Ning Yang, Chengzhi Wang, Yibo Liu, Baoliang Tian, Haijun Zhang
Prefill-only KV compression freezes a token subset at the end of prefill and decodes from it without further eviction. The retention decision is therefore irreversible, yet existing methods estimate the corrective signals it relies on, per-head reliability and prompt-level compression sensitivity, online from a single noisy prompt. We argue this is the wrong statistical unit: these signals exhibit far higher cross-prompt regularity than within-prompt signal-to-noise. We introduce \textsc{CompilerKV}, a KV-retention policy whose corrective tables are compiled offline from a calibration corpus, reducing online correction after the standard observation-window scan to $O(1)$ lookups plus a budget clamp. We find that compiled retention tables behave as portable architectural priors: rankings transfer across disjoint corpora on four backbones (mean Spearman $\barρ{=}0.90$), and direct model-to-model table transfer costs only $0.4$--$0.8$ LongBench points on average. At a 512-token budget, \textsc{CompilerKV} attains compressed-SOTA on all four backbones, improving over the strongest prefill-only baseline by $+1.67$ points on average (task-bootstrap 95\% CI $[+1.08,+2.37]$). Pressure regimes amplify the gap: under a fixed $512/32k$ cache ratio, \textsc{CompilerKV} remains the strongest compressed method through 128k RULER ($\sim\!73$ vs.\ FullKV $\sim\!79$, SnapKV $\sim\!38$); on 32k NIAH it reaches $0.89$ vs.\ SnapKV $0.42$; and at 32k input, retaining only $1.56\%$ of the prefill KV, batch-16 serving remains feasible where FullKV is OOM.
Multivariate time series anomaly detection (MTSAD) relies on modeling both temporal dynamics and inter-sensor dependencies to identify complex system faults. Although graph-based detectors explicitly model sensor topologies, they train task-specific networks from scratch on limited target-domain data, remaining vulnerable to spurious correlations and unable to reuse transferable priors. While pretrained Graph Foundation Models (GFMs) offer transferable topological priors and anomaly discriminability, directly adapting them to MTSAD is hindered by cross-modal feature-topology mismatch, unreliable pseudo-anomaly supervision, and decision distortion during node-to-window aggregation. To fill these gaps, we propose a novel adaptation framework, Gfm4Mtsad, which unleashes pretrained GFMs for MTSAD. We introduce the Structure-and-Information Contract Alignment (SICA) module to bridge continuous time series and the pretrained GFM space. This module harmonizes principal subspace token projection, normal-prior-guided graph construction, and residual recovery to suppress spurious edges while retaining fine-grained temporal details. We propose a Mechanism-Aligned Pseudo-Anomaly Screening strategy (MAPS) to steer unsupervised learning. It applies localized dependency breaks while preserving empirical residuals, selecting hard negatives near the normal boundary for targeted supervision. We develop a Tail-Gated Evidence Calibration mechanism (TGEC) to eliminate node-to-window aggregation distortion. It calibrates multi-granular graph responses against normal operational references, strictly preserving benign score rankings while selectively refining extreme anomaly candidates in the tail distribution for reliable window detection and interpretable sensor evidence. Extensive evaluations with two GFM backbones demonstrate that Gfm4Mtsad achieves superior detection performance over existing baselines.
Chain-of-Thought (CoT) reasoning substantially improves Large Language Models (LLMs), yet its out-of-distribution (OOD) generalization mechanism remains theoretically underexplored. We study CoT as compositional OOD generalization: the target task lacks complete reasoning trajectories or direct input-output pairs, while training provides only in-distribution (ID) atomic step data. We show that CoT generalizes by repeatedly applying a shared single-step predictor to compose reusable ID transitions. Theoretically, let $T$ denote the number of reasoning steps in the target task, $n$ the number of training samples per sub-task, and $k$ the number of output tokens per step; then the OOD task risk is controlled by the sum of the ID sub-task risks, scaled by a coefficient of order $\mathcal{O}(k)$ that is independent of $T$, although the summed risk itself may still grow with $T$. A Rademacher-complexity analysis further shows that the estimation term is of order $\mathcal{O}(\sqrt{T/n})$, rather than linear in $T$, due to predictor sharing under a uniform capacity control on the shared class. For autoregressive Next-Token Prediction, the final-task risk is bounded by the best-in-class step-level risk scaled by the same coefficient of order $\mathcal{O}(k)$, plus an estimation term of order $\widetilde{\mathcal{O}}\big(k\sqrt{T/n}\big)$. A parity-coupled construction further shows that, for the sub-task empirical risk minimizer, the expected final-task risk is at least the best in-class risk plus $Ω(\sqrt{T/n})$ when $T\le n$. Experiments on Atom-Task and GSM8K support the theory: CoT enables short-to-long and compositional generalization, while explicit state tracking is crucial. Our code is available at https://github.com/maxliang114514/CoT-OOD.
Aleksandros Sobczyk, Gioele Gottardo, Christos K. Matzoros, Mirko De Vita, Filip Skogh, Anastasios Zouzias, Jiawei Zhuang
Linear attention has emerged as a cornerstone for efficient long-context architectures, as evidenced by its integration into state-of-the-art open-source models including Qwen3.5/3.6, Kimi Linear/K3, and RWKV-7. Models that incorporate linear attention layers with the so-called Delta-Rule involve the inversion of triangular matrices as a core sub-routine. This operation often forms a performance bottleneck, and, due to its high-sensitivity to numerical errors, it can significantly deteriorate end-to-end model accuracy if it is not carefully implemented. This work provides a systematic analysis of both direct and iterative triangular inversion algorithms, targeting methods that are rich in matrix products, and, therefore, have the potential to efficiently utilize modern hardware. To that end, our analysis covers a broad spectrum of mathematical and practical aspects, with a heavy focus on numerical stability, computational complexity, and, ultimately, hardware efficiency and practical considerations. We provide a rigorous experimental evaluation to verify these properties in practical scenarios, and in low-precision floating-point representations, highlighting the strengths and limitations of each method. Performance benchmarks on NPUs reveal up to $4.3\times$ speed-up against the state-of-the-art implementations of SGLang for triangular matrix inversion, leading to significant performance improvements on the entire layer level, while maintaining end-to-end model accuracy.
We show that a hardwired distributed processing unit (HDPU) can take a neural decision at the sensor and so remove the inference from the host. Built for zebra-crossing detection on a low-cost Artix-7 FPGA, the unit decides on every camera frame in a fixed 2,269,477 clock cycles (34.04 ms depends upon Fpga clk ), inside the 37.49-ms frame period, and on the board it returns the same scores and decisions as its bit-exact software model on all 639 test images. The whole detection path, from camera pixels to decision, takes 1,814 LUTs, with no processor, no operating system and no full-resolution frame store. When the unit answers instead of reporting "not sure", 95.7% of its answers on street photographs are correct, and a hardware contrast stage lifts accuracy on camera-like dashcam frames from 47.7% to 78.2%. The host is spared all of the inference: 8.13 ms of a core in every frame period, where a full-precision ResNet-18 would need more than a frame period even on 16 threads. It also receives 16 bytes per frame instead of 614,400, 38,400 times less. The unit is an 80x60-pixel pipeline feeding a four-layer binarized CNN with 93,408 one-bit weights, verified bit-exactly from a Python model through RTL simulation to the board. Earlier HDPUs did pre-processing (edges, corners, steering); this one shows that the decision itself can move to the sensor.
Sajedeh Norouzi, Mostafa Rahmani Ghourtani, Eric Samikwa, Torsten Braun
Reliable channel estimation and symbol detection in 5G New Radio (NR), as well as emerging 6G radio systems, rely on demodulation reference signals (DMRS). The density and placement of these pilots on the orthogonal frequency-division multiplexing (OFDM) time--frequency resource grid affect both receiver reliability and spectral efficiency. Sparse DMRS layouts increase the time--frequency distance between pilot observations and data resource elements (REs), which can make pilot information harder to exploit with purely local processing. This paper proposes \textbf{PARGAR (Pilot-Aware Graph Attention Receiver)}, a graph-attention neural receiver that represents the OFDM resource grid as a graph for bit detection. The graph combines local time--frequency edges with pilot-anchor links from data REs to selected pilot REs, and we further study a channel-scaled pilot-anchor rule based on approximate coherence time and coherence bandwidth. Using Sionna-generated OFDM simulations, we evaluate \textsc{PARGAR} against DeepRx, a convolutional neural network (CNN)-based receiver, across dense, two-symbol sparse, and single-symbol sparse DMRS layouts under channel-mismatch scenarios. Under the considered SISO uncoded setup, channel-scaled \textsc{PARGAR} explicitly injects pilot-aware inductive bias and yields lower BER in selected sparse-pilot scenarios while remaining competitive under dense pilots.
Partially interpretable models extend simple interpretable models with the option to defer complicated inputs to an accurate but uninterpretable black-box model. By deferring more inputs, the model gets more accurate; by deferring fewer, it gets more interpretable. Existing procedures to train such partially interpretable models build on linear models or decision rules. We propose three algorithms for decision trees: a partially interpretable extension of the existing Branches algorithm, which performs exact search to optimize the tree; a conversion algorithm that we run on top of standard CART, which is computationally more efficient; and a generalized CART whose splitting criterion takes into account the option to defer while growing the tree. Perhaps surprisingly, the performance of the three algorithms is similar across many realworld datasets. We can explain this through a formal bound on the potential gain of smart tree construction, provided the black box accuracy is sufficiently homogeneous across the domain. We also construct a synthetic dataset where the accuracy of the black box is deliberately inhomogeneous. There, our deferral-aware generalization of CART does outperform regular CART and matches the performance of exact search. Finally, we evaluate an application to concept bottleneck models. Partially interpretable decision trees can replace uninterpretable neural networks after the bottleneck. On undeferred inputs, the model is fully interpretable, and deferring preserves accuracy.
Policy-gradient methods need a scalar reward, yet for many tasks no formula captures what we want, while judging which of two outcomes is better is easy. We study reinforcement learning that receives only such comparisons - no reward function, no reward model - in multi-step tasks. Groups of trajectories start from a shared state, and each action is credited by how its reward-to-go compares with those of other trajectories at the same time step. We show the expected update is policy gradient with reward-to-go replaced by its win rate against the population still running at that time, weighted by the chance an opponent is still running; in deterministic environments the update ranks actions by true value and the optimal policy is stable, whereas under stochastic outcomes it prefers actions that win more often rather than those with higher mean. Label noise near ties or from position bias leaves the expected update unchanged; only noise that also flips clear comparisons halves the signal. A critic trained from the same comparisons removes the state confound without changing the expected update. In a pre-registered study over six environments with ten seeds per claim, comparisons cost no detectable performance relative to scalar returns at the same granularity, learning a reward model from the same comparisons is not better, and ten percent of comparisons suffice when spent on a few time points with all pairs. With real LLM judges, comparisons are 7-10 points more accurate than pointwise scores, and the gap grows with the stronger judge. Trained end-to-end with an LLM judge on a requirement the hand-written reward omits, policies satisfy it in all five seeds but land far less often than with oracle comparisons, because the judge is near chance on the near-tied failures dominating early training. Comparisons do not close the gap to PPO with scalar rewards on HalfCheetah.
Target-conditioned diffusion drafters predict candidate tokens in parallel, yet each new block must wait for updated target-model hidden states. This forces drafting and verification to run sequentially, limiting speedup even though each block is generated in one pass. We introduce D3D, a standalone diffusion drafter that removes this inter-round dependency. D3D combines blockwise denoising with auxiliary autoregressive supervision to achieve long accepted prefixes without target hidden states. This independence lets D3D draft successor blocks during verification in one packed pass, enabling asynchronous speculative decoding. We progressively prune layers according to their effect on acceptance length to reduce the cost of packed drafting. Removing five of 36 layers retains over 99\% of mean acceptance length across three benchmarks with Qwen3-8B as the target. Across seven benchmarks, D3D averages 5.00$\times$ and 5.71$\times$ decoding speedup with Qwen3-8B and Qwen3-32B targets, compared with 4.03$\times$ and 3.67$\times$ for DFlash. On GSM8K with Qwen3-32B, speedup reaches 7.48$\times$. Code is publicly available at https://github.com/fudan-generative-vision/D3D.
Quentin Le Roux, Marti Jimenez, Yannick Teglia, Philippe Loubet Moundi
The protection of mature, multi-stage Face Recognition Systems is crucial to system providers. This technical report targets the threat model most relevant to industry practitioners: although model training is performed in-house, training data is sourced from possibly-compromised third parties. Given this setting, we present three practical defense "tricks" against data poisoning-based backdoor attacks. The first operates at training-time, detecting poisoned identities during training or prior to a pipeline's deployment. The last two operate at inference-time, filtering suspicious inputs before an answer can be sent back to a potential attacker. Each method is presented with their motivation, algorithms and computational impact, and evaluated on realistic face recognition workloads. Our goal is to provide actionable hardening techniques to system providers.
Adrian Hayler, Shashank Reddy Chirra, Andrei Lupu, Johannes Forkel, Bidipta Sarkar, Siheng Feng, Jakob Foerster
We study zero-shot coordination (ZSC), where independently developed agents must coordinate at test time. While ZSC has been well studied in the RL literature, far less is known about the performance of LLM agents despite their increasing deployment in such settings. Existing work on LLM coordination often relies on specialised scaffolding that independently developed agents are unlikely to share in practice. Moreover, evaluations on complex environments (e.g., Hanabi) make it difficult to pinpoint the sources of coordination failure. In contrast, we focus on simple, general-purpose scaffolds in minimal environments designed to isolate specific coordination challenges. Our results show that even in these controlled settings, frontier LLM agents struggle to coordinate, largely due to limited understanding of the coordination problem and weak reasoning about their partner's beliefs. While LLMs introduce semantic information as an additional axis for coordination, they nonetheless fail to exploit this structure effectively. Towards this, we propose Coordination-friendly definitions (CFDs) as a principled approach for enabling robust coordination among LLM agents. Finally, we show that CFDs can be discovered automatically, removing the need for manual engineering. Code is available at https://github.com/ahayler/zsc-llm-agents.
Sam Gijsen, Michał Łukomski, Marc-André Schulz, Kerstin Ritter
Flow matching and diffusion models enable conditional generation across domains ranging from images to proteins, with recent extensions to out-of-distribution contexts. Yet generative models of neural time series have largely remained restricted to categorical conditioning, precluding compositional and zero-shot generalization. In this work, we propose PINC-Flow, a per-timestep conditioned diffusion transformer for generating realistic fMRI brain dynamics during unseen cognitive tasks by injecting both compositional language and optional spatial priors in-context. Such zero-shot generation could enable counterfactual neuroscience by supporting in-silico design and evaluation of novel cognitive experiments before empirical validation. Leveraging this model, we evaluate across hundreds of held-out task conditions and characterize predictive performance in relation to the training manifold. From language alone, the model recovers region-specific recruitment across tasks and held-out spatial activation patterns. Spatial priors, when available, complement the text pathway by anchoring generation in regions of task space where language alone degrades, while retaining the compositional structure needed for counterfactual task specification. To our knowledge this is the first generative model of whole-cortex fMRI dynamics for unseen cognitive tasks, advancing counterfactual neuroscience and data-driven experimental design.
Sungha Kim, Gawon Lee, Jusuk Lee, Jonghae Park, H. Jin Kim, Daesol Cho
Maximum entropy reinforcement learning (MaxEnt-RL) enables robust exploration, yet practical implementations often restrict policies to simple Gaussians. While recent approaches incorporate expressive generative policies via importance-weighted supervised learning, they are prone to importance weight collapse, which limits their scalability in high-dimensional action spaces. Our key insight is to mitigate this limitation by localizing the sampling region, avoiding the weight degeneracy induced by importance sampling over the entire action space. To instantiate this insight, we introduce \textbf{FLAG} (\textbf{F}low policy with \textbf{L}atent-\textbf{A}ugmented \textbf{G}uidance). FLAG augments the state space with a flow latent variable and optimizes a provably consistent proxy MaxEnt-RL objective. We empirically demonstrate that FLAG enables expressive policy optimization with limited importance samples and scales to state-based high-dimensional control tasks. Furthermore, FLAG achieves state-of-the-art performance among expressive-policy MaxEnt online RL methods on state-based high-dimensional control benchmarks.
Toshitaka Hayashi, Dalibor Cimr, Hamido Fujita, Richard Cimler
One-class Support Vector Machine(OCSVM) is a variant of support vector machine(SVM) for one-class classification (OCC). However, OCSVM has scalability issues with large datasets. This paper proposes an acceleration strategy for OCSVM. The idea is to decompose the dataset into samples and train OCSVM models for single data points. Subsequently, ensemble learning is applied to combine all models to compute the OCSVM model for the dataset. In addition, further acceleration is achieved through a support vector selection strategy with an OCSVM model trained on the average of the training samples. The experiment compared the proposal and traditional OCSVM using the Python package. The proposed strategy is faster than traditional OCSVM, while achieving similar classification results. Moreover, the proposed strategy can create a one-to-one correspondence between samples and models. Source code is uploaded at https://github.com/ToshiHayashi/ODSVM.
Xuliang Huang, Tianxiang Zhao, Chengwei Qin, Yao Shu
Offline reinforcement learning with verifiable rewards (RLVR) trains on a fixed pool of sampled responses and typically supervises every token of every response, although responses and tokens are not equally informative. We study which ones to supervise through the Gibbs-optimal policy of KL-regularized RLVR, which offline training targets, and measure the mismatch between a policy and this target with a bidirectional KL divergence. At the reference policy, this mismatch factorizes exactly and separates the two outcomes: correct responses contribute mainly through the forward term, concentrated on their low-probability tokens, and incorrect responses through the reverse term, concentrated on the responses the policy generates often and on their high-probability tokens. Based on this analysis, we propose GIST (Gibbs-Targeted Hierarchical Data Selection), a two-level selection method that approximates these principles with signals available before training. GIST uses the sampling temperature to select correct responses on-policy and incorrect responses near the policy's mode, and it uses the initial policy's token probabilities to supervise only the tokens where the mismatch lies. Applied to two offline training objectives, GIST raises the average accuracy over GSM8K, MATH-500, and MBPP for all three models under both objectives. In a controlled comparison on Qwen3-0.6B/GSM8K, it outperforms seven selection baselines that choose from the same candidates while supervising 66.6% fewer tokens than training on all candidates of the same prompts.
Finding high-performing formulations is difficult when few have been measured and experimental budgets are limited. We propose KolTa, an active learning framework that selects batches to discover the top 5% of formulations at a target temperature under a fixed query budget. KolTa combines a three-stage predictor of component contributions, role-conditioned interactions, and anchored temperature response with task-specific soft physics constraints. Its Physics-Residual-Calibrated Upper Confidence Bound (PRC-UCB) acquisition rule combines predicted utility, ensemble disagreement, and a target-temperature physics residual, with nonnegative exploration weights fitted to out-of-fold signed errors. We evaluate the framework on three measured tasks and one simulated task covering binary-mixture viscosity, drug solubility, ionic-liquid viscosity, and electrolyte conductivity. KolTa ranks first or second in mean area under the learning curve (AULC) on every task and outperforms seven of the eight comparator strategies in paired tests, winning 75 of 78 paired comparisons across tasks and seeds against those seven strategies. Across the three measured tasks, its largest gap to the best mean AULC is 0.032, compared with at least 0.056 for every comparator. On Drug, reaching 80% average recovery of the remaining targets across drugs requires an estimated median of 348 additional formulation queries, about 18% fewer than the fastest comparator. Researchers can adapt the framework to diverse formulation problems through component roles and physical constraints.
Data-driven pre-fault dynamic security assessment (DSA) rapidly evaluates the dynamic risk of credible contingencies on a power system using machine learning (ML). Existing approaches face two limitations. First, they require a large labelled database for training, with a separate model trained, tuned, and maintained for each contingency in a potentially long list of credible contingencies. Second, the trained models generalize poorly to unseen contingencies. This work addresses these limitations by using a tabular foundation model (TFM) that assesses stability through in-context learning, and requires no retraining or hyperparameter optimization. We use a single TFM to assess many contingencies at once, removing the need to train one classifier per contingency. We also characterize when using electrical distance coordinates (EDC) as continuous features enables the TFM to generalize to unseen contingencies and when it does not, showing how a few labelled samples can reliably improve generalization. Through comprehensive case studies on the IEEE 68-bus system, we show that a single TFM attains an average Macro F1 score of about 90\% with only 120 labelled samples per contingency, outperforming both classical ML and deep learning baselines, without any model retraining or hyperparameter tuning. For held-out contingencies, we show that using just 10 labelled samples of the new contingency with a modified EDC encoding matches the balanced accuracy of the best achievable transfer learning oracle model, which requires fully labelled data and is not deployable in practice. Overall, this initial study paves the way for developing and deploying foundation models for power system operations, with possible applications across multiple operational tasks.
Nicolas Samuel Blumer, Julien Martinelli, Samuel Kaski
Black-box optimization in science and engineering often comes with side information: experts, simulators, pretrained predictors, or heuristics can suggest promising candidates. Such feedback can accelerate search, but can also be biased, input-dependent, or misleading. Feedback-aware Bayesian optimization methods typically handle tasks independently, limiting adaptation across feedback sources. In-context optimizers enable cross-task adaptation, but usually rely only on optimization history at test time. We study feedback-informed in-context black-box optimization (FICBO), where a pretrained optimizer conditions on both the observed history and cheap auxiliary feedback for the current candidate set. We introduce structured priors over how auxiliary feedback relates to the objective, and use them to pretrain a feedback-aware transformer across regimes ranging from highly informative to misleading. At test time, the model estimates source reliability in context by comparing observed objective values with auxiliary signals, improving query selection. On synthetic and real-world tasks, FICBO effectively exploits informative feedback while remaining robust to weak or misleading sources. Empirical investigations further illustrate how the model perceives test-time sources, offering insights into its interpretability and decision-making process.
Dooge asked whether hydrology has laws at the catchment scale. We ask whether a low-dimensional system built under physical laws can match the accuracy of neural networks that outperform process-based models on the same inputs. Three laws constrain a lumped catchment model. Conservation makes it a flux-connected storage network, dissipation fixes the sign of every flux and aggregation determines how the unresolved composition enters the fluxes, through effective laws over threshold distributions and a memory of the forcing. These laws admit four kinds of refinement. A parsimonious model under them, CaRT (11 parameters per catchment), is refined in a loop assisted by AI agents. Each refinement addresses an observed deficit through a hypothesis from the literature. The result, CaRT+, has fourteen states, 35 parameters per catchment and no daily neural network. Most refinements follow from aggregation over thresholds or from variables that carry no water but modulate fluxes. On 602 British and 593 US gauged catchments, single CaRT+ models reach median Nash-Sutcliffe efficiencies of 0.863 and 0.765 in held-out years. A stateful LSTM reaches 0.869 and 0.751 and an entity-aware LSTM 0.862 and 0.754, while three-seed LSTM ensembles remain ahead of the CaRT+ ensemble by 0.003-0.012 in median paired NSE. On 505 drier Australian gauged catchments unused in development, single CaRT+ models reach 0.778 against 0.762 and 0.753 for the two LSTMs. The results suggest that catchment-scale laws exist as laws of form, with the constitutive families shared and only their parameters learned.
Pengfei Qu, Lei Ye, Wenyu Ouyang, Chi Zhang, Yikai Chai, Shuolong Xu, Yongri Piao, Miao Zhang 等 9 位作者
Accurate prediction in ungauged basins (PUB) requires effective transfer across hydrologically diverse regions. While deep learning has advanced large-sample hydrological modeling, its cross-regional generalization depends in part on the static basin descriptors used to guide model behavior. Earth Foundation Models (EFMs) provide a promising alternative by learning environmental representations from multi-modal satellite data. In this study, we evaluate the 64-dimensional AlphaEarth Foundations (AEF) embeddings as a new form of static basin representation and assess their value for streamflow prediction under both in-sample (IS) and out-of-sample (OOS) settings using the CAMELS dataset. Results show that AEF embeddings offer higher OOS robustness than CAMELS attributes, despite nearly identical IS performance. Mutual-information analysis reveals statistical associations between AEF embeddings and conventional catchment attributes related to terrain, vegetation, and rooting characteristics. We further investigate similarity-based donor-basin selection for PUB using CAMELS attributes and AEF embeddings. Across 100 target basins, AEF static inputs generally improved PUB performance. With AEF inputs, similarity-based donor selection outperformed random selection for smaller donor pools, but its advantage diminished as pool size increased. Adding more donor basins did not consistently improve performance. Leave-one-cluster-out experiments showed AEF input benefits depended on basin partitioning, suggesting that transferability remains sensitive to the catchment conditions represented during training. Overall, this study demonstrates that Earth Foundation Model embeddings provide a transferable representation for PUB prediction and underscores the importance of basin representation quality and donor pool design, motivating future research toward unified hydrological foundation models.
Erik Nikulski, Yamen Habib, Vicenç Gomez, Anders Jonsson, Rubén Moreno-Bote, Javier Segovia-Aguas
Generalization in Reinforcement Learning (RL) refers to the ability to execute close-to-optimal policies in unseen tasks after the agent has been trained on a different set of tasks. Building on the seminal work of the successor representation and further adaptations with function approximation, Transfer in RL has traditionally focused on generalizing to tasks that only differ in the reward function. A decade after the introduction of the successor representation, Robust RL emerged simultaneously from several articles in the field of operations research. In Robust RL, the transition kernel is unknown, and the goal is to maximize the expected reward under this uncertainty. Our work unifies these two paradigms through robust successor features, which generalize across both the reward function and the transition kernel, under the assumption that tasks are linear Markov Decision Processes. We derive a bound on Generalized Policy Improvement (GPI) that explicitly quantifies how performance degrades with the mismatch between transition kernels, recovering existing successor-feature guarantees when dynamics are shared. Finally, the generalization capabilities of robust successor features are validated on several grid-based benchmarks and compared to previous alternatives that focus solely on either the reward or the transition kernel.
When many inputs share a budget of repeated attempts by a large language model (LLM), a common rule gives harder inputs more. On public math, code and formal-proof samples, the best published difficulty-proportional rule captures half or more of the oracle gain on benchmarks the model mostly solves, but at most 13% on hard workloads, even given true pass rates. Because the value of one more attempt peaks at intermediate difficulty, the optimal static allocation is inverted-U and gives little to inputs that are hopeless at the current budget. In a strict survey of 221 LLM allocation papers, at least 13 give the hardest inputs the most attempts; their easier settings sit at budgets where this is near-optimal, so the cost appears on harder workloads. Acting on it requires knowing which inputs are solvable, so information, not the policy, drives the gain. With a verifier and a prior from a related model's past runs on the same inputs, Thompson Sampling raises the solved share on the hardest workload from 14% to 27%; without such a prior, no policy we tested beats round-robin. For the policies we tested, the gain vanishes between prior rank correlations of about 0.3 and 0.55, which cheap text-based priors barely reach on the weakest models, and it requires a verifier whose false acceptances are rare relative to the pass rate. On new Qwen2.5 samples the pattern holds, and tempering the prior avoids losses on easy workloads. We turn these results into a decision guide.
Yuxiang Zeng, Kun Xie, Yang Wang, Jigang Wen, Xiaocan Li, Guangxing Zhang, Gaogang Xie, Jiannong Cao
Architecture attribute predictors reduce the cost of neural architecture search by estimating accuracy or latency from a small set of evaluated candidates. Neural architectures are operation-labeled directed acyclic graphs (DAGs), where edge directions determine computational roles and execution dependencies. Existing graph Transformer predictors incorporate topological and directional information in different ways, but primarily focus on specific mechanisms rather than coordinated encoder-wide modeling of these semantics. We introduce encoder-wide directional consistency, a design principle that coordinates the modeling of directed computational semantics across node initialization, global spectral encoding, pairwise interaction, and feature updates. To implement this principle, we propose DTA-GT, a Direction- and Topology-Aware Graph Transformer that integrates direction- and topology-aware node initialization, Magnetic Laplacian spectral encoding, direction-dual structural attention, and a Direction-Sensitive MoE-FFN. Across accuracy and latency prediction on NAS-Bench-101 and NAS-Bench-201, DTA-GT consistently outperforms representative sequence-, GNN-, Transformer-, and hybrid predictors under limited-budget settings. Ablations and Magnetic spectral controls support the contribution of the direction-aware stages and orientation-sensitive spectral information, while mechanistic diagnostics characterize module behavior. These results support encoder-wide directional consistency as an effective design principle for architecture DAG representation.
Minjae Oh, Sangjun Song, Gyubin Choi, Yunho Choi, Yohan Jo
On-Policy Distillation (OPD) has emerged as a dominant post-training paradigm for Large Language Models, especially in reasoning domains. However, OPD remains unstable in practice due to the high gradient variance of its single-sample Monte Carlo estimator, and recipes for stable training are still immature. We propose vOPD (On-Policy Distillation with a control variate baseline), which casts OPD as policy-gradient RL and stabilizes it by introducing a control variate baseline---canonically a value function---from the RL literature. We show that the OPD value function admits a closed form: the per-token negative reverse KL divergence between student and teacher, available directly from the existing forward pass with no additional critic or inference. Existing stabilization methods either compute the full token-level reverse KL over the entire vocabulary, adding significant overhead, or restrict it to a top-k support, biasing the objective. vOPD instead preserves the lightweight single-sample estimator and subtracts the value function as a detached baseline to keep the gradient unbiased while reducing variance. Moreover, we show that a top-k approximation of the baseline further lowers cost without compromising performance. Across mathematics, science, and instruction-following benchmarks, vOPD consistently outperforms base OPD and matches the most expensive full-vocabulary method, offering an efficient stabilization of OPD through principled RL variance reduction.
Large Language Models (LLMs) demonstrate remarkable reasoning capabilities, yet the internal mechanisms driving these multi-step processes remain opaque. Existing mechanistic interpretability approaches often rely on superficial text-pattern co-occurrences or capture only short-term effects, struggling to trace the long-horizon, outcome-driven influence of internal components such as neurons or sparse features in multi-step reasoning.In this paper, we introduce Integrated Policy Gradient (IPG), a novel, training-free framework that brings policy-based attribution into mechanistic interpretability for LLM reasoning. Grounded in \textit{outcome-oriented} and \textit{sequential-influence-aware} principles, it backpropagates outcome-level, non-differentiable behavioral signals (e.g., reasoning outcome correctness) through entire inference trajectories, utilizing path integration to isolate outcome-relevant mediators at component level for reliable attribution. This provides a linear approximation of Causal Mediation Analysis (CMA) in reasoning. Empirical evaluations demonstrate that \method achieves precise localization and reveals mechanistic insights by identifying sparse, concentrated subsets of internal components associated with specific reasoning behaviors. Our method also enables modulation of reasoning capabilities, offering a powerful and reliable framework for understanding LLM reasoning. The project page is at https://foundation-model-research.github.io/IPG .
Foundation models for electroencephalography (EEG) are pretrained on tens of thousands of hours of recordings, and their accuracy grows with that amount. Such data is hard to assemble. The largest sources are clinical archives that cannot be redistributed, and the public sources are spread over several archives, file formats, and licences. We release eegdash-corpus-v0, a training-ready corpus of 77 public EEG datasets, as the gated Hugging Face dataset EEGDash/eegdash-corpus-v0 with the SHA-256 of every file. REVE, the largest published EEG foundation model, names 89 datasets as its pretraining corpus. Twelve cannot be redistributed: four clinical archives that hold 82% of its hours, and eight under no-derivatives licences. The corpus holds the 76 others that REVE's sampling rate allows and one more from REVE's release card: 77 datasets, 14,075 hours from 7,272 subjects, 5.0 million 10-second samples and 6.21 billion tokens per epoch. Every dataset is fetched through EEGDash, converted once with REVE's recipe, and stored in its own buckets of a chunked Zarr store with its licence recorded beside it, so a model can train on the whole store or on any subset of datasets. To test that the store pretrains a model, we retrain REVE-Base from scratch twice with the authors' code, changed only in how it reads data, and score it on seven public OpenEEGBench tasks. It reaches 0.331 under linear probing and 0.273 after fine-tuning, against 0.333 and 0.325 for the released model and 0.200 and 0.205 for an untrained encoder. The converter, the conversion reports, and the loader changes are released with the corpus.
Temporal point processes (TPPs) are the canonical framework for event sequence data, with applications spanning seismology, social media, and electronic health records. Recently, neural TPPs have been reported to substantially outperform classical TPPs, such as Hawkes processes, but these comparisons pitted flexible neural models against Hawkes processes in their most restricted form, and did not consider large-scale medical data. Here, we revisit this comparison through a systematic evaluation of Poisson and Hawkes processes of increasing flexibility against neural TPPs, on five established benchmark datasets and on two large-scale medical evaluation tasks derived from UK Biobank disease histories and MIMIC-IV ICU records, under autoregressive next-event prediction. First, the previously reported gap between Hawkes processes and neural TPPs largely disappears once Hawkes processes are equipped with flexible, learnable baselines and kernels. Second, the flexibility of the intensity parametrization also limits prior neural TPPs: a Transformer with a spline-based intensity head outperforms previous intensity-based neural TPPs and is on par with the intensity-free model. Decomposing the log-likelihood into a time term $\log p(t)$ and a type term $\log p(d \mid t)$, we find that this flexibility advantage lies almost entirely in the time density: more flexibility does not improve the type term. On these benchmarks, the gains of neural TPPs thus lie mostly in modeling when events occur, not which. Third, our results underscore the value of complex datasets for evaluating TPPs: only on the MIMIC-IV ICU records do neural TPPs retain an advantage in the type term. Our results inform both future method development and model choice. We release our framework, TorchPP, as an open-source package with consistent implementations of classical and neural TPPs.
Majharulislam Babor, Giacomo Rossi, Annalisa Altavilla, Oliver Schlüter, Marina M. -C. Höhne
Edible insects offer an efficient source of alternative protein, requiring less land and water and emitting less greenhouse gas than conventional livestock. However, their successful integration into the food supply chain demands reliable species authentication to control allergen exposure, prevent adulteration, and meet regulatory standards. Near-infrared spectroscopy provides a rapid analytical tool, but its performance drops when applied to production batches unseen during training due to batch-to-batch variation in spectral measurements. We introduce the Batch-Invariant Spectral Network (BISN), an end-to-end framework that combines a learnable preprocessing module, initialised with Savitzky-Golay filtering, with an unsupervised, entropy-regularised auxiliary branch connected through a gradient reversal layer. The branch uses no batch labels and supplies opposing entropy gradients to the auxiliary head and preprocessing module. Using 2,700 spectra from three species (Acheta domesticus, Hermetia illucens, and Tenebrio molitor) collected across three independent production batches, BISN achieves a mean leave-one-batch-out accuracy of 0.93 (standard deviation 0.04), outperforming the strongest baseline by four percentage points (p=0.002). Further insights gained by using explainable AI confirm that model decisions consistently rely on the lipid and protein absorption regions across all folds, connecting predictive performance to known insect biochemistry. BISN addresses both cross-batch robustness and biochemical interpretability for automated insect species authentication in a controlled benchtop setting. The source code and dataset are publicly available at https://github.com/majharB/bisn.
The records that could train the most useful models, such as a hospital's patient files, are often the ones an institution cannot share. We propose CoRTeC, a method that releases synthetic versions of such records under differential privacy. Differential privacy places a proven limit on what the synthetic data can reveal about any person, and in particular on membership inference. CoRTeC spends the privacy budget once, on a small file of noisy counts designed for training a classifier. For each group of records, the file gives the group's size, its balance of outcomes, a histogram of every column for each outcome and the outcome rate in each of several non-overlapping cells. A reviewer can check every number before the file leaves the institution. Anyone can then turn the counts into synthetic rows, as often as they like and at no further privacy cost. A maximum-entropy decoder can do this without any language model, and so can each of the five frozen language models we tested. None of them ever sees a real record. A selection step then matches the rows to the released counts. Decoded without a language model, the automatically configured release at $\varepsilon = 2$ trains classifiers to within 0.02 AUC of a real sample of the same size on the Adult dataset, across three data splits. On NHANES and bank marketing, without the selection step, it stays within 0.05 AUC. On Adult, with a fixed cohort rule, the release carries the outcome rates a deployment names more closely than a real sample does, at some cost in accuracy on pairs of columns. Four controls show that most of the utility comes from the release and the selection step, and that a language model adds what it knows from the column names. No private row ever reaches a model, and no row CoRTeC generates copies a real one. The strongest membership attack stays far below the bound the budget allows.
Iterative feedback tuning (IFT) learns the parameters of a fixed-structure controller by stochastic gradient descent. At each iteration, two closed-loop experiments are run on the plant: the first one measures the tracking error, and the second one, driven by this error, provides the sensitivities from which the gradient of the cost is estimated. This estimate is unbiased for linear plants only. We study the biased case on automotive electronic throttle valves, whose friction and stiction vary between valves and runs, controlled by an intelligent proportional (model-free) controller, which is an incremental PI controller. We show that the mismatch of the controller state at the start of the two experiments, which stiction makes unavoidable, adds an error inversely proportional to the amplitude of the second experiment. On digital twins of ten valves, the IFT gradient is a descent direction in only 69\% of the low-cost region, friction being the cause, and all update rules plateau and then drift. On two hardware valves, the best update rule depends on the noise structure of the gradient. On the first valve, where one gradient component is buried in noise, Adam outperforms Gauss--Newton, its normalisation and its momentum each improving one parameter. On the second valve, where both components are consistent, the same normalisation amplifies a direction that Gauss--Newton rightly discounts, and Adam drifts to a controller worse than the initial one. The twins reproduce the mechanism but not this ranking.