histories. To characterize how it evolved, a node-relation path-signature encoder captures ordered structural changes. To predict what will evolve next, a dual-level joint-embedding objective predicts target-snapshot node and relation representations without reconstructing the raw graph. Experiments on 12 link-prediction and two node-classification datasets demonstrate the effectiveness of DyG-WM across both tasks.

Best-of-N sampling is a simple inference-time scaling strategy for large language models, yet its behavior under incomplete reward models remains poorly understood. We study Best-of-N as a competition between the extreme reward tails of correct and incorrect responses. This view shows that success depends not only on the probability of sampling a correct response, but also on whether correct responses dominate incorrect ones in the upper tail. Using extreme value theory, we characterize three asymptotic regimes: correct-tail-dominant, incorrect-tail-dominant, and critical explaining both monotonic gains and reward-hacking-induced degradation. We further derive finite-sample scaling laws for representative reward distributions, showing that, under Gaussian rewards, incorrect responses with sufficiently large variance can dominate selection even when their mean reward is lower. Unlike oracle selection, convergence with an incomplete selector can be polynomial rather than exponential. Finally, we propose a lightweight tail-tension criterion to estimate when additional sampling may become harmful. Experiments on synthetic data and real reward model scores are consistent with the theoretically predicted regimes, scaling laws, and diagnostic behavior, suggesting that tail competition provides a unified lens on Best-of-N sampling for verifiable problems.

Activation steering adds a vector to a model's hidden states to control a named behaviour. A gain on that behaviour, however, does not show that the vector controls it directly: it may act through a component shared with neighbouring behaviours, which unsteered and random baselines cannot separate. We introduce the competing-target test: a vector is compared with the vector built the same way for a competing target, and an equivalence test states when their difference is below a useful margin. We apply it in open-weight models to vectors for Kantian deontology, which judges an act by the rule it follows, and act utilitarianism, which judges it by the welfare it produces. On held-out situations, prompting a framework raises its score and lowers the other's, whereas vectors for either framework move both scores alike, and none reaches the margin. The two vectors share most of their direction, a request to apply some framework; the same construction separates an optimistic from a pessimistic tone, whose vectors share much less. What a vector shares with its competitors therefore sets what it controls: steering a perspective can deliver a mixture, and measuring a disposition can measure its neighbours, unless tested against the competing target.

In this paper we study the behavior of transformer networks generating text over long periods of time. We show that even a three-token, single-layer attention-only transformer can fail to converge. We then study a rich class of transformers with arbitrary finite depth, layer normalization, MLPs, RoPE encodings, and an arbitrary mix of local and global layers. Every transformer in this class has a softmax temperature threshold above which its empirical cache distribution converges to a unique limit, independent of initial conditions. We show further that, while each transformer has a temperature threshold, the supremum of this set is infinite, so there is no global minimal temperature that forces all transformers to converge.

Clinical assessments are often collected using heterogeneous assessment systems across centers and time, leading to records that mix different but related sets of measurements. This motivates harmonization beyond total-score linking. We formulate clinical harmonization at the measurement level as a bidirectional recoverability problem: given paired observations from two assessment systems, the goal is to identify which measurements can be reliably translated in both directions within an application-defined tolerance, while separating non-translatable components. We propose Bidirectional Feature Harmonization Decomposition (BFHD), a feasibility-driven framework that enforces bidirectionally coupled translation and uses feature-wise output gating to produce an explicit decomposition in the original measurement space. Experiments on synthetic data and real clinical assessment pairs show that BFHD achieves broader feasible harmonization coverage and improved subset stability compared to baselines.

Group Relative Policy Optimization (GRPO) provides no reward-relative learning signal when all rollouts receive identical rewards. In long-horizon tasks that require multiple interactions, rewards are given only upon successful task completion, making all-fail groups common early in training and limiting opportunities to practice later steps. Guide Assisted GRPO replays a successful reference prefix before the learner continues. We study how initial positions, handoff units, and movement rules affect unassisted task completion. Across three model sizes on BFCL, starting near completion is more effective under a fixed turn-level threshold rule. Finer handoffs alone do not consistently help, but outcome-informed tool-call selection outperforms GRPO on trained and untrained task categories. Guided starts also improve AppWorld success on both evaluation splits. These findings highlight where learning starts and how starting positions move as key curriculum choices.

Novelty detection is crucial in various technological and scientific domains. Its importance spans from ensuring the reliability of sensor networks to enhancing the adaptability of computer vision systems. As AI systems become increasingly sophisticated and complex, effective anomaly detection is essential for maintaining reliability and accuracy. This work addresses novelty detection in weather sensor networks and computer vision. In weather sensor networks, particularly for monitoring rainfall, detecting blocked rain gauges can significantly impact data reliability, food security, and disaster management. This work evaluates four existing quality control methods for detecting blocked rain gauges and demonstrates their limitations, particularly in detecting clusters of faulty sensors. By comparing methods such as the Conditional Mixture Model (CMM), PSWQ (Fz-filter), Kernel Regression, and Gaussian Method, we identify their shortcomings and propose three novel quality control techniques that integrate probabilistic fault detection and consistency-based diagnosis framework to improve timely detection of blocked sensors, while reducing false alarms. In the realm of computer vision, current systems often fail to handle multiple types of novelty due to their focus on single-novelty scenarios. We highlight the need for integrated systems capable of handling multiple types of novelty. To support this, we introduce the Novel Snapshot Serengeti (NOVEL-SS) dataset, derived from the Snapshot Serengeti camera trap data, which includes various forms of novelty. We also present the DCA (Detect, Characterize, and Accommodate) baseline system, designed to handle multiple novelty by detecting, characterizing, and accommodating them. This work aims to advance the robustness and adaptability of novelty detection systems across both weather sensor networks and computer vision applications.

Layer-ablation studies often perturb early or late layers of a language model, compare the damage to prediction on two corpora, and read a larger early-layer effect on one corpus as early-layer specialization for its task. Such a comparison averages over the tokens that each corpus asks the model to predict, and these target tokens differ: math solutions, for example, ask for far more whitespace and numbers than encyclopedia prose. We show that this difference can create the appearance of depth specialization. We split a cross-corpus early–late damage interaction by target-token type and recompute it under a target mixture shared by both corpora. Between GSM8K and Wikitext, a positive interaction replicates across five model instances. In the three fully decomposed models, only whitespace has a positive interaction. Removing whitespace reverses the pooled effect in all three, and matching the full target mixture reverses it in two. The diagnosis holds under position matching, adversarial reassignment of token types, and a continuous readout. The risk can also be anticipated from the tokenizer alone. A screen computed before any model runs places every held-out pair it scored in the region its label predicts, and on the pairs it rates at high or medium risk, removing whitespace reverses every positive pooled estimate. A smaller content-word effect one token ahead follows the share of content targets that the unablated model already predicts. A cross-corpus depth effect therefore supports an interpretation of task specialization only after the target-token mixture is reported, the effect is decomposed by token type, and a mixture-matched summary keeps its sign.

A training-data quality gate changes both which examples survive within regions of the input space and how training mass is distributed across those regions, so a gate-versus-random comparison does not isolate these mechanisms. We introduce Counterfactual Gate-Support Decomposition (CGSD), a four-arm, budget-matched design that crosses gate eligibility with support distribution and decomposes the observed gate effect into within-support selection, between-support redistribution, and their interaction. The practical constraint is finite-pool realizability: we characterize mass and token-assembly conditions under which the required arms can be constructed, together with a support-extinction case in which the ordinary design is undefined. On Korean-English and Korean-Japanese retrieval, the mean within-support selection component is the only one that exceeds its calibrated null band (+29.7 and +5.6 holdout items; 30/30 positive replicates), decisively in Korean-English and marginally in Korean-Japanese; redistribution and interaction remain unresolved at the available power. The same pattern holds for a second, unrelated gate family – a model-confidence threshold rather than a deterministic verifier – judged against its own null (+27.3 items, 9/9 positive), so the result is not a property of one gate's mechanism. On a mathematics gate, zero-eligibility supports leave the ordinary design undefined at every budget, while the extended design is constructible only inside narrow budget bands that an exhaustive scan locates and that the natural operating point misses – the assembly condition separating a property of the pool from a limitation of the packer. CGSD therefore returns a decomposition, a constructible-but-underpowered design, or an explicit non-constructibility result, and says which; it prevents support redistribution from being silently interpreted as sample-level gate quality.

In large language models (LLMs), attention concentrates heavily on the first token. This 'attention sink' is believed to protect deep models against over-mixing. Even though graph transformers face a related pressure, we find that a SoTA graph transformer, GRIT, rarely forms sinks on molecules, and that it sinks less when trained on larger graphs, the opposite of the trend in LLMs. In principle, a graph can prevent sinks: we prove and show empirically that its symmetry and locality limit the attention any node can receive. On molecules, however, with distinguishable atoms and global attention, neither limit prevents sinks; what stops them lies in the model. We find that the attention scoring function is the main cause. Replacing GRIT's gated logit with a dot product raises its sink rate from 0.09 to 0.68 on identical graphs. In Graph Transformers using dot-product attention, removing the sink gives about the same error as removing a random node, while GRIT's most-attended nodes matter to its predictions. Attention sinks are therefore a property of dot-product attention rather than of Transformers. In our experiments, they form when dot-product scoring has a distinguishable node to concentrate on.

Recent advancements in topic modeling have split into two main paths: efficient embedding-based methods that offer scalability but are often outperformed on standard clustering metrics, and Large Language Model (LLM) approaches that achieve state-of-the-art clustering performance but at the cost of prohibitive computational expense and poor scalability. This paper introduces TopicMerge, a novel multi-stage pipeline that bridges these approaches to achieve both state-of-the-art clustering performance and computational efficiency. TopicMerge first employs an efficient clustering pipeline to deliberately over-segment a document corpus into a large set of fine-grained, semantically pure micro-clusters, which are then given initial semantic labels. It then introduces a novel, iterative merging algorithm where an LLM, guided by local neighborhood context, arbitrates on consolidating candidate pairs. These approved merges are then selectively executed and semantically relabeled by the LLM, hierarchically building broader, more coherent topics. Extensive experiments on the Bills and Wiki datasets show that TopicMerge achieves state-of-the-art clustering performance while reducing required LLM calls by over 92% compared to document-level methods, offering a practical and scalable solution.

We present a global-in-time spectral Duhamel-Picard method with polynomial expansions for post-exposure bake (PEB) simulation. The coupled acid-quencher reaction-diffusion equations are reformulated as nonlinear Volterra integral equations using the Duhamel principle. Full trajectories are represented by finite polynomial expansions, with temporal coefficient fields determined over the bake interval by Anderson-accelerated Picard iteration, thereby avoiding sequential time marching. In 3-D PEB simulations, we employ Chebyshev and Legendre temporal bases while evaluating the Duhamel integrals using the corresponding Gaussian quadrature rules. We demonstrate that the proposed method simulates a large domain of $1200\times1200\times100$ voxels on the order of one minute. The representation is further incorporated into a physics-informed neural operator (PINO) with a mixed-transform variant of a factorized Fourier neural operator, which is trained without precomputed solution trajectories. Together, the integral formulation and its PINO extension enable accelerated, time-marching-free prediction of concentration fields at arbitrary times within the bake interval.

Sequential screening policies must use acquired feedback effectively under a fixed experiment budget. We study acquisition-time residual adaptation: a small ranking correction built from fixed historical features and anchored to base scores recorded before feedback is revealed. The interface accepts a history-conditioned scorer and a candidate feature table; its primary instantiation augments AssayFormer's ten-dimensional representation with 64 historical directions while retaining the pretrained model. In retrospective evaluation on 20 official-test screens from 15 recorded sources, a comparison fixed before target-outcome access increased source-macro fraction of hits recovered from 22.58% to 28.61%, exceeding all six prespecified comparators. A same-feature ordinary ranking head achieved 25.73%. Supplementary strong logistic, rank-fusion, and unprojected controls achieved 26.42%, 26.02%, and 26.78%, respectively. A separate model-independent pilot on 13 reused development screens reconstructs the features without AssayFormer assets. Across three seeds, residual adaptation improves random-forest hit recovery from 28.87% to 31.27%, while logistic recovery changes from 30.22% to 29.78%. The results show effective pretrained-policy adaptation and useful transfer to an independent nonlinear base, with gains that depend on the base model.

Jailbreak attacks against large language models (LLMs) are evaluated using attack success rate (ASR), which depends on correctly identifying responses that fulfill prohibited requests. Existing evaluators can mistake refusals or topically related responses for successful attacks. We show that changing only the judge LLM can alter reported ASR by up to $82$ percentage points. We introduce the semantic reconstruction game, a training-free evaluator that decomposes compliance judgments into constrained semantic comparisons. Given only a response, an LLM reconstructs the request answered by its substantive content, excluding information merely repeated or refused. The game compares this reconstruction against similar decoy requests to determine whether it matches the original request, allowing abstention when neither candidate matches. Across three off-the-shelf LLM judges, the game achieves the lowest mean error against human labels among evaluated methods and reduces variation in reported ASR. It prioritizes avoiding false positives, even at the cost of missing some successful jailbreaks.

On-policy distillation (OPD) provides dense token-level supervision on student rollouts, but this supervision can become unreliable when the student and teacher diverge, either because of a large initial policy gap or compounding errors in long-horizon reasoning and tool use. We introduce Divergence-Triggered Dynamic Distillation (DTDD), an adaptive OPD framework that uses segment-level student-teacher divergence to trigger teacher takeover during training. When divergence triggers an intervention during reasoning, the student prefix continues to receive standard OPD supervision. The teacher then generates a recovery suffix, which the student learns through behavior cloning, using cross-entropy on the teacher-generated tokens. We evaluate DTDD in three regimes: long-horizon reasoning, distillation from a base initialization without offline supervised fine-tuning (SFT), and multi-turn agents. We also compare alternative trigger designs and teacher-intervention rates. DTDD reduces the post-warmup peak gradient norm by $8.6\times$ and, from a base initialization, reaches $21.3\%$ held-out math accuracy in $2.6\times$ fewer optimizer updates than OPD; both methods finish near $22\%$. It improves best-observed performance by $4.8$ and $4.1$ percentage points on ALFWorld with LoRA and agentic code, respectively. Ablations show that segment-averaged triggers outperform token-spike triggers. Behavior cloning is also more effective than the tested OPD surrogate for training teacher-generated recoveries. Across 23 experimental arms, takeover tends to help when the teacher substantially outperforms the OPD-trained student, provided that the teacher can recover from student-visited states, the student can learn those recoveries, and teacher intervention is neither excessive nor insufficient.

Reinforcement learning for code usually rewards passing tests, which a model can do while remaining wrong on the inputs the tests do not cover. A program verifier instead decides whether a program satisfies a formal specification for all inputs. Language models are almost always paired with verifiers at inference time only, and turning a verifier into a reward raises two difficulties. It reports only acceptance or rejection, which gives no gradient when every attempt fails, and a policy can cheat it, most simply by telling it to assume what the proof was supposed to establish. We build an agentic environment and reward that address both. The reward reads the verifier's diagnostics rather than its verdict, scoring a candidate by the weighted fraction of proof obligations it discharges, extended with hidden probes from a reference proof's annotations. An integrity audit outranks the verifier, rejecting any candidate that alters the specification or assumes what it was meant to prove. The model works as an agent against a live verifier over many turns, with a separately trained proof specialist as a frozen collaborator. Across Qwen3 models from 1.7B to 14B and a 35B Qwen3.6 mixture-of-experts model, held-out verification on code generation rises from 66.1% to 93.5% on DafnyBench and from 56.3% to 100% on SWE-Proof. This reward leads every other training signal we compare, including a binary pass-or-fail reward trained under otherwise identical settings, with the widest margin on code generation. When the agent must also deliver the repository patch, graded by each project's withheld tests, the fraction of held-out issues it resolves rises from 59.6% to 72.5%. Evaluated unchanged on three verification suites withheld from training, the checkpoints improve on all of them, including one whose language and verifier are absent from training.

Understanding neural scaling laws is increasingly important for theory and practice, yet first-principles explanations remain limited even in simple settings. In this work, we study capacity-limited scaling in a bottlenecked bigram model with power-law (Zipf) token statistics, viewed as an analytically tractable toy model of a large language model. Fixed random embeddings serve as a proxy for representational capacity and induce an explicit bottleneck of width, allowing us to derive scaling laws for prediction regret as a function of model width, vocabulary size, and the Zipf exponent. Our results complement recent theoretical accounts based on optimization dynamics by isolating scaling due purely to representational constraints. In the appropriate regime, the theory recovers the functional form of classical neural scaling laws and suggests that heavy-tailed language statistics may underlie both optimization- and capacity-driven scaling. Experiments natural language datasets (WikiText2, OpenWebText) closely match the predicted curves.

Modern machine learning models increasingly operate on variable-size structured data, such as sets and graphs. A central requirement is size generalization: the ability of a model trained on inputs of one size to generalize to inputs of a different, potentially much larger, size. Classical statistical learning theory, however, primarily considers fixed input domains, and its guarantees often deteriorate with input dimension. Although extensive work has established when models on sets and graphs can be transferred across input sizes, such results do not characterize the statistical cost of doing so: the sample complexity of size generalization, and its relationship to classical in-domain learning, remain largely unknown. We develop a statistical learning theory of size generalization and establish a sharp exponential separation in sample complexity. Letting $n$ denote the number of training samples, we show that polynomial source decay is insufficient for efficient size generalization: while it yields error $n^{-\Theta(1)}$ and hence polynomial sample complexity for in-domain learning, size-generalization error decays only as $(\log n)^{-\Theta(1)}$, yielding an exponential separation. Thus, polynomial in-domain sample complexity can become exponential across sizes. We then identify an exponential source condition under which size generalization achieves error $n^{-\Theta(1)}$, recovering polynomial sample complexity. These guarantees are independent of input dimension and depend only on the number of training samples, yielding dimension-free scaling laws. We also provide matching minimax lower bounds, establishing optimality. To our knowledge, this work provides the first dimension-free minimax theory of size generalization and a rigorous characterization of when, and at what statistical cost, learning can transfer across input sizes.

On-policy distillation provides dense supervision on the student's own samples, but its targets still come from the teacher. Verifier reinforcement learning (RL) instead uses an outcome signal, but it provides no direct constraint on capabilities outside the verifier. Combining the two is difficult for diffusion language models (dLLMs), because bidirectional denoising makes the policy likelihoods required by ratio-based RL unreliable. We develop a likelihood-free pipeline for a mixture-of-experts dLLM. Each rollout group is routed by verifier outcome. Mixed groups use ratio-free, group-centred REINFORCE. If every student rollout fails but a teacher succeeds, the model receives dense KL supervision on the student's denoising canvases. All other groups are skipped. A KL term anchored to the distilled checkpoint limits changes outside the verifier reward. The tool-calling teacher succeeds on only 18.6% of the training prompts. At 24 denoising steps, distillation exceeds the teacher by 6.77 percentage points, and routing improves the distilled checkpoint by a further 18.31 points. Routing lowers reasoning-trace accuracy by 0.71 points but raises direct-answer accuracy on the same items by 13.21 points. The anchored arm restores trace accuracy. Its tool-calling score differs by -0.16 points, with a confidence interval that includes zero, although this comparison also removes a small fallback distillation route. A domain-conditional recipe retains a 16.37-point tool-calling gain; a second seed gives 18.23 points and reproduces both preservation results, but the math result is seed-dependent. We pre-registered every arm and evaluated each comparison on paired items in a single sweep on one machine.

The ecological insurance theory posits that diverse forests are more resilient and better able to adapt to changing environmental conditions. Increasing drought and heat stress in Central Europe promotes mixed forests for maintaining ecosystem services. However, a diverse species distribution complicates remote sensing-based species mapping, as distinguishing pure from mixed species compositions at the pixel level remains challenging. We evaluated the potential of recurrent neural networks (RNNs) for tree-species classification from Sentinel-2 time series and investigated the spectral-temporal features underlying class discrimination, with particular emphasis on separating pure from mixed Norway spruce pixels. We compared RNN classification models with Random Forest (RF) and Extreme Gradient Boosting (XGBoost) and used Integrated Gradients (IG) to quantify feature attributions across spectral bands and time. The best-performing RNN achieved a Macro-F1 score of 0.851 on the test data, outperforming XGBoost by 0.012 and Random Forest by 0.103. Beyond classification performance, the attribution analysis revealed distinct spectral-temporal patterns associated with the separation of pure and mixed Norway spruce pixels. Attribution values increased particularly after approximately day of year (DOY) 125, while B11 showed the strongest contribution during periods of increasing spectral divergence between the two classes. These findings show that recurrent models can provide competitive classification performance while also revealing the phenological periods and spectral regions that are most relevant for distinguishing closely related forest compositions. Combining time-series classification with explainability therefore provides additional insight into the spectral-temporal mechanisms underlying tree-species discrimination and may support more detailed mapping of mixed forest structures.

Physics-informed neural networks provide a flexible framework for learning parametric PDE solution families without labeled data, yet preserving accurate derivatives across heterogeneous regimes remains challenging. Structured neural solvers offer compact representations, but restrictive conditional parameterizations can limit adaptation across the solution family. We introduce BeTINN, a structured conditional architecture extending time-induced modulation through multi-directional time-parameter conditioning and a joint residual pathway. We evaluate BeTINN on the Heston stochastic-volatility PDE, a demanding parametric problem with mixed derivatives and economically relevant first- and second-order sensitivities. BeTINN improves joint accuracy across normalized option value and key sensitivity metrics relative to our baselines. Controlled ablations characterize the roles of its architectural components. We further find that validation physics loss is informative but does not fully determine differential fidelity, and that accurate solution values can coexist with localized derivative errors. These results support structured conditional adaptation and motivate derivative-aware evaluation when sensitivities are downstream quantities of interest.

Fairness audits with several sensitive attributes are often performed one coordinate at a time. This can miss purely intersectional disparities: a score may be independent of each sensitive coordinate while depending on their joint profile. We study distributional demographic parity for a real-valued score using one-dimensional Wasserstein geometry. We distinguish two complementary constructions. Conditional and sequential audits characterize joint parity through their zero sets, but their positive values do not generally decompose a fixed joint Wasserstein disparity. By contrast, full-cell conditional quantile functions define a Hilbert-valued random variable whose variance is the joint disparity. Conditional expectations of this quantile process induce a monotone information game: every reveal order yields an exact martingale decomposition of the same joint disparity, while Shapley values provide an order-free attribution across sensitive attributes. Under independent sensitive attributes, the game reduces to a Hoeffding-ANOVA decomposition and Shapley splits each interaction energy equally among its members; under dependence, orthogonality is lost and explicit formulas reveal suppression and redundancy effects that marginal audits can hide. We give fixed-grid estimators and permutation calibration, and illustrate the framework on synthetic examples and ACSIncome. The resulting framework separates three questions that are often conflated: whether joint distributional parity holds, how a fixed joint disparity decomposes along an information path, and how that disparity can be attributed across sensitive coordinates without privileging an ordering.

Equivariant structure-prediction pipelines alternate coordinate updates with the construction of a k-nearest-neighbour (kNN) graph whose membership is a discrete, piecewise-constant function of the coordinates. Rebuilding that graph at every block boundary (PB) is the expensive part of an epoch, whereas reusing one frozen graph (F) is fast but leaves a topology gap. We study a two-pass design in which the cheap pass records a small amount of state (margins, kNN radii, measured per-boundary displacements) and a second pass compensates conservatively. We formulate the choice of the graph-construction basis as a damped fixed-point iteration whose unique fixed point is exactly PB, and derive (i) exact non-flip certificates with the two-sided constant $m_i/4$ and one-sided constant $m_i/2$, (ii) the minimum-variance (BLUE) replay gain $\gamma^*$ and the resulting feasibility crossover, (iii) an $\ell_\infty$ trust-region projection that preserves convergence, and (iv) an analytic near-tie spacing law $E[m] \approx r_k/(3k)$ that matches measurements to 1%. Three compensation levels are implemented in C++ (L1: frozen set with live geometry; L2: margin-gated local repair; L3: damped, projected basis) and validated under a matched-conditions protocol (identical seed, MSA depth and trajectory family, only the topology mode differs). L2 reproduces the PB edge set exactly ($D = 0.0000$ at every boundary), L1 stays within 0.6% and L3 interpolates monotonically: L2 reproduces the PB edge set exactly when the trust region is removed, and L3 interpolates between the frozen and the PB topology. The recorded trajectory itself yields a measured $1.5\times$ Pass-1 and $1.76\times$ epoch speed-up with no measurable loss gap. An open-source implementation of the runtime and of all the experiments reported here is available at https://github.com/AlbertHuang96/ProteinPredictMachineLearning.

Financial disclosures express the underlying relationships of double-entry accounting through unstructured text. While financial LLM benchmarks predominantly emphasize question answering and classification, schema-guided evidence discovery remains relatively under-explored. We investigate whether open-weight LLMs can extract and classify evidence corresponding to predefined double-entry accounting relationships underlying the text. Using a financial accounting graph comprising 11 nodes (accounts) and 11 edges (transaction types), we evaluate mid-sized models from 5 open-weight families across 1,300 SEC 10-K Item 7 filings spanning 26 years, comparing their extraction performance against three proprietary models. Results reveal substantial differences in extraction volume, edge distributions, and cross-model agreement. Building on this study, we introduce a human-annotated relation classification benchmark comprising 483 passages across three financial accounting categories, evaluating 34 open-weight and 10 proprietary models. The strongest open-weight and proprietary models achieve 84.1% and 92.8% accuracy, respectively, with considerable variation across accounting categories. Our findings highlight the importance of evaluating evidence discovery and relation understanding separately, while investigating locally deployable models for scalable, privacy-sensitive financial analysis.

When writing code, the same computation can be expressed in different programming languages and through different implementation strategies. We show that the activations of LLMs capture this shared structure in reusable and composable abstract representations when they generate code. These representations emerge in the middle layers, are reusable across programming languages, and afford the model an impressive amount of generalization ability when generating code under various interventions. Using activation steering, we show that function, programming language, and implementation strategy can be independently manipulated. The independent interventions on language and function can be flexibly composed to control code generation, even in cases when the prompt contains no information about language and function. We additionally find that finetuning on examples of arbitrarily defined functions in one programming language transfers to other languages, consistent with the models reusing abstract features of code generation that are shared across different examples. Finally, to study how such abstract representations come to be, we implement a minimal model of shared, reusable, and composable structure in a synthetic setting, and show that transformers trained on such data learn representations that reuse these parts flexibly across contexts.

Automatic stuttering systems primarily detect dysfluencies after they have occurred, limiting their use for anticipatory intervention. We address this gap by formulating stuttering prediction as continuous causal time-to-event forecasting, where a model estimates the risk of an upcoming stuttering onset every 100 ms using only preceding speech. To support this task, we augment FluencyBank Timestamped for 36 speakers with manually reviewed repetitions, prolongations, and blocks, yielding 2,872 forecasting events. We then model the temporal evolution of frozen Whisper-large-v3 speech representations using a compact causal Conformer with 3.67M trainable parameters. The proposed approach achieves an AUROC of 0.63. Temporal analysis further shows that explicitly modeling preceding speech states outperforms simply increasing the static acoustic window, while a discrete-time survival formulation provides coherent forecasts from 300 ms to 1.5 s using a single model. These results provide initial evidence that stuttering onset is preceded by measurable temporal information in speech and establish a framework for future anticipatory stuttering systems.

Large language models impose substantial memory demands and arithmetic costs during inference. Dense-to-mixture-of-experts (MoE) conversion is a model optimization technique that aims to reduce these arithmetic costs. Dense-to-MoE conversion partitions pretrained feed-forward network (FFN) neurons into experts and selects a subset for each token. Leading methods fall into two categories: analytic methods produce a converted model quickly using heuristics, while learning methods produce a model much more slowly by optimizing via gradients. Both approaches, however, fix the same active expert count, or "utilization," at every layer. We introduce Spectral-MoE, which learns to allocate a utilization budget heterogeneously across layers. Its name draws a metaphor between separating a dense FFN into experts and separating a light beam into its spectrum. Spectral-MoE consists of two distinct approaches: a static approach that uses a single per-layer scalar which is shared across tokens, and a dynamic approach which emits a per-token schedule. We demonstrate that both allocators improve performance over uniform on both the analytic-based and learning-based SOTA dense-to-MoE methods. At 25% FFN utilization, dynamic allocation reduces WikiText-2 perplexity relative to uniform by 1.04, 2.96, and 2.71 points on the learned base for Llama-2-7B, Llama-3-8B, and Qwen2.5-7B, respectively. The corresponding reductions on the analytic base are 2.45, 6.26, and 2.17 points. On the learned base, dynamic allocation improves six-task mean accuracy over uniform by an average of 1.5, 4.2, and 3.6 percentage points across the three models at FFN utilizations of 12.5%, 25%, and 50%, respectively. These results establish per-layer expert utilization as a novel optimization axis in dense-to-MoE conversion.

The Continuous Thought Machine (CTM) is a recurrent network that attends to inputs and makes predictions based on the synchronization between its neurons. On tasks spanning many domains, it displays emergent thought processes that have not been explained. We report a number of findings about how the CTM classifies images, focusing on ImageNet. Each thinking step is an attracting but strongly non-normal map exhibiting balanced amplification (Murphy and Miller, 2009); while attraction is already present at CTM initialization, training highly promotes non-normality. By Banach's theorem, the neural state has a locally unique attracting fixed point, which moves at every step, and the thought trajectory is the curve these fixed points trace, followed one tick behind. Most of that curve's motion is shared across images, and visible as a previously reported traveling wave of activity across neurons. The attention schedule along the thought is largely generic: shuffling queries across images at every tick costs only 7 percent of classification accuracy, while image-independent increments predict the next query with R^2 = 0.98 (held out). Our results suggest that the CTM has found a single strategy to process all ImageNet images.

Large language models are increasingly deployed as teams of agents that hold different private information about a shared task. How should such agents concisely communicate in a way that promotes aggregation of decision-relevant information? Communicating only proposed actions can discard decision-relevant uncertainty: two agents may prefer the same action for different reasons. We study an alternative interface in which agents communicate calibrated beliefs over actions. We first analyze a simple synthetic setting that captures the structure of multi-round reasoning under partial information. In this setting, we show theoretically and empirically that communicating unbiased probabilities can be strictly more powerful than communicating actions. We then test the same principle in a collaborative maze-solving task with trained transformers and pretrained LLM agents. Our results demonstrate that probability communication outperforms action communication when the exchanged predictions preserve calibrated uncertainty. When communicated probabilities are biased, we show that the benefits of probability communication can disappear. We give a post-hoc conversational calibration intervention that consistently improves decision-making by correcting these biases both in a toy transformer setting and when pretrained LLMs use the corrections in context. These results suggest that probabilistic communication is useful for multi-agent collaboration not merely because it transmits more information, but because calibrated uncertainty gives other agents a reliable object to update on.

While causal discovery methods have matured for complete data, such algorithms for handling missing values remain limited. We establish general sufficient conditions under which causal graphs remain identifiable (i.e., uniquely determined from the available data distribution) under various missingness categories, including Missing Not At Random, and provide a framework to extend correctness guarantees from causal discovery algorithms for complete data to missing data settings. Our results show that identifiability of causal graphs with no unobserved confounders is preserved when (1) missingness mechanisms do not causally interact, (2) no variable directly causes its own missingness, and (3) either faithfulness holds for missingness mechanisms, or they follow generalized linear models. Guided by these theoretical results, we introduce two asymptotically correct algorithms for Additive Noise Models (ANMs) that recover the underlying causal graph and provide consistent parameter estimates.