Papers

Enterprise AI agents in production often need to be bounded, stateful, observable, and governable rather than fully autonomous. We present PinSieve, a production case study in a large-scale content-quality pipeline. Its deployed component is a selective vision-language-model (VLM) Serving Agent that operates only on the grey-zone slice left unresolved by lightweight upstream models, exposes a scalar routing score online, and preserves controlled human escalation. On this slice, the deployed system filters 2.05x more non-actionable items than the previous production module while slightly reducing estimated miss rate; after promotion, it improves review productivity by 25.7%, reduces normalized operating cost by 16.2%, and moves signal delivery from next-day to same-day. We then study maintenance through a governed memory flywheel under selective feedback, where escalated items are reviewed by default and auto-passed items are labeled mainly through audit sampling. Feedback Memory records routing traces, observation paths, audit propensities, and replay metadata for evaluation and debugging. The Data Curation Agent uses a bounded proposal-verifier loop over representative, uncertainty, recency, and fresh-review replay, with positive-rate and score-bin guardrails before batch acceptance. In chained monthly refresh over six months of production data, this design reduces average FNR@50% from 17.73% under representative random replay to 13.29%. A Reasoning Review Agent audits teacher-generated rationales and supports keep/repair/drop decisions. Production claims are attributed only to the deployed Serving Agent; replay and rationale-review results are offline or sampled-governance evidence. The same serving-agent recipe has been adopted to several additional internal signals, suggesting transferability beyond one task.

Time series classification underpins applications in healthcare, sensing, and industrial monitoring. Although time series foundation models support forecasting and transferable representation learning, classification still typically requires fitting a task-specific classifier on each target dataset, while individual channels of multivariate inputs are often encoded independently. We introduce ChorusTIC, a classification-native foundation model for in-context classification across heterogeneous channel configurations without target-task parameter updates. ChorusTIC combines episode-consistent Random Subchannel Slot Concatenation with a shared dual-axis encoder to model temporal and cross-channel interactions and map variable channel configurations into a fixed-width representation independent of the original channel count. It then calibrates feature axes using context-derived distributions and predicts query labels through leakage-protected in-context learning. We pretrain ChorusTIC solely on synthetic labeled episodes comprising context and query sets that share a task background, with classes distinguished by sparse temporal or cross-channel rules. Evaluations on the complete UEA-30 and UCR-128 archives show strong full-context and low-label performance without target-specific classifier fitting.

Understanding how neural networks learn and organize features is central to understanding their behavior. Much existing theory of feature learning has focused on the emergence of a global low-dimensional representation. We show that this picture is incomplete. In regression problems with clustered data, we demonstrate that multilayer perceptrons (MLPs) naturally develop monosemantic specialized neurons: individual neurons become strongly aligned with a specific predictive feature relevant to a particular region of the input space. Rather than learning a single global low-dimensional representation, MLPs learn a collection of local low-dimensional representations. We show that this ability to specialize gives MLPs a provable data-efficiency advantage over feature-learning methods based on a global low-dimensional representation.

Virtual sensing enables digital twins and safety-critical systems to reconstruct and forecast spatial-temporal physics in real time. However, conventional computational and data-driven methods often face challenges in generalization, latency, and energy efficiency for edge deployment. Neural operators offer a promising alternative but remain reliant on power-intensive hardware. Spiking neurons and neuromorphic computing can improve efficiency, yet surrogate-gradient training and multi-step spiking introduce convergence and latency challenges. We propose the Sparse-Activation-ReLU (SAR) layer, a single-step alternative that promotes activation sparsity without surrogate-gradient training while remaining compatible with event-based computing. Within a trunk-based NOMAD architecture, SAR achieves over a fivefold improvement in the combined Latency-Error-Energy (LEE) metric compared with Variable Spiking Neuron (VSN) and Leaky Integrate-and-Fire (LIF) implementations. We further analyze spiking entropy and feature usage and introduce synthetic knowledge distillation, reducing the LEE score by more than twofold. Finally, we improve VSN through a ReLU-based spiking loss and graph-neighbor thresholding. On the Heat Exchanger dataset, these approaches reduce L2 error by more than twofold and nearly sevenfold, respectively, while reducing spiking and spatial aggregation. Overall, the work presented is a step towards energy-efficient virtual sensing by providing an alternative framework that can be positioned towards neuromorphic or other edge device integration that can be a gold standard to compare latency, energy, and error performance for future efficient designs that are sparsity or brain-inspired spiking based.

Scientific reasoning requires language models to retrieve specialized knowledge and incorporate it reliably into multi-step computation. Conditional memory provides an explicit lookup pathway that complements dense neural representations, but its usefulness is inherently input- and computation-dependent: retrieved information may repair missing scientific associations, yet it may also introduce distracting shortcuts or interfere with reasoning that the base model can already perform correctly. In this work, we systematically investigate when, where, and to what extent conditional memory should participate in scientific reasoning. We characterize the scientific knowledge boundary and controlled interventions on memory-enabled knowledge-circuit nodes. Based on these analyses, we propose a Knowledge Boundary-Aware Router that uses task-specific input proxies available before generation to determine whether memory is activated, which layer-stage nodes receive memory signals, and how strongly these signals contribute. Experiments on biological and chemical reasoning benchmarks, covering two backbone families and six task types, show that memory effects vary substantially across inputs, tasks, and injection locations. Compared with static and activation-rate-matched random routing, our approach more consistently preserves beneficial memory contributions while suppressing memory-induced regressions, establishing selective memory allocation as an important principle for reliable scientific reasoning.

Multimodal large language models (MLLMs) have made significant progress in understanding and interpreting mul- timedia content. However, their ability to generate me- dia remains limited. Recent approaches have attempted to bridge this gap by translating the hidden representations of token sequences into the embedding space of visual models or directly into raw image data. However, these methods often represent each image using multiple specialised to- kens which significantly increases the input length. This be- comes a major limitation for tasks such as graphic design generation where the output typically involves a seamless blend of thousands of tokens across text, multiple images, and layout information. To address this challenge, a novel architecture is proposed that maps hidden token represen- tations to the embedding space of visual models, such as CLIP ViT-L/14, using a single [IMG] token per image. The architecture employs two shallow MLP blocks, each with a separate compression module followed by a shared expan- sion module, trained with six distinct loss functions. One block aids the other during training and is omitted during inference, resulting in a lightweight solution. Strong perfor- mance is demonstrated in both image-to-design and text-to- design generation tasks.

Retrieval-augmented generation (RAG) improves the factuality of large language models by grounding responses in external documents, but it also exposes a critical security vulnerability: adversarial documents injected into the knowledge database can enter the context window and steer the model toward targeted incorrect answers. Existing post-retrieval defenses rely on instruction following, parametric knowledge, or text-level consistency, all of which can be imitated or optimized against by adaptive attackers. We propose RAGSentinel, a training-free, label-free defense for black-box RAG systems. RAGSentinel uses a surrogate encoder to measure query-conditioned hidden-state shifts induced by retrieved documents, removes shared topic directions, and filters poisoned documents as geometric outliers from a robust majority consensus. We prove that, under an honest-majority assumption and a representation-level separation condition, RAGSentinel exactly recovers a poison-free majority-sized context. Experiments across three question-answering datasets, three LLM families, and multiple poisoning attacks show that RAGSentinel consistently achieves low attack success rates while preserving competitive accuracy and remaining effective against adaptive attacks with full pipeline knowledge.

Safety alignment in large language models (LLMs) remains brittle against a growing spectrum of attacks. Jailbreak attacks bypass safety mechanisms through crafted prompts, while neuron-level attacks directly prune safety-critical neurons post-deployment. Both exploit a common weakness: safety-relevant information concentrates in a sparse neuron subset. We present NeuronGuard, a fine-tuning-stage defense that simultaneously hardens LLMs against both attack classes by redistributing safety signals across a broader set of neurons. NeuronGuard dynamically identifies safety-critical neurons via periodically refreshed per-layer linear classifiers, forces refusal behavior under deliberate neuron ablation, and applies KL-divergence regularization for distributional consistency. A randomized gradient projection strategy preserves downstream task utility by resolving conflicts between the defense and task objectives. We provide a formal guarantee that NeuronGuard strictly reduces the attack success rate (ASR) upper bound, and experiments across three LLMs, six state-of-the-art attack strategies, and multimodal settings confirm near-zero ASR while maintaining task accuracy, including against white-box adaptive adversaries.

Offline reinforcement learning is intrinsically multi-objective: a policy must remain compatible with the behavioral support of a fixed dataset while preferentially selecting high-value actions. We recast these objectives in a common form by viewing each as an action-space motion field that specifies how generated actions should move. This perspective enables heterogeneous learning objectives to be combined directly through field composition. Inspired by drifting models, we propose CoDrift, a compositional framework for one-step generative policy learning. CoDrift combines three objective-level fields into a unified policy field. The conditional field preserves state-dependent behavioral structure, while the marginal field pools actions across states to provide a more stable generative signal in the single-positive-sample regime of continuous-control offline RL. The value field moves generated actions toward higher-value regions. The composed field is absorbed into a stochastic generator that produces an action with a single forward pass at deployment. We evaluate CoDrift on 73 tasks from OGBench and D4RL in both offline and offline-to-online settings. CoDrift compares favorably with state-of-the-art methods and achieves the best average rank in both settings.

Modern score-based generative models have achieved remarkable empirical success in high-dimensional tasks such as image, audio, and video synthesis. These models reduce distribution learning to a sequence of regression problems that, if solved exactly on finite data, would ultimately reproduce the training samples. Their ability to generalize must therefore arise from the implicit or explicit regularization during training. In this work, we develop a generative counterpart to the theory of benign overfitting and algorithmic regularization for overparameterized neural networks in the supervised lazy-training regime. We study denoising score matching in a vector-valued reproducing kernel Hilbert space with an inner-product kernel. In the proportional high-dimensional regime $n\asymp d$, we derive exact risk trajectories under gradient flow training. These trajectories exhibit three phases governed by qualitatively distinct estimators: a spectral estimator that generalizes, a pure-noise score with localized peaks that interpolate the training objective, and an empirical Bayes estimator that memorizes the data. We then analyze how these estimators combine along the reverse-time SDE and characterize the distribution of the resulting samples. The analysis reveals familiar mechanisms from supervised learning, including kernel linearization and self-induced regularization from the nonlinear part of the kernel, but also reveals a distinct phenomenology specific to generative modeling.

We present a dynamical-systems based model for resting-state functional magnetic resonance imaging (rs-fMRI), trained on a dataset of roughly 40K rs-fMRI sequences covering a wide variety of public and available-by-permission datasets. While most existing proposals use transformer backbones, we utilize multi-resolution temporal modeling of the dynamics across parcellated brain regions. We show that MnemoDyn is compute efficient and generalizes very well across diverse populations and scanning protocols. When benchmarked against current state-of-the-art transformer-based approaches, MnemoDyn consistently delivers superior reconstruction quality. Overall, we find that with such large-scale pre-training on (non-proprietary) rs-fMRI datasets, we get a highly performant model for various downstream tasks. Our results also provide evidence of the efficacy of the model on small sample size studies which has implications for neuroimaging studies at large where resting state fMRI is a commonly acquired imaging modality.

Diffusion inference repeatedly evaluates the same large network. We ask whether pretrained weights alone can identify residual branches that need not be recomputed throughout the trajectory. Our Spectral Concentration Ratio (SCR) measures leading-versus-tail singular-value energy. Combined with Frobenius magnitude, it yields an offline sensitivity proxy and a deterministic lifetime for each scheduled unit. A frozen unit reuses its cached residual-branch update while the current residual stream and all external conditioning continue to propagate. The method needs no router, calibration prompts, or input-dependent search. At matched layer-step budgets, SCR/Frobenius preserves quality better than random, depth, norm, stable-rank, and Frobenius–stable-rank schedules on LLaDA-8B, DiT-XL/2, U-ViT-L, and SDXL. Broader LLaDA tests cover retrieval, reasoning, code, summarization, and open-ended generation; matched-horizon controls retain the ranking down to ten denoising steps. The complete captured-graph system reaches $2.8\times$–$3.0\times$ wall-clock speedup over eager inference. This is a systems-level number: on LLaDA, padded graph execution already gives $2.7\times$, while eliminating inactive branch work raises it to $3.0\times$. The perturbation analysis motivates pre-norm attention and MLP components under explicit local assumptions; results on AdaLN, U-shaped, convolutional, and cross-attention blocks are empirical transfer, not certified guarantees.

Static pruning imposes one sparse structure on every prompt, even though reasoning, retrieval, generation, coding, and translation can depend on different parts of a language model. We introduce Task-Aware Spectral Pruning (TASP), a post-training framework that calibrates module-level spectral descriptors against measured task-specific ablation effects, closes grouped-query-attention and SwiGLU dependencies during sparse-mask construction, and routes each user turn to one compiled mask that remains fixed throughout prefill and decoding. A module-disjoint pilot first determines whether the spectral signal is informative before full calibration. Under the stated retrospective operating rule, the pilot passes on the evaluated Llama-3-8B and Llama-3-70B checkpoints but rejects Qwen2.5-1.5B, demonstrating that applicability is model-dependent rather than universal. At a 43% active-FLOP reduction, the Llama-3-70B benchmark harness retains 97.7 +/- 0.2% of the dense BF16 score. In the deployment-matched INT8-weight/BF16-compute runtime on a single A100 80GB, the compiled sparse path retains 97.3 +/- 0.2% relative to dense BF16 and reduces decode latency from 45.2 +/- 0.4 to 31.3 +/- 0.4 ms/token, yielding a 1.44x speedup. Factorized ablations, disjoint-module tests, compiled structured baselines, routing-corruption studies, and an explicit 136-GPU-hour calibration audit further delimit the source and operating regime of these gains

Conformal Bayes combines Bayesian posterior predictive scores with conformal calibration, but under continuous label shift both the score and calibration weight depend on the unknown response-marginal density ratio. Existing methods typically estimate one shift parameter from pseudo-labels or predictive samples and plug it into calibration. We instead propose Joint Tilt-Sensitivity Conformal Bayes (JTS-CB), which performs sensitivity analysis over a prespecified set of plausible tilts; its split-conformal realization is JTS-SCB. Each tilt jointly determines the Bayesian conformal score and conformal importance weight. JTS-SCB forms a bounded sensitivity envelope over candidate tilts, but its calibration-only construction does not inherit the exact finite-sample weighted-conformal guarantee. We therefore study a separate candidate-weighted exact counterpart and show that its usefulness depends sharply on tail behavior. For scalar linear exponential tilts, any nonzero candidate tilt makes the exact set unbounded. More generally, tail-growing density ratios produce the same pathology, whereas quadratic tilts with a negative coefficient on \(y^2\) have vanishing tail weights and admit bounded exact inference on the original target. Ratio clipping provides a complementary bounded exact construction for a surrogate target when tails grow. Experiments show that strong plug-in predictive sampling can match the oracle when the shift is well identified, while sensitivity analysis is most useful for richer, weakly identified, or systematically biased shift models, at the cost of wider prediction sets.

We treat options pricing as a representation problem: can machine learning detect systematic deviations from Black-Scholes using 2.6M real option contracts? We compare three regimes: learned abstract embeddings (Kernel PCA), preserved domain structure (tree-based ensembles), and neural network validation. Tree-based methods outperform kernel dimensionality reduction by 21.5 percentage points (93.8% vs 72.3%), and domain-expert features (Greeks, moneyness) outperform engineered features. NN-based and BS-based deviation labels agree 99.9974% of the time, suggesting deviations reflect market structure rather than model artifact. We conclude that in domains with expert-designed symbolic features, preserving structure beats learning abstractions. We make no claim of exploitable mispricings.

Central bank press conferences are not merely information releases — they are structured narratives. We study whether the shape of sentiment within a statement, not just its average tone, carries policy-relevant signals. Constructing sentiment arcs for ECB and Fed press conferences along three dimensions — monetary stance, economic outlook, and uncertainty — we assess their predictive content for policy rate changes, inflation expectations, and forecaster disagreement. Our findings show that arc shape robustly predicts rate decisions beyond lexicon-based benchmarks at both institutions — it is not merely whether a statement sounds hawkish or economically optimistic on average, but how these sentiments are sequenced and emphasized across the statement, that carries the policy signal. Arc features also shape how professional forecasters update inflation expectations and how much they disagree, pointing to a receiver-side effect distinct from the direct policy signal. These findings suggest that communication design — the sequencing and emphasis of policy language across a statement — is a first-order feature of the policy signal, not a second-order refinement.

Text anomaly detection, the task of identifying text instances that deviate from normal language patterns, is crucial for language-driven applications. However, most existing methods can only perform document-level anomaly detection, making it hard to locate harmful phrases or support targeted prevention. Recently, there has been an emerging trend toward token-level text anomaly detection, which aims to address the above limitation by identifying anomalous words or fragments within a document. Nevertheless, one representative method mainly relies on representation-space distance measurement, neglecting the complementary roles of different anomaly cues in capturing diverse abnormal patterns. To bridge the gaps, we propose a Dual-evidence framework with adaptive Fusion and Aggregation (DiFA) for token-level anomaly detection. DiFA derives anomaly scores from form-structural and semantic views to capture visible structural abnormality and contextual inconsistency, respectively, thereby providing complementary evidence for identifying diverse anomalies. To combine these two scores with varying numerical scales, DiFA incorporates a calibration and fusion mechanism to adaptively balance the two views. Moreover, to obtain a discriminative document-level score, a multivariate aggregation method is designed to summarize token-level anomaly scores from multiple perspectives, preventing rare anomalous tokens from being diluted. Extensive experiments across various text anomaly detection benchmarks demonstrate that DiFA consistently achieves top performance while maintaining strong efficiency, robustness, and interpretability. The code and scripts are available at: https://github.com/qyy11-com/DiFA.

Walsh-sequence-based receiver architectures offer an alternative approach for the reception and reconstruction of multiple simultaneous RF signals over wide bandwidths. While previous works have focused on the architecture and theoretical operation of Walsh-domain processing, limited experimental results have been reported so far on hardware operation under realistic conditions. This paper presents the experimental evaluation of a Walsh-sequence-based receiver prototype. A measurement campaign was conducted under various signal placement, interference, and receiver configuration scenarios. Reconstructed spectra, Walsh-lane power spectral densities (PSDs), and Error Vector Magnitude (EVM) measurements were used to investigate Walsh-domain signal representation, interference effects, and receiver performance. Finally, dense- spectrum experiments involving up to ten simultaneously active signals demonstrate the potential of the architecture for low power wideband spectrum sensing applications.

Wireless embedded systems increasingly rely on wireless channel information for decision making, yet practical platforms operate under severe constraints, including few antennas, narrow bandwidth, and sparse, noisy measurements. While neural field based approaches inspired by Neural Radiance Fields (NeRFs) have recently been explored for continuous wireless channel modeling, existing approaches depend on dense measurements or external priors such as known geometry, visual context, or angle-of-arrival (AoA) information, limiting their practicality in real-world deployments. We present WiNeRF, a neural field framework that learns a spatially continuous, complex-valued wireless channel representation directly from sparse channel state information (CSI) collected by commodity WiFi devices. WiNeRF embeds intrinsic system constraints, such as antenna geometry, limited spatial resolution, and phase uncertainty, as inductive biases through a 3D conical wave sampling model, a multi-resolution implicit scene representation, and a differentiable optimization framework for complex-valued channel learning. Across diverse indoor environments with non-line-of-sight (NLoS) regions, WiNeRF achieves a median prediction SNR of 5.3 dB, outperforming prior neural baselines by 4.9 dB on average (approximately 3x higher prediction SNR), and produces a task-agnostic channel representation that can be directly reused in standard signal-processing pipelines, including beamforming, AoA estimation, and RSSI coverage mapping, without modifying existing hardware or wireless protocols.

We consider a federated reinforcement learning setting involving $M$ agents, all of whom interact with a common Markov Decision Process (MDP). The agents exchange information via a central server to learn the optimal value function. Our goal is to understand to what extent one can hope for collaborative sample-complexity speedups in such a setting, when a small fraction of the agents are adversarial and can act arbitrarily. To that end, we propose Robust Fed-Q}, a federated Q-learning algorithm that blends ideas from both model-based and model-free RL, along with the median-of-means device from robust statistics. We prove that despite corruption, with high-probability, Robust Fed-Q (i) guarantees exact convergence to the optimal value function in the limit of infinite samples, and (ii) enjoys near-optimal finite-time rates that benefit from collaboration. In addition, our approach requires just $\tilde{O}(1)$ rounds of communication to achieve each of the above guarantees, a feature of independent interest in FL where communication is the major bottleneck.

Deploying vision perception models in autonomous vehicles requires that we prioritize inference speeds, resulting in a model with shallower architectures and lesser model parameters (i.e., more pruned). Such small models do not generalize well, which could result in poor performance when encountered with novel scenarios. We propose a system that overcomes this by continuously retraining the vision models on the cloud with data uploaded by vehicles. We leverage the abundant compute resources, including machine learning accelerators, of the cloud to run a highly-accurate oracle model that will guide the retraining process of the on-vehicle model. This newly trained model is transmitted to the vehicle over the network and is utilized by the vehicle for perceptions, leading to improved inference accuracy over time.

Autonomous vehicles operate in dynamic, ever-changing environments where new scenarios and edge cases constantly emerge. As a result, static learning models are inadequate for ensuring safe and reliable operation. Continuous learning is essential for adapting to these evolving conditions and maintaining robust performance across diverse real-world settings. However, autonomous vehicles generate massive streams of visual data during operation, and existing continuous learning approaches typically rely on heuristic sampling methods that fail to capture temporal dynamics, often overlooking critical learning opportunities or selecting redundant frames. In this paper, we introduce FrameScope, a temporal data valuation framework for continuous learning in autonomous vehicles. FrameScope extends neural tangent kernel theory to temporal domains, enabling principled valuation of streaming visual data. Unlike cloud-centric methods that transmit all video data for processing, our approach performs principled, local frame selection on the vehicle and queries a cloud-based oracle model only for labels of those high-value frames. Extensive experiments across multiple domain shifts show that FrameScope consistently outperforms existing methods, achieving higher sample efficiency and significantly reducing catastrophic forgetting in autonomous vehicle perception. By valuing data on the vehicle and querying only labels for selected frames, FrameScope reduces bandwidth requirements, enabling scalable operation with a lightweight cloud labeling service.

Protein fitness landscapes contain nonlinear interactions in which mutation effects depend on other residues. We introduce ORBIT, an Order-Resolved Benchmarking of Interaction Transformations framework that separates interaction presence, representation accessibility, and functional recovery. ORBIT first validates Walsh-based diagnostics on synthetic landscapes with known interaction order, then analyzes the experimentally measured GB1 fitness landscape under the FLIP 2-vs-rest setting. We compare ridge regression, a standard MLP, independent tokens, nonlinear independent tokens, and Residual Interaction Tokenization (RIT). Across 20 paired training seeds, the primary two-hidden-layer comparison found no significant architecture differences in FLIP test R^2, third- or fourth-order functional recovery, or final-layer third- or fourth-order accessibility. However, RIT significantly increased pairwise accessibility at the token stage relative to both independent-token controls (Delta A_tok,2 = 0.2468, d_z = 1.67, Holm-adjusted p = 1.14 x 10^-5), without a detectable downstream higher-order advantage. A pre-specified depth/capacity analysis showed that deeper MLPs improved FLIP prediction, third-order functional recovery, and final-layer third-order accessibility; fourth-order accessibility also improved relative to the shallow MLP but remained below zero in absolute held-out R^2. ORBIT therefore reveals representation-level changes hidden by conventional prediction metrics and distinguishes early interaction-aware encoding from higher-order structure constructed by downstream nonlinear capacity.

Systematic dialectal performance gaps in language models (LMs) are well documented, but the source of these disparities within the modern language modeling pipeline remains unclear. Our study traces this "dialect tax" across the natural language processing pipeline. Using parallel English dialect corpora that hold meaning fixed while varying surface form, we first confirm that LMs recognize matched Standard American English (SAE) and dialectal texts as semantically equivalent. However, we discover further representational gaps corresponding to downstream performance gaps. Across model families and generations, modern LMs still encode dialectal texts unequally during tokenization, pre-training, post-training, and inference. Strikingly, bypassing traditional subword segmentation via a character-level counterfactual tokenizer removes neither input and output asymmetries nor dialectal accuracy gaps. During pre-training, dialect pairs induce more divergent gradient updates than pairs of entirely unrelated SAE documents, indicating that models find semantically equivalent dialectal content harder to learn from than unrelated SAE documents. During post-training, reward models show contextual, unstable dialect preferences, assigning higher values to isolated AAVE-exclusive tokens than to SAE-exclusive tokens, while full reasoning contexts receive task- and model-dependent dialect penalties. Overall, our findings suggest that the dialect tax is encoded and accumulated not by any one step in isolation, but at every step of the language modeling process.

Reinforcement learning (RL) post-training has emerged as a powerful framework for enhancing the capabilities of large language models (LLMs), enabling impressive reasoning, math, and coding capabilities. Yet for many researchers and practitioners, the principles behind classical RL remain a "black box". In this work, we deconstruct the RL post-training algorithm, investigating each step to clarify what is actually happening beneath the surface. By isolating the mechanics of RL with Verifiable Rewards in a controlled and simplified environment, we examine how RL outcomes are shaped by the base model's prior distribution, the granularity of the reward signal, the diversity of the prompt distribution, and model scale. We use the entropy of the policy's output distribution as a lens to compare the distributions learned through pretraining, SFT, and RL post-training, revealing how each stage shapes model certainty. Our investigation sheds light on how these choices interact to affect post-training success. For example, we show that the effect of so-called 'spurious rewards' depends on the prompt distribution used for post-training. We also provide insight into why the success of RL post-training depends on whether the base model already places sufficient probability mass on the desired behavior, linking it to the classical concept of exploration in RL. Ultimately, we provide this primer as a resource to those in the NLP community wishing to incorporate RL as a tool in their toolbox.

End-to-end training of multimodal neural networks often exhibits unstable neural dynamics characterized by three coupled failure modes that degrade learning: (i) modality imbalance, where one branch dominates gradient-based optimization; (ii) unstable gating, where noisy confidence cues induce erratic modality selection; and (iii) fusion interference, where modality-specific gradients conflict at the shared fusion layer. We propose CAT-GS (Calibrated, Adaptive, Thresholded Gating with Fusion Surgery), a neural dynamics-based optimization controller for intelligent computing applications. CAT-GS operates during backpropagation without modifying model architectures, fusion modules, or task losses. Through calibration of teacher-derived reliability via temperature scaling and EMA smoothing, CAT-GS stabilizes neural dynamics using a margin-thresholded policy to switch between warm-up dropout, weak-modality prioritization, and weak-biased blending, stabilizes gradient magnitudes under aggressive gating via capped gradient-budget renormalization, and applies fusion-only PCGrad to reduce destructive cross-modal interference at the primary shared bottleneck. We evaluate CAT-GS on audio–visual multimodal pattern recognition benchmarks (CREMA-D, AV-MNIST, and VGGSound), a tri-modal setting (UR-FUNNY), controlled synthetic data (CG-MNIST), and additional cross-domain benchmarks (AVE and CMU-MOSI). CAT-GS improves or matches fused multimodal accuracy against strong imbalance-aware baselines (including OGM-GE, G$^2$D, and UMT) across settings, and yields smoother gating behavior with fewer conflicting fusion gradients.

Macros constitute a large part of the core area in modern very large-scale integration (VLSI) designs. Moreover, macro positions have a significant impact on the final quality of result (QoR), and macro legalization is typically the final step in determining the macro positions. However, existing approaches related to macro legalization either lack robustness or incur substantial computational costs or neglect the regularity between macros. To address these limitations, we introduce MacroAgent. The novel framework is a four-stage approach: clustering, contour generation, template matching, and inter-cluster refinement. We propose leveraging Large Language Models (LLMs) to discover multiple, effective heuristic regularity-aware contour algorithms. This framework successfully generates robust and effective algorithmic solutions for macro legalization. Compared with state-of-the-art macro legalization works, experimental results on TILOS and Chipyard benchmarks demonstrate a 2 to 8 fold improvement in layout regularity, a 3% to 5% reduction in routed wirelength with comparable congestion after global routing, and significantly better robustness with an acceptable runtime. Furthermore, end-to-end evaluation through Cadence Innovus place-and-route confirms that the regularity improvements translate into tangible PPA gains, including 2.9% lower routed wirelength and 68.3% TNS improvement over the DREAMPlace macro legalization baseline; it also achieves 1.8% lower routed wirelength when integrated into the Innovus macro placement flow.

Recent years have witnessed remarkable achievements of Large Language Models (LLMs) in multiple domains, while the excessive resource requirements of LLMs hinder the deployment on resource-constrained devices. Although model quantization stands out as an effective approach, conventional quantization approaches typically incur severe performance degradation due to uniform bit-width or simple heuristic sensitivity evaluation. In this paper, we propose a novel Fisher information-based Adaptive Mixed Precision Weight Quantization approach, i.e., FAMPWQ, which performs layer-adaptive weight quantization for effective LLM inference on commodity GPUs. First, we propose a system model with a novel Fisher information metric to measure the layer-wise sensitivity to quantization. Second, we propose a reinforcement learning-based bit-width allocator in FAMPWQ, which generates an adaptive bit-width allocation strategy based on the Fisher information sensitivity metric. Extensive experiments on 7 models and 5 benchmarks demonstrate that FAMPWQ significantly outperforms 7 baseline approaches in terms of PPL (up to 3.39 smaller), accuracy (up to 6.87% higher), and LLM-as-a-judge comparison (up to 76% win rate).

Partial differential equations (PDEs) often have high-frequency and multi-scale features that neural networks struggle to approximate. Physics-Informed Neural Networks (PINNs) build the governing equations directly into training, but suffer from spectral bias: they learn low-frequency components faster than high-frequency ones. Techniques such as Fourier feature embeddings and sinusoidal activations address this, but most studies assume they help across the board without checking which spectral regimes actually benefit. We introduce a dual-branch, spectrally-gated architecture (DBSG-PINN) that splits low- and high-frequency components into separate subnetworks joined by an adaptive gate, and use it to run a partially controlled ablation of frequency decomposition and spectral routing. We test this on five one-dimensional benchmark PDEs, ranging from smooth, single-scale problems to oscillatory, multi-scale ones. Frequency decomposition helps most on the spectrally complex benchmarks, cutting relative $L_2$ error by up to $59.2\%$ on a multimodal wave problem, but gives little benefit on smoother PDEs. On one benchmark (1D Wave), it performs substantially worse than a simpler fixed-combination variant. The gate's benefit scales with how spectrally rich the target solution is: the full model's advantage over the ablations is largest on multi-scale benchmarks and smallest (or negative) on single-scale ones, consistent with the gate exploiting frequency structure rather than acting as noise,though we do not directly visualize or quantify its spatial activations in this study. All results come from a single training seed across five 1D benchmarks, so we present this as an exploratory study meant to raise questions rather than answer them, and outline the additional seeds and benchmarks needed to test whether the pattern holds.

Mixture-of-Experts (MoE) models scale capacity for strong quality while keeping per-token compute bounded through sparse expert activation. Yet low-latency MoE serving is increasingly challenging, because it spans two inference phases with fundamentally different bottlenecks: prefill is dominated by token-wise expert computation, whereas decode is constrained by memory traffic from the batch-wise activated expert set. However, existing training-free acceleration methods optimize only a single resource proxy, either the experts each token executes or the experts a batch activates, and either discard the excluded experts' contribution or leave it only implicitly approximated. In this paper, we propose ExFold, a unified training-free expert-folding framework for jointly accelerating MoE prefill and decode. ExFold casts both prefill and decode as one budgeted output-approximation problem: execute only a phase-specific constrained expert set while projecting the contribution of budget-excluded experts onto retained experts using calibrated scalar projectors. Motivated by the observation that many expert outputs are directionally aligned but differ in magnitude, ExFold calibrates a pairwise scalar-projector matrix on unlabeled data and uses it at inference time to fold excluded expert contributions into retained experts. Under this view, prefill acceleration becomes token-level Top-K folding, and decode acceleration becomes batch-level expert-pool folding. The two phases differ only in how retained experts are selected, while excluded contributions are recovered by one shared folding mechanism. We implement ExFold as a plug-and-play plugin in vLLM, with a lightweight expert-folding CUDA kernel, delivering up to 1.41x TTFT and 2.45x TPOT speedups while retaining about 99% of the original average quality.