Sparse Autoencoders (SAEs) decompose residual-stream activations into interpretable features. Recent latent-space defenses increasingly rely on these decompositions, assuming that identified "unsafe" SAE features serve as actionable handles for monitoring and intervention. In this paradigm, clamping a specific harmful feature is expected to reliably prevent model misbehavior. However, we show that this success may hide a recoverable failure mode: the clamp may block one visible route to a behavior without eliminating the behavior itself. We formulate this vulnerability as post-intervention recovery, a constrained residual-space optimization problem. Starting from the post-intervention residual state, we optimize residual perturbations to recover the pre-intervention behavior while preserving the post-intervention values of the targeted SAE features. Even under a strong threat model where the intervention remains active throughout optimization and generation, recovery remains possible. To rule out that recovery simply undoes the intervention, we use encoder-orthogonal updates for single-layer interventions and the corresponding feature-map Jacobian in the cross-layer setting. Across
The emergence of quantum computing presents a fundamental challenge to the security of current Internet communication systems. Transport Layer Security (TLS), which forms the backbone of secure web communication, predominantly relies on classical public-key cryptographic algorithms such as RSA and elliptic curve cryptography (ECC), both of which are susceptible to quantum attacks. This paper conducts a large scale empirical evaluation of post-quantum readiness across 32,011 domains, with a primary focus on real-world TLS deployments across diverse sectors by analysing negotiated TLS parameters, including protocol versions, cipher suites, key exchange mechanisms, and certificates. The results indicate that while modern protocols like TLS 1.3 and QUIC are gaining adoption, 15.70% of domains especially in critical sectors such as banking and government still rely on TLS 1.2. Furthermore, 49.3% of domains support hybrid post-quantum key exchange mechanisms (e.g., MLKEM768 with X25519), whereas 50.7% continue to use classical key exchange, reflecting partial transition. Notably, 0% adoption of hybrid post-quantum certificates was observed, leaving the authentication layer vulnerable to
(Abridged) Mechanisms for quenching star formation in galaxies remain hotly debated, with galaxy mergers an oft-proposed pathway. In Ellison et al. (2022) we tested this scenario by quantifying the fraction of recently and rapidly quenched post-starbursts (PSBs) in a sample of post-merger galaxies identified in the Ultraviolet Near Infrared Optical Northern Survey (UNIONS). With our recent development of the Multi-Model Merger Identifier (MUMMI) neural network ensemble (Ferreira et al. 2024a,b), we are now additionally able to predict the time since coalescence (T_PM) for the UNIONS post-merger galaxies up to T_PM = 1.8 Gyr, allowing us to further dissect the merger sequence and measure more precisely when quenching occurs. Based on a sample of 5927 z<0.3 post-mergers identified in UNIONS, we find that the post-coalescence population evolves from one dominated by star-forming (and starbursting) galaxies at 0 < T_PM < 0.16 Gyr, through to a population that is dominated by quenched galaxies by T_PM ~ 1.5 Gyr. We find a PSB excess throughout the post-merger regime, but with a clear peak at 0.16 < T_PM < 0.48 Gyr. In this post-merger time range PSBs are more common than
Large language models (LLMs) are increasingly used for long-document question answering, where reliable attribution to sources is critical for trust. Existing post-hoc attribution methods work well for extractive QA but struggle in multi-hop, abstractive, and semi-extractive settings, where answers synthesize information across passages. To address these challenges, we argue that post-hoc attribution can be reframed as a reasoning problem, where answers are decomposed into constituent units, each tied to specific context. We first show that prompting models to generate such decompositions alongside attributions improves performance. Building on this, we introduce DecompTune, a post-training method that teaches models to produce answer decompositions as intermediate reasoning steps. We curate a diverse dataset of complex QA tasks, annotated with decompositions by a strong LLM, and post-train Qwen-2.5 (7B and 14B) using a two-stage SFT + GRPO pipeline with task-specific curated rewards. Across extensive experiments and ablations, DecompTune substantially improves attribution quality, outperforming prior methods and matching or exceeding state-of-the-art frontier models.
The widespread availability of generative artificial intelligence tools poses new challenges for school mathematics education, particularly regarding the formative role of traditional mathematical tasks. In post-AI educational contexts, many activities can be solved automatically, without engaging students in interpretation, decision-making, or mathematical validation processes. This study analyzes a secondary school classroom experience in which open mathematical tasks are implemented as a didactic response to this scenario, aiming to sustain students' mathematical activity. Adopting a qualitative and descriptive-interpretative approach, the study examines the forms of mathematical work that emerge during task resolution, mediated by the didactic regulation device COMPAS. The analysis is structured around four analytical axes: open task design in post-AI contexts, students' mathematical agency, human-AI complementarity, and modeling and validation practices. The findings suggest that, under explicit didactic regulation, students retain epistemic control over mathematical activity, even in the presence of generative artificial intelligence.
Post-training has greatly improved reasoning in frontier vision-language models, yet its gains for perception remain comparatively limited, creating a bottleneck for end-to-end visual reasoning. To investigate this gap, we introduce a controlled diagnostic framework with two synthetic tasks that disentangle perception from reasoning. Our analysis reveals a consistent perception-reasoning asymmetry: posttraining improves reasoning more substantially than perception, though the underlying mechanism differs by training paradigm. For supervised fine-tuning (SFT), this asymmetry stems from token imbalance in chain-of-thought supervision, where perception occupies fewer tokens and thus receives a weaker training signal. Dynamically reweighting the loss mitigates this imbalance and boosts end-to-end performance by up to 18.2. For reinforcement learning (RL), the asymmetry instead arises from reward coupling: outcome rewards correlate more strongly with reasoning than with perception, weakening the signal for perception learning. Adding a perception-aware reward alleviates the imbalance and improves end-to-end accuracy by up to 6.0; even without groundtruth perception rewards, a reliable s
Post-AGI information systems won't merely distract governance from important problems. They will systematically transform how institutions make decisions in ways that progressively remove humans from meaningful participation in resource allocation. We show that policy myopia -- the tendency to prioritize visible crises over invisible structural risks -- is not a symptom of poor attention management but a mechanism producing irreversible human disempowerment. Through three entangled mechanisms (salience capture displaces consequentialist reasoning, capacity cascade makes recovery structurally infeasible, value lock-in crystallizes outdated preferences), policy myopia couples with institutional dynamics to create a self-reinforcing equilibrium where human disempowerment becomes the rational outcome of institutional optimization. We formalize these mechanisms through coupled dynamical systems modeling and demonstrate through numerical simulation that these mechanisms operate simultaneously across economic, political, and cultural systems, amplifying each other through feedback loops.
Large Language Models (LLMs) have transformed the natural language processing landscape and brought to life diverse applications. Pretraining on vast web-scale data has laid the foundation for these models, yet the research community is now increasingly shifting focus toward post-training techniques to achieve further breakthroughs. While pretraining provides a broad linguistic foundation, post-training methods enable LLMs to refine their knowledge, improve reasoning, enhance factual accuracy, and align more effectively with user intents and ethical considerations. Fine-tuning, reinforcement learning, and test-time scaling have emerged as critical strategies for optimizing LLMs performance, ensuring robustness, and improving adaptability across various real-world tasks. This survey provides a systematic exploration of post-training methodologies, analyzing their role in refining LLMs beyond pretraining, addressing key challenges such as catastrophic forgetting, reward hacking, and inference-time trade-offs. We highlight emerging directions in model alignment, scalable adaptation, and inference-time reasoning, and outline future research directions. We also provide a public reposito
ECG foundation models are increasingly popular due to their adaptability across various tasks. However, their clinical applicability is often limited by performance gaps compared to task-specific models, even after pre-training on large ECG datasets and fine-tuning on target data. This limitation is likely due to the lack of an effective post-training strategy. In this paper, we propose a simple yet effective post-training approach to enhance ECG foundation models. We evaluate it on a publicly available Transformer-based foundation model. Experiments across multiple ECG tasks show that our method consistently outperforms baseline fine-tuning. On the PTB-XL benchmarks, it improves macro AUROC by 0.7%-8.9% and macro AUPRC by 23.3%-77.9%, also outperforming several recent state-of-the-art approaches, including task-specific and advanced architectures. Further analyses demonstrate improved training dynamics and data efficiency, with only 30% of the training data outperforming the baseline trained on the full dataset. Ablation studies highlight the importance of stochastic depth and preview linear probing. These findings underscore the potential of post-training strategies to improve EC
As generalizations of Rota--Baxter groups, Rota--Baxter Clifford semigroups have been introduced by Catino, Mazzotta and Stefanelli in 2023. Based on their pioneering results, in this paper we first continue to study Rota--Baxter Clifford semigroups. Inspired by the corresponding results in Rota--Baxter groups, we firstly obtain some properties and construction methods for Rota--Baxter Clifford semigroups, and then study the substructures and quotient structures of these semigroups. On the other hand, as generalizations of post-groups, Rota--Baxter Clifford semigroups and braided groups, in this paper we introduce and investigate post Clifford semigroups, relative Rota--Baxter Clifford semigroups and braided Clifford semigroups, respectively. We prove that the categories of strong post Clifford semigroups, dual weak left braces, bijective strong relative Rota-Baxter Clifford semigroups and braided Clifford semigroups are mutually pairwise equivalent, and the category of post Clifford semigroups is equivalent to the category of bijective relative Rota--Baxter Clifford semigroups, respectively. As a consequence, we prove that both post Clifford semigroups, relative Rota--Baxter Cliff
Large-scale autoregressive models have demonstrated remarkable capabilities in image generation. However, their sequential raster-scan decoding relies on strictly next-token prediction, making inference prohibitively expensive. Existing acceleration methods typically either introduce entirely new generation paradigms that necessitate costly pre-training from scratch, or enable parallel generation at the expense of a training-inference gap or altered prediction objectives. In this paper, we introduce FlashAR, a lightweight post-training adaptation framework that efficiently adapts a pre-trained raster-scan autoregressive model into a highly parallel generator based on two-way next-token prediction. Our key insight is that effective adaptation should minimize modifications to the pre-trained model's original training objective to preserve its learned prior. Accordingly, we retain the original AR head as a horizontal head for row-wise prediction and introduce a complementary, lightweight vertical head for column-wise prediction. To facilitate efficient adaptation, we branch the vertical head from an intermediate layer rather than the final layer, bypassing the inherent horizontal head
After providing a short review on the recently introduced notion of post-group by Bai, Guo, Sheng and Tang, we exhibit post-group counterparts of important post-Lie algebras in the literature, including the infinite-dimensional post-Lie algebra of Lie group integrators. The notion of free post-group is examined, and a group isomorphism between the two group structures associated to a free post-group is explicitly constructed.
Sudden attention on social media, and how users navigate these contextual shifts, has been a focus of much recent work in social media research. Even when this attention is not harassing, some users experience this sudden growth as overwhelming. In this workshop paper, I outline how growth infuses the design of much of the modern social media platform landscape, and then explore why applying a post-growth lens to platform design could be productive.
Galaxy mergers can enhance star formation rates throughout the merger sequence, with this effect peaking around the time of coalescence. However, owing to a lack of information about their time of coalescence, post-mergers could only previously be studied as a single, time-averaged population. We use timescale predictions of post-coalescence galaxies in the UNIONS survey, based on the Multi-Model Merger Identifier deep learning framework (\textsc{Mummi}) that predicts the time elapsed since the last merging event. For the first time, we capture a complete timeline of star formation enhancements due to galaxy mergers by combining these post-merger predictions with data from pre-coalescence galaxy pairs in SDSS. Using a sample of $564$ galaxies with $M_* \geq 10^{10} M_\odot$ at $0.005 < z < 0.3$ we demonstrate that: 1) galaxy mergers enhance star formation by, on average, up to a factor of two; 2) this enhancement peaks within 500 Myr of coalescence; 3) enhancements continue for up to 1~Gyr after coalescence; and 4) merger-induced star formation significantly contributes to galaxy mass assembly, with galaxies increasing their final stellar masses by, $10\%$ to $20\%$ per mergi
The importance of the post-merger epoch in galaxy evolution has been well-documented, but post-mergers are notoriously difficult to identify. While the features induced by mergers can sometimes be distinctive, they are frequently missed by visual inspection. In addition, visual classification efforts are highly inefficient because of the inherent rarity of post-mergers (~1% in the low-redshift Universe), and non-parametric statistical merger selection methods do not account for the diversity of post-mergers or the environments in which they appear. To address these issues, we deploy a convolutional neural network (CNN) which has been trained and evaluated on realistic mock observations of simulated galaxies from the IllustrisTNG simulations, to galaxy images from the Canada France Imaging Survey (CFIS), which is part of the Ultraviolet Near Infrared Optical Northern Survey (UNIONS). We present the characteristics of the galaxies with the highest CNN-predicted post-merger certainties, as well as a visually confirmed subset of 699 post-mergers. We find that post-mergers with high CNN merger probabilities (p(x)>0.8) have an average star formation rate that is 0.1 dex higher than a
We show how the effects of large numbers of measurements on many-body quantum ground and thermal states can be studied using Quantum Monte Carlo (QMC). Density matrices generated by measurement in this setting feature products of many local non-unitary operators, and by expanding these density matrices as sums over operator strings we arrive at a generalized stochastic series expansion (SSE). Our `post-measurement SSE' is based on importance sampling of operator strings contributing to a measured thermal density matrix. We demonstrate our algorithm by probing the effects of measurements on the spin-$1/2$ Heisenberg antiferromagnet on the square lattice. Thermal states of this system have \SU{2} symmetry, and at first we preserve this symmetry by measuring \SU{2} symmetric observables. We identify classes of post-measurement states for which correlations can be calculated efficiently, as well as states for which \SU{2} symmetric measurements generate a QMC sign problem when working in any site-local basis. For the first class, we show how deterministic loop updates can be leveraged. Using our algorithm we demonstrate the creation of long-range Bell pairs and symmetry-protected topol
Affordable desktop CNC routers are common in education, prototyping, and makerspaces, but most lack a rotary axis, limiting fabrication of rotationally symmetric or multi-sided parts. Existing solutions often require hardware retrofits, alternative controllers, or commercial CAM software, raising cost and complexity. This work presents a software-only framework for indexed rotary machining on GRBL-based CNCs. A custom post-processor converts planar toolpaths into discrete rotary steps, executed through a browser-based interface. While not equivalent to continuous 4-axis machining, the method enables practical rotary-axis fabrication using only standard, off-the-shelf mechanics, without firmware modification. By reducing technical and financial barriers, the framework expands access to multi-axis machining in classrooms, makerspaces, and small workshops, supporting hands-on learning and rapid prototyping.
Construction of a general-purpose post-recognition error corrector poses a crucial question: how can we most effectively train a model on a large mixture of domain datasets? The answer would lie in learning dataset-specific features and digesting their knowledge in a single model. Previous methods achieve this by having separate correction language models, resulting in a significant increase in parameters. In this work, we present Mixture-of-Experts as a solution, highlighting that MoEs are much more than a scalability tool. We propose a Multi-Task Correction MoE, where we train the experts to become an ``expert'' of speech-to-text, language-to-text and vision-to-text datasets by learning to route each dataset's tokens to its mapped expert. Experiments on the Open ASR Leaderboard show that we explore a new state-of-the-art performance by achieving an average relative 5.0% WER reduction and substantial improvements in BLEU scores for speech and translation tasks. On zero-shot evaluation, NeKo outperforms GPT-3.5 and Claude-Opus with 15.5% to 27.6% relative WER reduction in the Hyporadise benchmark. NeKo performs competitively on grammar and post-OCR correction as a multi-task model.
Major mergers between galaxies are predicted to fuel their central supermassive black holes (SMBHs), particularly after coalescence. However, determining the prevalence of active galactic nuclei (AGNs) in mergers remains a challenge, because AGN diagnostics are sensitive to details of the central structure (e.g., nuclear gas clouds, geometry and orientation of a dusty torus) that are partly decoupled from SMBH accretion. X-rays, expected to be ubiquitous among accreting systems, are detectable through non-Compton-thick screens of obscuring material, and thus offer the potential for a more complete assessment of AGNs in mergers. But, extant statistical X-ray studies of AGNs in mergers have been limited by either sparse, heterogeneous, or shallow on-sky coverage. We use new X-ray observations from the first SRG/eROSITA all-sky data release to characterize the incidence, luminosity, and observability of AGNs in mergers. Combining machine learning and visual classification, we identify 923 post-mergers in Dark Energy Camera Legacy Survey (DECaLS) imaging and select 4,565 interacting galaxy pairs (with separations <120 kpc and mass ratios within 1:10) from the Sloan Digital Sky Surve
Post-hoc interpretability methods are critical tools to explain neural-network results. Several post-hoc methods have emerged in recent years, but when applied to a given task, they produce different results, raising the question of which method is the most suitable to provide correct post-hoc interpretability. To understand the performance of each method, quantitative evaluation of interpretability methods is essential. However, currently available frameworks have several drawbacks which hinders the adoption of post-hoc interpretability methods, especially in high-risk sectors. In this work, we propose a framework with quantitative metrics to assess the performance of existing post-hoc interpretability methods in particular in time series classification. We show that several drawbacks identified in the literature are addressed, namely dependence on human judgement, retraining, and shift in the data distribution when occluding samples. We additionally design a synthetic dataset with known discriminative features and tunable complexity. The proposed methodology and quantitative metrics can be used to understand the reliability of interpretability methods results obtained in practica