Papers

Building state-of-the-art (SOTA) predictive models for drug discovery requires expensive search over tools, architectures, and training strategies. Current LLM-based agents can find SOTA solutions through extensive trial and error, but they do not retain the experience accumulated along the way and therefore pay the full search cost on every new task. We propose \method (Self-evolving Agent Experience), a framework that accumulates and reuses experience across tasks to build SOTA drug discovery models efficiently. \method maintains a cross-task memory of verified skills, statistical evidence about effective strategies, and a record of recurring errors and their fixes. In some cases, \method transfers a working solution directly without test-time search. In 33 molecular property prediction tasks, \method ranks first among nine SOTA agents in a single-task setting. With memory accumulated from 16 smaller tasks, \method achieves an averaged normalized score of 0.935 on 17 held-out tasks in a cross-task evaluation setting and outperforms all baseline agents by 10-30% in a zero-test-time search regime. In summary, our work shows the advantage of cross-task memory for efficient SOTA model development in drug discovery.

The minima of modern neural network loss functions are typically not isolated, rather they form connected components of reparameterization invariant solutions on the training data. Analytically characterizing these solutions is a hard problem, but sampling approaches are feasible. By construction, existing methods either spread over low-loss regions, and thus do not sample reparameterization invariant solutions exactly, or are inherently local, which limits exploration of other minima valleys. We propose sampling such reparameterization invariant models using a dynamical system based on kinetic energy, subject to a gravitational pull and a friction term that dissipates energy from the system. Our proposed sampler, DiMS, is guaranteed to sample exactly from the minimum level sets and depends on physically motivated hyperparameters which allows control over the exploration capabilities of the sampler. We consider uncertainty quantification in Bayesian inference as the motivating problem and observe improved performance compared to previously proposed approaches.

Reasoning-trained language models often spend more tokens on harder problems, but longer chains of thought do not show whether a model is merely computing for more steps or following a different internal trajectory. We study this distinction through hidden-state trajectories during chain-of-thought generation across competitive programming, mathematics, and Boolean satisfiability. Raw trajectory geometry is strongly shaped by generation length: longer generations mechanically alter path statistics, so difficulty-dependent comparisons are misleading without adjustment. After residualizing trajectory statistics on length, difficulty remains systematically coupled to corrected trajectory geometry across all domains studied. The clearest reasoning-specific separation appears in the code domain, where harder problems show more direct corrected trajectories and less heterogeneous local curvature in reasoning-trained models than in matched instruction-tuned baselines. Corrected difficulty-geometry coupling is weaker, but still present, in mathematics and Boolean satisfiability. Prompt-stage linear probes do not mirror the code-domain separation, and behavioral annotations show that stronger corrected coupling co-occurs with strategy shifts and uncertainty monitoring. Together, these findings establish length correction as a prerequisite for generation-time trajectory analysis and show that reasoning training can be associated with distinct corrected trajectory geometry, with the strength of the effect depending on the domain.

Concealed Object Segmentation (COS) encompasses a family of dense-prediction tasks, including camouflaged object detection, polyp segmentation, transparent object detection, and industrial defect inspection, where targets are visually entangled with their surroundings through different physical mechanisms. Existing methods either operate directly on RGB images or employ heterogeneous decompositions (\eg, Fourier, wavelet) that redistribute spatial evidence across scale/frequency coefficients, making pixel-aligned cues less direct. We introduce a fundamentally different perspective: homogeneous image decomposition via Retinex theory, which factorizes an image into illumination and reflectance components within the same spatial domain. Our key insight is that visual entanglement enforces appearance matching in the composite space, but this does not necessitate simultaneous matching in both component spaces, a phenomenon we formalize as the Discriminability Gap Theorem. Crucially, we show that across diverse COS sub-tasks, the underlying physical processes systematically anti-correlate illumination and reflectance differences, yielding theoretical guarantees that Retinex decomposition preserves or strictly improves total foreground–background discriminability across the full physical regime, with anti-correlation maximizing the gain. Building on this, we propose RIDE comprising: (i) a Task-Driven Retinex Decomposition module that learns segmentation-optimal factorizations end-to-end; (ii) a Discriminability Gap Attention mechanism that adaptively exploits where decomposition helps; and (iii) a Camouflage-Breaking Contrastive loss operating in reflectance feature space.

This paper presents a comprehensive analysis of neural activation patterns across six distinct large language model (LLM) architectures, examining their performance on twelve cognitive task categories. Through systematic measurement of final activation values, attention entropy, and sparsity patterns, we reveal fundamental differences in how encoder and decoder architectures process diverse cognitive tasks. Our analysis of 144 task-model combinations demonstrates that mathematical reasoning consistently produces the highest attention entropy across all architectures, while decoder models exhibit significantly higher sparsity patterns compared to encoder models. The findings provide critical insights into the computational characteristics of modern language models and their task-specific neural behaviors, with implications for model selection and optimization in big data applications.

Standard deep-learning pipelines usually choose the network architecture before training and keep it fixed throughout optimization. In contrast, a model can also be adapted by editing its structure during training, for example by pruning existing hidden units or growing new ones; however, growth is not simply the inverse of pruning. Pruning selects among units trained from initialization, whereas growth inserts new capacity into an already specialized optimization trajectory. We isolate this insertion problem and show that newborn units can be forward-active yet backward-starved, receiving substantially weaker gradient signal than incumbent units. This asymmetry is weak in small MLPs but emerges in more challenging convolutional settings, where Grow reaches competitive final architectures despite weaker trajectory-level performance. Extending structural edits to residual networks reveals a strong dependence on edit location: head-localized growth produces concentrated but task-aligned representations and faster bounded label remapping, whereas growth throughout the residual hierarchy yields lower effective feature support, and a lower bounded adaptation endpoint. Interventions targeting optimizer state, insertion, selection, and trainability improve newborn integration, but not necessarily final subnetwork quality. Across continual-learning benchmarks, growth is most competitive when newborn units have sufficient time and trainability to integrate. Therefore, Grow should be evaluated not only by the architecture it ultimately discovers, but by whether late-arriving capacity becomes usable before a new structural change occurs during training.

Electroencephalograph (EEG) timeseries signals are characterized by significant noise and coarse spatial resolution, which complicates the classification of neurodegenerative diseases. Even SOTA deep learning architectures struggle to distinguish between healthy controls and diseased subjects, or between different disease types, due to high intergroup similarity. In this paper, we show that a spectrally selective approach to feature construction enhances class separability. By isolating signal strengths within the primary brainwave bands, we transform high dimensional raw data into high value spectral features. Our results demonstrate that in small datasets a) features derived from frequency and time frequency domain allow traditional machine learning models to match or exceed the performance of SOTA deep learning models, b) Attention mechanism is unable to distill the stable feature signatures that characterize healthy neural activity in both resting and task EEGs, and c) the limitations of attention based models in finding relevant spectral features appear to be robust in that providing frequency selective time domain input do not appreciably improve their performance. We validate our methodology across three open source resting EEG datasets and one task EEG dataset, providing robust empirical evidence for our claims.

Modern RL post-training methods such as GRPO and DAPO train on N response sequences of R tokens sampled from a shared prompt of P tokens, but standard FlashAttention replicates all P prompt tokens N times across both forward and backward passes – duplicating compute and memory on identical hidden states. In large-rollout, long-context RL training (N>=16, P>=8K), this redundancy dominates the policy update cost. We observe that in decoder-only models, causal masking makes prompt representations invariant across sequences at every layer, so all per-token operations (norms, projections, MLP) and attention can process the prompt once – a property not yet exploited at the kernel level for training. We propose DualKV, the first FlashAttention kernel variant that eliminates shared-prompt replication during RL training, via (1) fused CUDA forward and backward kernels that iterate over two disjoint KV regions – shared context and per-sequence response – in a single kernel launch, and (2) a data-pipeline redesign in veRL that repacks N(P+R) tokens into P+NR tokens per micro-batch, extending the token reduction from attention to the entire model by a factor rho = N(P+R)/(P+NR). DualKV is mathematically equivalent to standard attention and introduces no approximation. On Qwen3-8B GRPO training with 8xH100 GPUs (N=32, 8K-context), DualKV achieves 1.63–2.09x policy-update speedup, enables 2x larger micro-batches, and raises MFU from 36% to 76%. Similar gains hold for DAPO (2.47x speedup, 77% MFU). At 30B MoE scale on 16xH100, DualKV achieves 3.82x policy-update and 3.38x end-to-end step speedup over FlashAttention (which requires 4-way Ulysses sequence parallelism to avoid OOM). DualKV also extends to hybrid sliding/global attention with head dimension 512 (which FA2 does not support) and integrates with Ulysses sequence parallelism, demonstrated on Gemma-4-31B GRPO at 64K context.

Flow matching trains a neural velocity field by regression against a target velocity associated with a prescribed probability path connecting a simple initial distribution to the data distribution. A central design choice is the path itself. Existing constructions, including rectified and optimal-transport-based paths, transport samples along straight lines between coupled endpoints and thus cover only a narrow class of dynamics. We observe that this corresponds to the simplest case of the least-action principle in classical mechanics, in which the kinetic Lagrangian yields free-particle straight-line trajectories. Building on this observation, we propose Lagrangian flow matching, a physics-based framework in which the probability path and velocity field are determined by minimizing the action of a general Lagrangian subject to the continuity equation and the prescribed endpoints. We show that this dynamic problem admits an equivalent static optimal transport (OT) formulation, yielding a family of simulation-free training objectives that recover OT-based flow matching as the kinetic special case and the trigonometric variance-preserving diffusion path as the harmonic-oscillator case. More general Lagrangians give rise to new probability paths and velocity fields, and numerical experiments show that they induce meaningful changes in the learned dynamics while remaining competitive with existing conditional flow matching models.

In GFlowNets and variational inference, it has been shown that the mean square error between target and model log probabilities is an effective, low variance, surrogate loss for training generative models. This loss has the property that when evaluated on-policy its gradients correspond to those of the KL divergence, while off-policy it remains a valid loss with the same global minimizer. In this work, we demonstrate that this construction can be extended to the whole family of $f$-divergences, leading to a family of losses whose on-policy gradients are that of the corresponding $f$-divergence, but retain the same global minimizer off-policy. Specifically, we show that the on-policy gradients lead to a one to one correspondence between translation invariant loss functions on the target and model log probabilities, and $f$-divergences. This equivalence allows us to design new surrogate loss functions for tuning a wide class of generative models that inherit the properties of the corresponding $f$-divergence, such as being more mode covering, whilst being applicable to off-policy data. We apply our losses on a range of tasks, including classic synthetic examples, SynFlowNets for molecule discovery, and asynchronous large language model (LLM) tuning, demonstrating that our models retain their predicted properties on- and off-policy in a wide class of generative models.

Jung et al. (2025) introduce a hypothesis testing framework for guaranteeing agreement between large language models (LLMs) and human judgments, relying on the assumption that the model's estimated confidence is monotonic with respect to human-disagreement risk. In practice, however, this assumption may be violated, and the generalization behavior of the confidence estimator is not explicitly analyzed. We mitigate these issues by learning a dedicated confidence estimator instead of relying on heuristic confidence signals. Our approach leverages simulated annotator diversity and a margin-based ranking formulation to explicitly model how confidently an LLM distinguishes between human-agreement and human-disagreement cases. We further derive generalization guarantees for this estimator, revealing a margin-dependent trade-off that informs the design of an adaptive estimator training procedure. When integrated into fixed-sequence testing, the learned confidence estimator yields improved ranking accuracy and empirically strengthens the monotonic relationship between confidence and disagreement risk, leading to higher success rates in satisfying target agreement levels across multiple datasets and judge models.

Despite the remarkable success of transformer architectures in natural language processing, their scalability limitations remain poorly understood through systematic empirical analysis. This paper presents the first comprehensive large-scale evaluation of 118 transformer models across seven distinct architectural categories, revealing fundamental performance walls that manifest as hard deployment constraints. Our systematic benchmarking methodology uncovers a critical scalability crisis: while 88.1% of models successfully process sequences up to 512 tokens, this drops dramatically to 44.9% at 1024 tokens, with complete failure (0%) at 2048 tokens. Through rigorous analysis of loading times, memory consumption, and computational efficiency across sequence lengths from 128 to 2048 tokens, we demonstrate that compressed models achieve superior parameter efficiency (649.2 tokens/sec/M parameters) compared to large generative models (12.5 tokens/sec/M). Our findings challenge prevailing scaling assumptions and provide the first quantitative evidence that the theoretical O(n2) attention complexity translates into measurable performance walls. This work establishes new benchmarking methodologies for transformer evaluation and provides critical insights for practical deployment decisions in production environments.

We study contextual dynamic pricing in a semiparametric scalar-index valuation model where the latent value is $v_t=\mu_\ast(\mathsf c_t)+\xi_t$, with an unknown utility map $\mu_\ast$ and an unknown additive noise distribution. The key decision object is the one-dimensional oracle price map $u\mapsto p^\ast(u)$ induced by the scalar index $u=\mu_\ast(\mathsf c)$ and the noise tail. Under the $\beta$-Hölder smoothness of the tail function for $\beta\geq 2$ and a revenue-geometry condition that gives a unique, stable, interior maximizer, this oracle map is itself $(\beta-1)$-smooth. We exploit such structure through $\mathsf{ORBIT}$, a modular coarse-to-fine policy that takes a scalar pilot index as input, localizes a benchmark price in each active bin, and learns a local polynomial approximation of the oracle map inside a trust region via bandit convex optimization. For the baseline linear utility model $\mu_\ast(\mathsf c)=\mathsf c^\top\theta_\ast$, an adaptive elliptical exploration scheme constructs the required scalar pilot online without distributional assumptions on the contexts. The resulting policy achieves regret $\widetilde{O}\big(T^{\frac{2\beta-1}{4\beta-3}}+\sqrt{dT}\big)$. For fixed $d$, we establish a matching lower bound in the horizon dependence, unveiling that the nonparametric oracle-map learning term is minimax sharp. The same scalar-pilot interface also yields extensions to sparse high-dimensional linear utility and nonparametric Hölder utility.

Adaptive Non-local Observables (ANOs) have shown that making quantum observables dynamic can substantially enlarge the function space of Variational Quantum Algorithms, partly shifting hardware demands from circuit synthesis to measurement design. However, this advantage is accompanied by a steep increase in the number of parameters, as well as the classical optimization cost for varying general Hermitian observables. We propose a special form of ANO that significantly reduces this burden by considering only diagonal observables paired with quantum circuits. Mathematically, this is equivalent to the full ANO of a large parameter space since diagonal matrices are canonical representatives of the ANO space modulo unitary similarity. As a result, Diagonal ANO retains the same capability of full ANO while reducing $k$-local observable complexity from $O(4^k)$ to $O(2^k)$ and lowering the corresponding measurement-side classical computation. In this sense, diagonal ANO preserves much of the benefit of full ANO while encompassing conventional VQCs as a special case.

Mixture-of-Experts (MoE) models rely on balanced expert utilization to fully realize their scalability. However, existing load-balancing methods are largely heuristic and operate on noisy mini-batch assignment statistics, introducing bias relative to population-level objectives. We propose $\phi$-balancing, a principled framework that directly targets population-level expert balance by minimizing a strictly convex, symmetric, and differentiable potential of the expected routing distribution. Using convex duality, we derive an equivalent min-max formulation and obtain a simple online algorithm via mirror descent, yielding an efficient EMA-based routing adjustment with negligible overhead. Across large-scale pretraining and downstream fine-tuning, $\phi$-balancing consistently outperforms prior Switch-style and loss-free baselines, demonstrating more stable and effective expert utilization.

Neural surrogate solvers of partial differential equations (PDEs) promise dramatic speedups over numerical methods, especially in scenarios requiring many solves. However, current accuracy-based evaluations do not fully consider two central issues: (1) neural solvers incur substantial up-front costs for data generation, training, and tuning; and (2) classical solvers can also generate low-fidelity solutions at a sufficiently low simulation cost. To explicitly account for these realities and fully incorporate end-to-end costs, we propose an evaluation framework centered on breakeven complexity, a metric that counts the forward solves before a learned solver is cost-effective relative to an error-equivalent traditional solver. To evaluate this measure, we apply scaling laws to determine how much training budget to allocate to data generation and discuss how to achieve smooth error-matching in diverse settings. We evaluate the breakeven complexity of multiple neural PDE solvers on three PDEs on 2D periodic domains from APEBench and a novel benchmark of flows past multiple obstacles generated by the GPU-native PyFR code. Among other findings, our results suggest that neural PDE solvers become more effective as problems get harder in terms of cost, dimension, rollout, physics regime (e.g. higher Reynolds number), etc.

Joint-embedding predictive architectures (JEPAs) propose that a model should learn more useful abstractions when trained to predict latent representations rather than observed outputs. For autoregressive language-model fine-tuning the principle entails a stricter requirement: the induced hidden-state geometry must reach the language-model head and improve the decoded task metric. We test that requirement under a fixed Llama-3.2-1B-Instruct LoRA harness on natural-language-to-regex generation, comparing twenty-two training-time auxiliaries across trajectory-shape regularisation, distributional constraints, predictor/target asymmetry, Fisher-metric Jacobi residuals, and a decoder-visible JEPA objective constructed to lie in cross-entropy's positive cone. The empirical answer is a structured null: several auxiliaries clear single-cell paired $\alpha = 0.10$ without correction (T3-Local at $\Delta = +2.53$~pp, $p = 0.003$ being the strongest), but none survives Bonferroni or Holm–Bonferroni at the relevant family-wise threshold, even though many change curvature, anisotropy, variance, and gradient direction. Decoder-visible JEPA yields the first positive auxiliary–cross-entropy gradient cosine in the study, yet exact match remains inside seed noise; a full-fine-tuning replication of the same auxiliary at $n = 5$ seeds reproduces the null on both benchmarks (TURK: $\Delta = +0.04$~pp, $p_{\text{paired}} = 0.96$; SYNTH: $\Delta = +0.52$~pp, $p_{\text{paired}} = 0.28$), so the null is robust across LoRA and full fine-tuning for the decoder-visible construction. Hidden-state representation work and decoded-task accuracy are therefore weakly coupled in this regime; we accordingly reframe LLM-domain JEPA evaluation as a coupling problem, in which the operative question is under which metrics useful hidden geometry becomes decoder-visible task signal.

As large language models (LLMs) are increasingly deployed to perform tasks with minimal human oversight, it is crucial that these models operate robustly. In particular, a model that can solve a given problem should not fail simply because certain entities$\unicode{x2013}$such as names, numbers, or other contextual details$\unicode{x2013}$have changed while the underlying problem logic remains the same. Prior work suggests that current LLMs still struggle with this form of robustness: they often succeed on some variations of a problem but fail on others. However, existing evaluations often lack a systematic way to identify which logic-preserving variations are most likely to induce failure. Instead, they typically test a random subset of allowable variations, which can overstate robustness. To address this gap, we introduce logic-preserving difficulty scaling (LPDS), a framework that (i) quantifies the difficulty of a problem variation and (ii) systematically searches the space of allowable variations to find those that maximize difficulty and expose failures. We show that as difficulty increases, performance declines and errors in the models' reasoning chains become more pronounced. We further demonstrate that LPDS efficiently finds difficult problem variations for a model, resulting in performance drops up to 5 times larger compared to random sampling. Finally, we show that fine-tuning on more difficult variations leads to more consistent robustness gains than training on easier ones.

Stochastic estimators are fundamental to large-scale optimization, where population quantities must be inferred from noisy oracle observations. Although influential methods such as momentum, SPIDER, STORM, and PAGE have been highly successful, their analyses are largely estimator-specific and expectation-based, obscuring the structural tradeoffs that determine reliability. In this paper, we develop a unified framework for stochastic variance-reduced estimation based on a recursion with three components: memory retention, reset probability, and a correction term for iterate movement. This framework recovers several classical estimators, motivates new second-order variants, and yields a bias-variance decomposition of estimation error. Our main result is a unified high-probability bound proved using a new dimension-free vector-valued Freedman inequality, valid for smooth normed spaces involving random sums of vector martingales. The result applies in both Euclidean and non-Euclidean settings, including the analysis of mirror-descent-based methods in Banach spaces. As applications, we obtain high-probability oracle complexities for unconstrained optimization with mirror descent, establishing the logarithmic dependence on the confidence level. We also derive the first $\tilde{\mathcal{O}}(\varepsilon^{-3})$ oracle-complexity bounds for stochastic optimization with expectation constraints, improving upon the existing $\tilde{\mathcal{O}}(\varepsilon^{-4})$ complexity by leveraging variance-reduced estimation for the first time in this setting.

Memory plays a central role in enabling large language models (LLMs) to operate over sequential tasks by accumulating and reusing experience over time. However, existing evaluations of LLM memory mostly rely on aggregate metrics such as final hold-out accuracy or cumulative online performance, which can obscure critical failure modes such as forgetting and negative transfer. In this paper, we introduce SeqMem-Eval, a diagnostic evaluation framework for sequentially evolving LLM memory. Drawing inspiration from continual learning, it targets a test-time setting in which memory is external, prompt-mediated, and updated without modifying model parameters. Rather than focusing only on final performance, SeqMem-Eval evaluates how memory states evolve, generalize, consolidate experience, and retain useful information during sequential inference. Specifically, it measures online utility, hold-out generalization, backward transfer, and forgetting, providing a finer-grained view of memory quality. Through extensive experiments across diverse tasks and memory methods, we show that higher final or cumulative accuracy does not necessarily imply better memory quality: many methods exhibit strong performance gains while suffering from substantial forgetting or negative transfer. Moreover, different memory designs exhibit distinct trade-offs between adaptability and stability that remain invisible under standard evaluation metrics.

Accurate salt-body delineation is essential for seismic interpretation because salt structures distort wave propagation, complicate velocity-model building, obscure reservoir geometry, and increase uncertainty in exploration and drilling decisions. Although hybrid quantum-classical models have shown competitive performance on small-scale image-classification tasks, their value for dense, pixel-level geophysical prediction remains largely untested. This work introduces quantum feature gating, a hybrid segmentation architecture that embeds a parameterized quantum circuit (PQC) at feature-fusion points within an encoder-decoder pipeline. A 4-qubit, 2-layer PQC with data re-uploading computes a learned convex combination of lateral and top-down features at each Feature Pyramid Network merge point. A global-average-pooling layer maps encoder features to a fixed 4-dimensional quantum input, decoupling the 72-parameter quantum budget from backbone size and image resolution. The method is evaluated on the 2018 TGS Salt Identification Challenge using 4,000 seismic images at 101 x 101 resolution, across two integration topologies, eight circuit variants, and six encoders with 8M to 118M parameters under five-fold cross-validation. In a controlled EfficientNetV2-L ablation at 256 x 256 resolution, replacing the three Quantum FPN Gates with element-wise addition while holding the encoder, loss schedule, splits, and threshold search fixed reduces mean IoU from 0.9389 to 0.8404, a 9.85 percentage-point gap. Inserting the same circuit as skip-connection attention in a custom U-Net improves IoU by 0.88 points over the SolidUNet baseline, showing that the PQC contribution depends on where and what it gates. These results provide controlled evidence that quantum feature fusion can improve dense seismic segmentation.

Group-convolutional neural networks (GCNNs) are among the most important methods for introducing symmetry as an inductive bias in deep learning: In each linear layer, GCNNs sample a transformation group $G$ densely and correlate data and filters in different poses (with suitable anti-aliasing for steerable GCNNs) to maintain equivariance with respect to $G$. Unfortunately, applying filters to many data items resulting from this sampling is expensive (even for translations alone, i.e., in ordinary CNNs), and costs grow exponentially with increasing degrees of freedom (such as translations and rotations in 3D), which often hinders practical applications. In this paper, we propose sampling in feature space, i.e., replacing geometrically dense samples with representative samples selected by feature similarity. This decouples geometric resolution from memory and processing costs during training and inference, providing a novel way to trade off computational effort and accuracy. Our main empirical finding is that a coarse feature-space sampling already preserves classification accuracy remarkably well, which permits precomputation based on geometric similarity, accelerating the training of equivariant 3D classifiers substantially.

Accurate forecasting of residual Physical Resource Blocks (PRBs) is critical for proactive network slice provisioning, energy-efficient operation, and spectrum-aware decision making in cellular systems, where residual PRBs serve as a practical proxy for short- and medium-term spectrum availability. Existing PRB prediction methods typically rely only on historical PRB values and are trained independently per carrier or sector, limiting their ability to capture cross-carrier dependencies and providing no measure of forecast uncertainty. Moreover, point forecasts alone are insufficient for robust spectrum-aware control under highly variable traffic conditions. This paper proposes PRB-RUPFormer, a recursive unified probabilistic Transformer for residual PRB forecasting. The proposed model jointly processes multivariate KPI time series using temporal, seasonal, and carrier-aware embeddings, preserving inter-metric temporal coupling during recursive rollout and stabilizing long-horizon forecasting. A single shared model is trained across all carriers and sectors of an eNB, enabling efficient learning of joint traffic dynamics with low computational overhead. Forecast uncertainty is captured through quantile-based prediction intervals, providing confidence-aware estimates of future PRB availability. Evaluations on six months of commercial LTE network data from multiple U.S. locations demonstrate median MAE below 0.05 and hit probabilities above 0.80 for both one-day and seven-day recursive forecasts. These probabilistic predictions directly support spectrum-aware RAN functions such as dynamic carrier activation, congestion avoidance, and proactive spectrum sharing, making the proposed framework well-suited for dynamic spectrum access scenarios.

Spiking neural networks (SNNs) are biologically inspired energy-efficient models that use sparse binary spike-based communication between neurons, making them attractive for resource-constrained edge devices. Federated learning enables such devices to train collaboratively without sharing raw data. In time-series applications, edge devices often collect data at different time resolutions due to hardware and energy constraints. This temporal heterogeneity poses a fundamental challenge for federated learning: parameters learned at one temporal resolution do not necessarily transfer directly to another, which might result in the naive federated averaging being ineffective. Targeting SNNs and, more broadly, deep networks with stateful neurons, we propose a federated learning framework that addresses this temporal resolution mismatch. We investigate how neuron parameters learned from data at different temporal resolutions and model aggregation should be integrated. We evaluate the proposed framework across two SNN-native benchmark datasets (SHD and DVS-Gesture) under a range of resolution heterogeneity scenarios. Our results show that the proposed adaptation methods can substantially recover accuracy lost due to temporal mismatch, hence enabling each client to train at their local temporal resolution while remaining compatible with the global model.

Despite the success of foundation models in language and vision, molecular graph generation still lacks a unified framework for heterogeneous design tasks with reliable controllability. While reinforcement learning (RL) offers a natural post-training mechanism for task-specific optimization, applying it to graph generative models is hindered by the vast atom-wise action spaces and chemically invalid intermediate states. We propose Controllable Molecular Generative Foundation Models (CoMole), built with a unified motif-aware graph diffusion pipeline. By learning a motif-aware graph space, CoMole transfers pretrained structural priors into controllable generation, where RL optimizes conditional reverse policies over chemically meaningful decisions. We theoretically characterize the bottleneck of atom-level RL and justify motif-aware policy optimization. Across three heterogeneous benchmarks spanning materials and drug discovery, CoMole ranks first in controllability on all nine targets, reduces MAE by up to 48.2% relative to the strongest baselines, and maintains validity above 0.94 without rule-based correction or post-hoc filtering. We further show that CoMole transfers controllability to unseen properties by optimizing only task embeddings with the generator frozen, achieving performance competitive with strong task-specific baselines.

Inferring the structure of directed acyclic graphs (DAGs) from data is a central challenge in causal discovery, particularly in modern high-dimensional settings where large-scale interventional data are increasingly available. While interventional data can improve identifiability, existing methods remain limited by soft acyclicity constraints, leading to optimization over invalid cyclic graphs, numerical instability, and reduced scalability. We introduce PACER (Perturbation-driven Acyclic Causal Edge Recovery), a scalable framework for causal discovery that guarantees acyclicity by construction. PACER parameterizes a distribution over DAGs through a joint model of variable permutations and edge probabilities, enabling direct optimization over valid causal structures without surrogate penalties. The framework supports a unified likelihood-based treatment of observational and interventional data, flexible conditional density models, and the incorporation of structural prior knowledge. For linear-Gaussian mechanisms, we derive closed-form expressions for the expected interventional log-likelihood and its gradients, yielding substantial computational gains. Empirically, PACER matches or exceeds state-of-the-art methods on protein signaling and large-scale genetic perturbation benchmarks, while scaling efficiently to networks with thousands of variables and achieving up to two orders of magnitude speedups over penalty-based differentiable approaches. These results demonstrate that exact and scalable causal discovery from high-dimensional perturbation data is achievable through principled search space design.

Stochastic compositional optimization minimizes $F(f(x), x)$ over a compact convex set when the inner map $f$ is available only through noisy samples. We propose Hybrid Momentum Stochastic Frank–Wolfe, a projection-free method for outer functions $F$ that are jointly convex and Lipschitz but not necessarily differentiable, such as maxima, CVaR, and norms. The method tracks the Jacobian and the value of $f$ with momentum, using a Taylor correction for the value, and passes the tracked linearization to a generalized linear minimization oracle. Under finite-variance noise, it attains an expected generalized Frank–Wolfe gap of order $K^{-1/4}$ for non-convex objectives and, when the linearized models minorize the objective, a last-iterate primal error of order $K^{-1/3}$. The analysis extends to heavy-tailed noise without clipping, with separate moment orders for values and Jacobians. With an exact oracle, the method reduces to the deterministic method of Vladarean et al. (2023).

LLM-based agents are increasingly used to simulate deliberative interactions such as negotiation, conflict resolution, and multi-turn opinion exchange. Yet generated transcripts often do not reveal why an agent's stance changes: movement may reflect evidence uptake, anchoring, role drift, echoing, or changed prompt and retrieval context. We introduce the Belief Engine (BE), an auditable belief-update layer that treats "belief" as an evidential state over a proposition and exposes it as scalar stance. BE extracts arguments into structured memory and updates stance with a log-odds rule controlled by evidence uptake u and prior anchoring a. Across multiple base LLMs, parameter sweeps show that these controls reliably shape stance dynamics while preserving an evidence-level update trail. On DEBATE, a human deliberation dataset with pre/post opinions, BE best reconstructs participants whose final stance follows extracted evidence; stable and evidence-opposed cases instead point to anchoring or factors outside the extracted evidence stream. BE provides configurable infrastructure for studying evidence-grounded deliberation, where openness, commitment, convergence, and disagreement can be tied to explicit update assumptions rather than hidden prompt effects.

Video reasoning models are a core component of egocentric and embodied agents. However, standard benchmarks for assessing models provide only evaluation of the output (e.g. the answer to a question), without evaluation of intermediate reasoning steps, and most provide answers only in the text domain. We introduce Minerva-Ego, a benchmark for evaluating complex egocentric visual reasoning. We extend recent high-quality video data sources recorded from egocentric / embodied settings with a set of challenging, multi-step multimodal questions and spatiotemporally-dense human-annotated reasoning traces. Benchmarking experiments show that state-of-the-art models still have a large gap to human performance. To investigate this gap in detail, we annotate each reasoning trace in the dataset with the objects of interest required to solve the question, as spatiotemporal mask annotations. Through extensive evaluations, we identify that prompting frontier models with hints of 'where' and 'when' to look yields substantial improvements in performance. Minerva-Ego can be downloaded at https://github.com/google-deepmind/neptune.

Autonomous labs increasingly deploy large language models (LLMs) as closedloop biology optimizers that propose a design, read the outcome, then propose again. This is partly because of an assumption that biological knowledge helps an LLM choose better designs than Bayesian optimization, which has no access to biological context. We test that assumption on 45 biology tasks reconstructed from 2,485 experimental rows across 341 published papers, scored by oracles fitted to those rows, each run over 30 iterations under two prompt conditions. The domain-aware condition provides the published parameter names, units, and labels, and the domain-agnostic renames them so they are not biologically meaningful, with search space and feedback identical. Across 19 LLMs domain-agnostic outperforms domain-aware on 58.7% of tasks by trajectory and 61.9% by endpoint. The failure is specific to biology, reversing on education tasks built the same way. Domain-aware prompting tends to help models propose better initial designs, while making them respond worse to feedback. Part of the reason is anchoring, with models keeping the conventional categorical choice even when the oracle scores a different (published best) value higher. Post-training Llama 3.1 8B to frontier-level raw optimization does not transfer to biology-aware optimization improvements. Biology-aware optimization therefore looks like a separate capability that autonomous laboratory loops should target directly.