arXiv:2512.03594v2 Announce Type: replace
Abstract: Detailed routing remains one of the most complex and time-consuming steps in modern physical design due to the challenges posed by shrinking feature sizes and stricter design rules. Prior detailed routers achieve state-of-the-art results by leveraging iterative pathfinding algorithms to route each net. However, runtimes are a major issue in detailed routers, as converging to a solution with zero design rule violations (DRVs) can be prohibitively expensive.
In this paper, we propose leveraging reinforcement learning (RL) to enable rapid convergence in detailed routing by learning from previous designs. We make the key observation that prior detailed routers statically schedule the cost weights used in their routing algorithms, meaning they do not change in response to the design or technology. By training a conservative Q-learning (CQL) model to dynamically select the routing cost weights which minimize the number of algorithm iterations, we find that our work completes the ISPD19 benchmarks with 1.56x average and up to 3.01x faster runtime than the baseline router while maintaining or improving the DRV count in all cases. We also find that this learning shows signs of generalization across technologies, meaning that learning designs in one technology can translate to improved outcomes in other technologies.
Science Journals
arXiv:2512.03829v2 Announce Type: replace
Abstract: Reliable operation of high-power proton cyclotrons is a critical requirement for Accelerator Driven Systems (ADS) and other large-scale applications. Beam tuning in such machines is traditionally performed manually, a process that can be slow, non-optimal, and difficult to execute in the presence of faults or changing conditions. To address this, we developed and deployed a machine learning (ML) based tuning framework on the Injector 2 cyclotron at PSI, chosen as an ideal testbed for high-power operation. The system combined a tailored reinforcement learning (RL) algorithm with real-time diagnostics and control, and incorporated accelerator-physics inspired adaptations such as an overshoot strategy that reduced magnetic field settling times by nearly a factor of six. Over an extensive 12-day operational test campaign, relatively long in the context of real-time ML experiments, the RL agent successfully tuned the machine across multiple operating points. For each investigated configuration, stable policies were obtained within a few hours of online training and subsequently demonstrated reliable low-loss operation during overnight evaluation runs. Crucially, the learned policy remained effective when transferred from low-current training to operation at beam currents up to 800 {\mu}A, demonstrating robust generalization under appropriately adapted operational constraints. These results constitute the first demonstration of RL-assisted tuning on a high-power cyclotron, with direct relevance to ADS-class drivers.
arXiv:2512.05673v2 Announce Type: replace
Abstract: Residual minimization in dual norms is central to Weak Adversarial Neural Network (WAN) approaches for solving partial differential equations (PDEs). This framework naturally leads to saddle-point problems whose numerical solutions can be highly unstable depending on the underlying iterative scheme. Motivated by this structure, we propose and analyze the Uzawa Double Deep Ritz Method, a deep PDE solver that integrates neural network approximations with the classical Uzawa iteration. The proposed method is built around two coupled update rules performed at each iteration: a residual update, obtained by minimizing a Ritz functional associated with the dual problem, and a solution update, obtained by minimizing a Ritz functional driven by the current residual. Both variables are represented by neural networks, mirroring the classical Uzawa architecture for saddle-point problems. By replacing the adversarial min-max optimization of WAN with a sequence of Deep Ritz minimization problems, our study theoretically proves that the proposed method acts as an iterative scheme for solving the WAN formulation. Furthermore, we establish a comprehensive convergence theory for an inexact Uzawa scheme where both subproblems are solved approximately. This analysis extends to practical gradient-based implementations, providing rigorous stability and convergence guarantees for both single and multiple-gradient step update strategies. Numerical experiments validate our theoretical findings and demonstrate the robustness of the proposed approach.
arXiv:2512.07588v3 Announce Type: replace
Abstract: Analysing learning in Multi-Agent Reinforcement Learning (MARL) environments is challenging, in particular with respect to \textit{individual} decision-making. Practitioners frequently struggle to compare training runs due to the inherent stochasticity in algorithms arising from random dithering exploration, environment transition noise, and stochastic gradient updates to name a few. Traditional analytical approaches, such as replicator dynamics, oft rely on mean-field approximations to remove stochastic effects, but this simplification, whilst able to provide general overall trends, can lead to dissonance between analytical predictions and actual agent realisations. We propose modelling MARL training as a \textit{coupled stochastic dynamical systems}, capturing both agent interactions and environmental characteristics. Leveraging tools from dynamical systems theory, we pragmatically analyse the stability and sensitivity of agent behaviour, which are key dimensions for their practical deployments, for example, in presence of strict safety requirements. This framework allows us to rigorously study the inherent stochasticity of MARL, providing a deeper understanding of system behaviour.
arXiv:2607.16085v1 Announce Type: new
Abstract: Increasingly, speech and language processing tasks take either audio or text directly rather than extracting features from these as the input to the classifier or regressor. Often these systems make use of complex, for example transformer-based, processes that have the ability to derive highly non-linear mappings between the input and the output. Unfortunately these systems can also learn ''shortcuts'' where the classifier is overly reliant on particular aspects of the input to yield the output. For the task of language proficiency assessment, this over-reliance can enable learners to increase their score by exploiting the shortcut rather than improving their ability. This paper introduces a novel training criterion that is able to reduce the classifier's reliance on shortcuts, thus for example limiting this option for malpractice in language assessment. This process is illustrated on two forms of assessment system, one based on the audio the other on the speech recognition text. The results show that, for both systems, there is higher correlations with features that could be exploited for malpractice than expected from the human reference, indicating an over-reliance on these features. By introducing the modified training criterion, this correlation can be reduced to be closer to the reference correlation.
arXiv:2512.08356v2 Announce Type: replace
Abstract: The formation of polarized signaling domains on cell membranes is a fundamental example of biological pattern formation. While such patterns resemble structures from equilibrium phase separation, they are intrinsically non-equilibrium, driven by energy-consuming enzymatic cycles that switch molecules like phosphoinositides or small GTPases between distinct states. Here, we develop a minimal model of this enzyme-driven phase ordering process. Starting from microscopic reaction kinetics, we derive a mesoscopic theory that belongs to the class of active Model A with a global constraint. This framework yields an explicit mean-field phase diagram and closed-form expressions for key observables, such as interfacial tension, domain fractions, and phase coexistence boundaries, in terms of kinetic rates. In this context, phase coexistence is controlled by non-equilibrium parameters like catalytic rates and enzymatic asymmetry, rather than equilibrium parameters such as saturation concentrations. The resulting phase-separated domains rapidly exchange material with their surroundings. Their maintenance requires a continuous power input determined by enzymatic kinetics. The predicted phenomenology is consistent with experimental observations on reconstituted systems of phosphoinositide and Rab5 membrane patterning. We further study how metastable uniform states decay via nucleation of minority-phase domains and subsequent coarsening, driven by an effective interfacial tension. Using large deviation theory, we derive the critical nucleation radius under the action of the intrinsic, multiplicative chemical noise. The analytical results are quantitatively confirmed by stochastic simulations of the process. Our work provides a theoretical framework identifying key biochemical parameters controlling active phase separation on membrane scaffolds, offering testable predictions for experiments.
arXiv:2512.10113v3 Announce Type: replace
Abstract: Dark personality traits have long been associated with antisocial and toxic online behaviors, yet their relationship with observable online activity remains unclear. We investigate the association between validated dark personality measures, self-reported experiences of online incivility, and linguistic and behavioral features extracted from real-world user activity. To this end, we developed a Web application that securely links responses to validated psychological questionnaires collected via Amazon Mechanical Turk with participants' Reddit activity. This yielded a dataset of nearly 57K comments (2.2M tokens) from 114 users, represented through a broad set of linguistic and behavioral features. Our analyses reveal a clear distinction between self-reported and observed behavior. Dark personality traits show consistent associations with self-reported engagement in uncivil interactions. However, no validated dark personality dimension significantly predicts text-derived toxicity or linguistic features. In contrast, self-reported experiences of engaging in or being targeted by toxic behavior are robustly reflected in users' language, exhibiting consistent associations with measures of negativity, moral framing, and emotional intensity. Taken together, these findings highlight a gap between stable personality traits and their manifestation in surface-level linguistic signals. While computational features effectively capture behavioral engagement in online incivility, they do not provide reliable proxies for underlying personality constructs within the present framework. Our results underscore the importance of grounding computational approaches in validated psychological measures and point to the need for richer, context-aware representations to better understand the relationship between personality and online behavior.
arXiv:2512.13636v4 Announce Type: replace
Abstract: Current Vision-Language-Action (VLA) paradigms in autonomous driving primarily rely on Imitation Learning (IL), which introduces inherent challenges such as distribution shift and causal confusion. Online Reinforcement Learning offers a promising pathway to address these issues through trial-and-error learning. However, applying online reinforcement learning to VLA models in autonomous driving is hindered by inefficient exploration in continuous action spaces. To overcome this limitation, we propose MindDrive, a VLA framework comprising a large language model (LLM) with two distinct sets of LoRA parameters. The one LLM serves as a Decision Expert for scenario reasoning and driving decision-making, while the other acts as an Action Expert that dynamically maps linguistic decisions into feasible trajectories. By feeding trajectory-level rewards back into the reasoning space, MindDrive enables trial-and-error learning over a finite set of discrete linguistic driving decisions, instead of operating directly in a continuous action space. This approach effectively balances optimal decision-making in complex scenarios, human-like driving behavior, and efficient exploration in online reinforcement learning. Using the lightweight Qwen-0.5B LLM, MindDrive achieves Driving Score (DS) of 78.04 and Success Rate (SR) of 55.09% on the challenging Bench2Drive benchmark. To the best of our knowledge, this is the first work to demonstrate the effectiveness of online reinforcement learning for the VLA model in autonomous driving.
arXiv:2512.20609v2 Announce Type: replace
Abstract: Divalent atoms have emerged as powerful alternatives to alkalis in ultracold atom platforms, offering unique advantages arising from their two-electron structure. Among these species, ytterbium (Yb) is especially promising, yet its anionic properties and its Rydberg spectrum remain comparatively unexplored. In this work, we perform a first and comprehensive experimental and theoretical investigation of ultralong-range Rydberg molecules (ULRMs) of $^{174}$Yb in $6sns\,^1S_0$ Rydberg states across nearly two decades in principal quantum number $n$ and three orders of magnitude in molecular binding energy. Using the Coulomb Green's function formalism, we compute Born-Oppenheimer molecular potentials describing the Rydberg atom in the presence of a ground-state perturber and achieve quantitative agreement with high-resolution molecular spectra. This enables the extraction of low-energy electron-Yb scattering phase shifts, including the zero-energy $s$-wave scattering length and the positions of two spin-orbit split $p$-wave shape resonances. Our results provide strong evidence that the Yb$^{-}$ anion exists only as a metastable resonance.% We additionally show the sensitivity of ULRM spectra to the atomic quantum defects, using this to determine the quantum defect of the $6s23f\, ^1F_3$ state. Together, these findings establish Yb ULRMs as a powerful probe of electron-Yb interactions and lay essential groundwork for future Rydberg experiments with divalent atoms.
arXiv:2512.21428v2 Announce Type: replace
Abstract: We report high-precision frequency ratio measurements between optical atomic clocks based on $^{27}$Al$^+$, $^{171}$Yb, and $^{87}$Sr. With total fractional uncertainties at or below $3.2 \times 10^{-18}$, these measurements meet an important milestone criterion for redefinition of the second in the International System of Units. Discrepancies in $^{87}$Sr ratios at approximately $1\times10^{-16}$ and the Al$^+$/Yb ratio at $1.6\times10^{-17}$ in fractional units compared to our previous measurements underscore the importance of repeated, high-precision comparisons by different laboratories. A key innovation in this work is the use of a common ultrastable reference delivered to all clocks via a 3.6 km phase-stabilized fiber link between two institutions. Derived from a cryogenic single-crystal silicon cavity, this reference improves comparison stability by a factor of 2 to 3 over previous systems, with an optical lattice clock ratio achieving a fractional instability of $1.3 \times 10^{-16}$ at 1 second. By enabling faster comparisons, this stability will improve sensitivity to non-white noise processes and other underlying limits of state-of-the-art optical frequency standards.
arXiv:2512.22274v4 Announce Type: replace
Abstract: We introduce GeCo, a geometry-grounded metric for jointly detecting geometric deformation and occlusion-inconsistency artifacts in static scenes. By fusing residual motion and depth priors, GeCo produces interpretable, dense consistency maps that reveal these artifacts. We use GeCo to systematically benchmark recent video generation models, uncovering common failure modes, and further employ it as a training-free guidance loss to reduce deformation artifacts during video generation.
arXiv:2512.24069v2 Announce Type: replace
Abstract: We consider the design of mixing matrices to minimize the operation cost for decentralized federated learning (DFL) in wireless networks, with focus on minimizing the maximum per-node energy consumption. As a critical hyperparameter for DFL, the mixing matrix controls both the convergence rate and the needs of agent-to-agent communications, and has thus been studied extensively. However, existing designs mostly focused on minimizing the communication time, leaving open the minimization of per-node energy consumption that is critical for energy-constrained devices. This work addresses this gap through a theoretically-justified solution for mixing matrix design that aims at minimizing the maximum per-node energy consumption until convergence, while taking into account the broadcast nature of wireless communications. Based on a novel convergence theorem that allows arbitrarily time-varying mixing matrices, we propose a multi-phase design framework that activates time-varying communication topologies under optimized budgets to trade off the per-iteration energy consumption and the convergence rate while balancing the energy consumption across nodes. Our evaluations based on real data have validated the efficacy of the proposed solution in combining the low energy consumption of sparse mixing matrices and the fast convergence of dense mixing matrices.
arXiv:2601.00898v3 Announce Type: replace
Abstract: Diffusion-based policies have gained growing popularity in solving a wide range of decision-making tasks due to their superior expressiveness and controllable generation during inference. However, effectively training large diffusion policies using reinforcement learning (RL) remains challenging. Existing methods either suffer from unstable training due to directly maximizing value objectives, or face computational issues due to relying on crude Gaussian likelihood approximation, which requires a large amount of sufficiently small denoising steps. In this work, we propose DIPOLE (Dichotomous diffusion Policy improvement), a novel RL algorithm designed for stable and controllable diffusion policy optimization. We begin by revisiting the KL-regularized objective in RL, which offers a desirable weighted regression objective for diffusion policy extraction, but often struggles to balance greediness and stability. We then formulate a greedified policy regularization scheme, which naturally enables decomposing the optimal policy into a pair of stably learned dichotomous policies: one aims at reward maximization, and the other focuses on reward minimization. Under such a design, optimized actions can be generated by linearly combining the scores of dichotomous policies during inference, thereby enabling flexible control over the level of greediness.Evaluations in offline and offline-to-online RL settings on ExORL and OGBench demonstrate the effectiveness of our approach. We also use DIPOLE to train a large vision-language-action (VLA) model for end-to-end autonomous driving (AD) and evaluate it on the large-scale real-world AD benchmark NAVSIM, highlighting its potential for complex real-world applications.
arXiv:2601.09735v5 Announce Type: replace
Abstract: Software transactional memory (STM) allows programmers to easily implement concurrent data structures. STMs simplify atomicity. Recent STMs can achieve good performance for some workloads but they have some limitations. In particular, STMs typically cannot support long-running reads which access a large number of addresses that are frequently updated. Multiversioning is a common approach used to support this type of workload. However, multiversioning is often expensive and can reduce the performance of transactions where versioning is not necessary. In this work we present Multiverse, a new STM that combines the best of both unversioned TM and multiversioning. Multiverse features versioned and unversioned transactions which can execute concurrently. A main goal of Multiverse is to ensure that unversioned transactions achieve performance comparable to the state of the art unversioned STM while still supporting fast versioned transactions needed to enable long running reads. We implement Multiverse and compare it against several STMs. Our experiments demonstrate that Multiverse achieves comparable or better performance for common case workloads where there are no long running reads. For workloads with long running reads and frequent updates Multiverse significantly outperforms existing STMS. In several cases for these workloads the throughput of Multiverse is several orders of magnitude faster than other STMs.
arXiv:2607.15394v1 Announce Type: new
Abstract: Black-box models limit the adoption of artificial intelligence in medicine due to their lack of interpretability and reproducibility. We introduce a statistically grounded framework that provides fully interpretable, rule-based clinical classification using the Bernoulli Na\"ive Bayes (BNB) model. The method applies supervised $\chi^2$-guided statistical binarization to continuous variables, identifying thresholds that maximize association with clinical outcomes within the training data. This transformation allows BNB to operate effectively on continuous medical data without sacrificing its inherent transparency. The approach was evaluated on three benchmark datasets, Pima Indians Diabetes, Wisconsin Breast Cancer, and Heart Failure Prediction, achieving area-under-the-curve (AUC) scores of 0.800 for the Pima analysis, 0.984 for Wisconsin Breast Cancer, and 0.919 for Heart Failure Prediction. In addition to discrimination, probabilistic reliability was assessed using leakage-safe cross-validated calibration analysis including Brier score, calibration intercept/slope, and post-hoc beta calibration, which improved probability calibration across datasets. These results suggest that a statistically interpretable framework can achieve performance comparable to more complex models while providing explicit, clinically meaningful decision rules and calibrated risk estimates. To illustrate this transparency concretely, a complete worked example demonstrates that model inference can be reproduced using only a reference table and basic arithmetic, without access to software or proprietary tools. This work offers a practical approach to supporting trustworthy and generalizable AI in real-world healthcare settings.
arXiv:2607.15397v1 Announce Type: new
Abstract: Electronic health record audit logs record timestamped actions through which clinical work is carried out. Generated as operational metadata, they now support research on clinician effort, patient outcomes, care-team coordination, and workflow structure. This Perspective explains that breadth by articulating audit logs as multi-axial event streams and drawing implications for representation learning, evaluation, and governance. Each logged action belongs simultaneously to multiple clinically meaningful relations: a clinician's work, a patient's trajectory, a team's activity, and a recurring workflow. This structure motivates foundation-model pretraining to learn reusable representations over the raw stream. Reading audit logs as multi-axial traces specifies what such representations must preserve, how their value should be tested, and how their use should be governed.
arXiv:2607.15462v1 Announce Type: new
Abstract: We study a single-machine scheduling problem in which each job $j$ has a nonnegative processing time $p_j\ge 0$ and a due date $d_j\in\mathbb{R}$. For a non-idling schedule $S$, let $C_j(S)$ be the completion time and let $L_j(S)=C_j(S)-d_j$ be the (possibly negative) lateness. The objective is to minimize the sum of the $k$ largest lateness values, interpolating between maximum lateness ($k=1$) and total lateness ($k=n$).
We prove that the decision version is weakly NP-complete when $k$ is part of the input. For fixed $k$, we give an $O(k^2 n^{k+2})$ algorithm. As a consequence, we resolve a conjecture of Woeginger on the top-$k$ tardiness problem and obtain an $O(n^{k+2})$ algorithm for every fixed $k$.
Our main structural result shows that there exists an optimal schedule that admits a block-island decomposition. Outside a suitable top-$k$ set, jobs form due-date blocks ordered by due date. Within each due-date class, the top-$k$ jobs form a suffix in lexicographic shortest-processing-time (SPT) order. This structure also yields an FPT algorithm parameterized by $D+k$, where $D$ is the number of distinct due dates.
Independently, a standard dual representation of the top-$k$ objective reduces the problem to a family of total-tardiness instances with uniformly shifted due dates. For integral data, this gives a pseudopolynomial algorithm and a fully polynomial additive approximation scheme with error at most $\varepsilon M$, where $M=\max\{1,\max_j p_j,\max_j |d_j|\}$. The same route also gives XP algorithms for fixed $P$ and fixed $D$, where $P$ is the number of distinct processing times.
arXiv:2601.12222v2 Announce Type: replace
Abstract: Music generative artificial intelligence (AI) is rapidly expanding music content, necessitating automated song aesthetics evaluation. However, existing studies largely focus on speech, audio or singing quality, leaving song aesthetics underexplored. Moreover, conventional approaches often predict a precise Mean Opinion Score (MOS) value directly, which struggles to capture the nuances of human perception in song aesthetics evaluation. This paper proposes a song-oriented aesthetics evaluation framework, featuring two novel modules: 1) Multi-Stem Attention Fusion (MSAF) builds bidirectional cross-attention between mixture-vocal and mixture-accompaniment pairs, fusing them to capture complex musical features; 2) Hierarchical Granularity-Aware Interval Aggregation (HiGIA) learns multi-granularity score probability distributions, aggregates them into a score interval, and applies a regression within the interval to produce the final score. We evaluated on two datasets of full-length songs: SongEval dataset (AI-generated) and an internal aesthetics dataset (human-created), and compared with two state-of-the-art (SOTA) models. Results show that the proposed method achieves stronger performance for multi-dimensional song aesthetics evaluation. The inference code and checkpoint are publicly available at https://github.com/yisan33/song-aesthetics-evaluation.
arXiv:2601.13020v2 Announce Type: replace
Abstract: Continual instruction tuning (CIT) requires multimodal large language models (MLLMs) to adapt to a stream of tasks without forgetting prior capabilities. A common strategy is to isolate updates by routing inputs to different LoRA experts. However, existing LoRA-based Mixture-of-Experts (MoE) methods often jointly update the router and experts in an indiscriminate way, causing the router's preferences to co-drift with experts' adaptation pathways and gradually deviate from early-stage input--expert specialization. We term this as Misaligned Co-drift, which blurs expert responsibilities and exacerbates forgetting. To address this, we introduce the pathway activation subspace (PASs), a LoRA-induced subspace that reflects which low-rank pathway directions an input activates in each expert, providing a capability-aligned coordinate system for routing and preservation. Based on PASs, we propose a fixed-capacity PASs-based MoE--LoRA method with two components: PAS-guided Reweighting, which calibrates routing using each expert's pathway activation signals, and PAS-aware Rank Stabilization, which selectively stabilizes rank directions important to previous tasks. Experiments on a CIT benchmark show that our approach consistently outperforms a range of conventional continual learning baselines and MoE--LoRA variants in both accuracy and resistance to forgetting, without increasing model parameters. Our code is publicly available at https://github.com/yueluoshuangtian/PASs-MoE.
arXiv:2601.15470v4 Announce Type: replace
Abstract: We develop low distortion embeddings with outliers from arbitrary metrics into hierarchically separated trees (HSTs). In particular, we develop an efficient algorithm that for any $\epsilon>0$, given an input metric $(X,d)$, and a probabilistic embedding of all but $k$ points from $X$ into HSTs with distortion $c$, samples from a probabilistic embedding of all but $O(\frac{k}{\epsilon}\log k)$ points into HSTs that achieves distortion at most $(32+\epsilon)c$.
Our results are based on two key technical components. First, we extend an algorithm of Munagala et al. [2023] for minimizing the distortion of embeddings without outliers into HSTs to the setting with outliers. We combine this with new results on bi-Lipschitz extensions into trees and $\ell_1$ space. In particular, we show that any probabilistic embedding into HSTs can be extended to $k$ additional points with only a factor $O(\log k)$ of additional distortion. This bi-Lipschitz extension result utilizes a new probabilistic partitioning scheme that we call onion partitioning.
arXiv:2601.20147v3 Announce Type: replace
Abstract: The rapid advancement of Large Language Models (LLMs) has revolutionized software engineering automation, particularly in automated code summarization, which enhances program comprehension and supports development activities. However, training LLMs for code summarization remains computationally expensive, with performance deteriorating on longer inputs-challenges that intensify when handling millions of code-comment pairs. We investigate strategic data optimization through targeted token reduction to minimize computational overhead while maintaining summary quality. We compare three token-level reduction techniques-(i) Abstract Syntax Tree (AST) representations, (ii) Function Signatures, and (iii) CrystalBLEU-guided pruning-combined with semantic filtering, evaluating them on Java and Python in standalone and cascaded reduction settings. Our findings reveal highly language-dependent optimal strategies: AST-based optimization achieves 37% performance improvements in Java with 56-73% token reduction but shows up to 49% degradation in Python. Conversely, Function Signatures perform poorly in Java but optimally in Python, achieving 83% token reduction while maintaining quality. CrystalBLEU demonstrates cross-language robustness with 60-72% reduction. These results challenge assumptions about cross-language transferability, demonstrating that which tokens are kept matters more than how many are removed, making language-aware token curation essential for efficient code summarization.
arXiv:2607.15470v1 Announce Type: new
Abstract: Optimization models are built in a variety of modeling languages and solved by a variety of solvers, but once a solution exists, the information needed to understand it is fragmented: each solver exposes a partial, differently named set of native diagnostics, and the modeling language has already canonicalized the formulation the user wrote. We present pyoptexplain, a practitioner-first Python library for post-optimality analysis of optimization models that sits above this layer. It adapts a model authored in any of five modeling front ends, namely cvxpy, Pyomo, gurobipy, docplex, and OR-Tools, into a normalized internal representation, solves it through a choice of backends, and answers the why and what-if questions of an optimization decision through one uniform interface. The design rests on two observations. First, a post-optimality quantity requested from different backends for the same problem can come back as an exception, a structurally meaningless zero or a basis-dependent value that disagrees across solvers, so reporting whatever one solver returns is unreliable. pyoptexplain reports a quantity only when both the representation and the chosen backend can justify it, and does not approximate unavailable information. Second, repeated scenario analysis can amortize its cost by extracting the model once and reusing a warm solver session across a batch of scenarios. pyoptexplain builds a single scalable what-if interface, uniform across its modeling languages and backends and returning a certified report for every scenario, at a cost within a small constant factor of the bare solver. A reproducible computational study substantiates both claims. Source code is available at https://github.com/h-fellahi/pyoptexplain and installation can be done through the Python Package Index https://pypi.org/project/pyoptexplain/.
arXiv:2601.22402v2 Announce Type: replace
Abstract: Rotary Positional Embeddings (RoPE) have become the standard for Large Language Models (LLMs) due to their ability to encode relative positions through geometric rotation. However, we identify a significant limitation we term ''Spectral Rigidity'': standard RoPE utilizes a fixed geometric decay ($\theta^{-i}$) optimized for local syntactic coherence, which fails to capture the long-range, periodic structures inherent in recursive logic and algorithmic reasoning. This results in a ''Structure Gap'', where models trained on shallow reasoning chains fail to extrapolate to deeper recursive steps. In this work, we introduce Bifocal Attention, an architectural paradigm that decouples positional encoding into two distinct modalities: Geometric Eyes (Standard RoPE) for precise token-level manipulation, and Spectral Eyes (Learnable Harmonic Operators) for tracking long-range recursive depth. We propose a novel training protocol, Spectral Evolution, which initializes positional frequencies as static geometric parameters but allows them to evolve via gradient descent into a harmonic basis optimized for the specific algorithmic topology of the task.
arXiv:2601.22818v2 Announce Type: replace
Abstract: Fine-tuned LLMs can covertly encode prompt secrets into outputs via steganographic channels. Prior work demonstrated this threat but relied on trivially recoverable encodings. We formalize payload recoverability via classifier accuracy and show previous schemes achieve 100\% recoverability. In response, we introduce low-recoverability steganography, replacing arbitrary mappings with embedding-space-derived ones. For Llama-8B (LoRA) and Ministral-8B (LoRA) trained on TrojanStego prompts, exact secret recovery rises from 17$\rightarrow$30\% (+78\%) and 24$\rightarrow$43\% (+80\%) respectively, while on Llama-70B (LoRA) trained on Wiki prompts, it climbs from 9$\rightarrow$19\% (+123\%), all while reducing payload recoverability. We then discuss detection. We argue that detecting fine-tuning-based steganographic attacks requires approaches beyond traditional steganalysis. Standard approaches measure distributional shift, which is an expected side-effect of fine-tuning. Instead, we propose a mechanistic interpretability approach: linear probes trained on later-layer activations detect the secret with up to 33\% higher accuracy in fine-tuned models compared to base models, even for low-recoverability schemes. This suggests that malicious fine-tuning leaves actionable internal signatures amenable to interpretability-based defenses.
arXiv:2602.01244v3 Announce Type: replace
Abstract: Training agentic models for terminal-based tasks critically depends on high-quality terminal trajectories that capture realistic long-horizon interactions across diverse domains. However, constructing such data at scale remains challenging due to two key requirements: \textbf{\emph{Executability}}, since each instance requires a suitable and often distinct Docker environment; and \textbf{\emph{Verifiability}}, because heterogeneous task outputs preclude unified, standardized verification. To address these challenges, we propose \textbf{TerminalTraj}, a scalable pipeline that (i) filters high-quality repositories to construct Dockerized execution environments, (ii) generates Docker-aligned task instances, and (iii) synthesizes agent trajectories with executable validation code. Using TerminalTraj, we curate 32K Docker images and generate 50,733 verified terminal trajectories across eight domains. Models trained on this data with the Qwen2.5-Coder backbone achieve consistent performance improvements on TerminalBench (TB), with gains of up to 20\% on TB~1.0 and 10\% on TB~2.0 over their respective backbones. Notably, \textbf{TerminalTraj-32B} achieves strong performance among models with fewer than 100B parameters, reaching 35.30\% on TB~1.0 and 22.00\% on TB~2.0, and demonstrates improved test-time scaling behavior. All code and data are available at https://github.com/Wusiwei0410/TerminalTraj.