Aggressive weight quantization to 2-bit precision offers substantial throughput and memory gains for large language model (LLM) inference, but typically incurs severe accuracy degradation. These gains are particularly relevant for edge and on-device deployment, where memory capacity and bandwidth are primary constraints. In this work, we extend Recover-LoRA -- a lightweight, data-free accuracy recovery method originally developed for general model weight corruption -- to the setting of ultra-low-bit quantization. We propose a selective mixed-precision strategy in which only gate and up projection layers of the MLP are quantized to 2-bit (W2), while all other linear layers remain at higher precision, yielding a mixed-precision GateUp configuration. We demonstrate via roofline analysis across three model families (4B--20B) and two hardware platforms that a W4/W2-GateUp deployment (4-bit base with 2-bit gate/up) delivers 7.5--23.3\% TPS improvement over uniform W4 depending on model and context length, while confining quantization error to a predictable subset of layers. We then apply Recover-LoRA -- training low-rank adapters on the quantized layers via logit distillation with synthetic data -- to recover accuracy lost from 2-bit quantization of the gate and up layers. In a case study on Qwen3-4B, Recover-LoRA achieves 80--95\% accuracy recovery on 9 of 12 benchmarks, using only 10k synthetic training samples and no labeled data. We further demonstrate that synthetic data performs comparably to curated labeled data for distillation-based recovery, and that recovery generalizes to out-of-distribution evaluation tasks. Our results present Recover-LoRA as a practical post-quantization accuracy recovery tool for aggressive weight compression in deployment settings.

Discrete diffusion language models can generate text efficiently by updating multiple masked positions in parallel, but this parallelism introduces a quality-latency trade-off. Aggressive decoding may commit mutually dependent tokens too early, while conservative decoding requires many denoising steps. Existing methods address this tension by deciding which tokens are safe to reveal using confidence or dependency criteria. However, avoiding unsafe commits does not necessarily make the remaining masked sequence easy to decode, since uncertain tokens may depend on masked tokens, creating a bottleneck for denoising steps. We propose AXON, a training-free module that can be added on top of existing parallel decoding strategies for diffusion language models. Rather than replacing the base decoder, AXON monitors the remaining uncertain masked tokens and intervenes only when their current state suggests that additional context is needed. It then shifts the criterion from which tokens are safest to reveal to which confident reveals would best support later denoising. AXON selects anchors, confident masked tokens that uncertain positions attend to, using attention, uncertainty, and confidence signals. Experiments on reasoning and code-generation benchmarks across multiple diffusion language models show that AXON improves the quality-latency trade-off of existing parallel decoders, often reducing the number of function evaluations while maintaining or improving accuracy.

A robotics benchmark score measures success under one fixed evaluation setup, yet is routinely treated as evidence of general manipulation capability. We identify four failure modes, each of which weakens or invalidates a benchmark's role as a valid proxy for that capability: shortcut solvability, lack of statistical significance, creeping overfitting, and data-source dependence. We propose one diagnostic per failure mode. We audit LIBERO, CALVIN, SimplerEnv, RoboCasa, and RoboTwin 2.0 under these diagnostics. LIBERO and CALVIN fail multiple diagnostics. RoboCasa and RoboTwin 2.0 fail fewer, despite appearing far less often in recent progress claims. On LIBERO, a 0.09B probe with no language encoder scores at or near reported SOTA, and most reported gains are not provably statistically significant. On CALVIN, randomizing block poses within the training range drops performance for every tested policy. We release the four diagnostics with reference implementations for authors and reviewers to apply before treating a benchmark score as evidence of progress. Code and artifacts are available at https://ripl.github.io/manipulation_benchmark_audit/.

Recent advances in multimodal retrieval-augmented generation (MM-RAG) have shifted toward minimal parsing, relying on page-level images for producing retriever embeddings and for answer generation. While efficient, this trend often neglects explicit handling of the rich, structured information in complex enterprise documents, instead depending on pre-trained embeddings or vision-language models to implicitly capture such structure. In this work, we take a more direct approach: MM-BizRAG proactively extracts and represents document structure via a document structure-aware split that dynamically routes documents through orientation-specific ingestion pipelines, applying explicit layout-aware parsing for vertically structured documents (e.g., reports) and holistic page-level representations for horizontally structured documents (e.g., slide decks). A unified LLM-driven artifact transformation pipeline with placeholder-based positional alignment preserves natural reading order, while inference-time multimodal assembly decouples retrieval representations from generation context, enabling richer, more grounded answers without any finetuning requirement. Through experiments on a large, heterogeneous enterprise dataset and two public benchmarks (SlideVQA and FinRAGBench-V), MM-BizRAG consistently outperforms state-of-the-art vision-centric baselines by up to 32% points, with especially strong gains on report-style layouts. Furthermore, we introduce FastRAGEval, a single-call LLM Judge metric for fine-grained generative recall that halves RAGChecker's cost while achieving stronger human alignment.

Simulation environments are useful for both robot policy learning and planning verification and validation. Traditionally, the process of creating a simulation was onerous. Creating a bespoke simulation environment for each individual environment that a robot would operate in was simply infeasible. In this work, we introduce PerceptTwin, a fully automatic pipeline that constructs interactive simulations directly from semantic scene representations produced by a robot's perception stack. PerceptTwin combines open-vocabulary object maps with 3D asset generation, affordance prediction, and commonsense condition checking. These interactive simulations can be used to validate and refine plans before they are executed on the robot hardware. Borrowing from the AI alignment literature, we also introduce an LLM judge that verifies plan correctness and alignment with human preferences. Experiments show that PerceptTwin feedback allows LLM planners to refine plans, enhance safety, and resist harmful black-box prompting attacks. In our suite of tasks, PerceptTwin improves plan success by an average of approximately 39% for GPT5, GPT5Mini, and GPT5Nano planners. Additionally, PerceptTwin also improves human plan verification by up to 18% on average for plans that fail due to unfilled skill preconditions. Our results demonstrate the potential of open-vocabulary scene simulation from robot perception as a foundation for safer, more reliable robot planning.

Multi-agent systems are commonly designed to reduce disagreement through voting, consensus protocols, debate, or fault-tolerant aggregation. We argue that this objective is insufficient for value-laden tasks, where disagreement may reflect genuine normative uncertainty rather than agent error. Building on prior work on reasoning-trace disagreement in human-AI collaborative moderation, we propose a knowledge-representation layer in which reasoning traces and agent decisions are abstracted into symbolic disagreement states. Given agents producing explicit reasoning traces and binary decisions, we distinguish four states according to reasoning similarity and conclusion agreement: convergent agreement, divergent agreement, convergent disagreement and divergent disagreement. These states support defeasible strategic routing rules. We instantiate the framework in content moderation and argue that disagreement-aware routing provides a bridge between sub-symbolic LLM deliberation and symbolic knowledge representation for multi-agent strategic reasoning.

Prosthetic socket fitting remains largely manual and iterative, and objective fit metrics are still limited. Part of the challenge is the lack of long-term real-life pressure data at the residual limb--socket interface. Traditional pressure sensors are prone to drift over time, and capture only normal pressures at sparse locations within the socket, missing a critical component for biomechanical analysis: shear. Although some sensors can report both normal and shear interface stresses, these components are often difficult to decouple because of measurement crosstalk. One potential path forward is to develop models that can augment available measurements. This work introduces a testbed to evaluate model performance under sparse pressure sensing using two complementary validation signals: (i) the global wrench (\ie, total forces and moments expressed in an orthonormal frame) transmitted through the socket, by an artificial residual-limb, and (ii) local interface loads (\ie, decoupled normal and shear pressure components in a right-hand-rule orthogonal frame that lives in each instrumented location) measured by sparse sensing clusters, each composed of four capacitance-sensing channels. Rather than presenting full-field pressure estimates, the focus is on an analysis sequence that quantifies how well candidate mechanical models explain both global and local measurements under controlled conditions. A quasi-static spring--mass contact model is evaluated, and its parameters are identified via a two-stage convex least-squares problem. Validation under static loading shows that estimating constant bias terms reduces steady offsets in the wrench channels and improves agreement with local measurements. A Pareto-front sensitivity analysis further illustrates how the trade-off between global and local objectives changes when bias terms are included.

Hearing aids impose strict latency and power constraints that current DNN-based speech enhancement systems struggle to meet on embedded hardware. We characterize this gap by deploying both speech separation and denoising using the lightweight SuDoRM-RF++ architecture on the AMD-Xilinx Kria KV260, evaluated at FP32 and 16-bit fixed-point precision for each task. Across these configurations, first-sample latency tracks with on-chip parameter caching rather than arithmetic throughput, identifying data movement as the primary bottleneck. Precision reduction halves the model memory footprint without compromising objective speech quality. The fixed-point denoising accelerator achieves a first-sample latency of 9.7~ms, meeting the 10~ms clinical threshold, while speech separation reaches 16.0~ms. These measurements establish concrete resource requirements for embedded DNN-based speech enhancement and quantify the remaining gap to hearing aid deployment.

Counterfactual explanations seek small, semantically meaningful changes to an input that alter a model's prediction, and are widely used to interpret and audit machine learning systems. In modern vision, language, and multimodal systems, pretrained encoders map inputs to representation spaces, and downstream classifier heads impose decision boundaries within those spaces. As a result, the feasibility and distance of nearby counterfactuals depend on boundary placement relative to the data. Yet models with similar predictive performance can differ substantially in whether such changes are achievable and how far representations must move. This work examines this variation using a standardized local search probe across several pretrained encoders and linear classifier heads. Results show that despite similar predictive performance, models differ substantially in their counterfactual behavior. Under fixed representations, varying only the classifier head alters counterfactual outcomes while leaving predictive performance largely unchanged. This variation is explained by the interaction of decision-boundary proximity and local data support, which jointly determine whether prediction changes are both feasible and lie in regions supported by the data, and can also improve counterfactual search within fixed models. Together, these findings identify counterfactual behavior as a distinct dimension beyond predictive performance and show that it can be altered without changing accuracy, with implications for model selection, robustness, and the reliability of counterfactual methods.

We address the challenge of enabling robots to manipulate deformable linear objects (DLOs), such as ropes, cables, and rubber bands. Prior work has primarily focused on narrow, task-specific problems, often relying on real-world demonstrations or handcrafted heuristics. Such approaches, however, struggle to scale to the wide variety of materials and tasks encountered in practice, and collecting sufficiently diverse real-world data is often impractical. Additionally, existing simulation environments offer limited support for the broad spectrum of material behaviors necessary for generalizable DLO manipulation. To overcome these limitations, we introduce a differentiable simulator explicitly designed for versatile DLO manipulation. Our simulator models a wide range of material properties-including (in)extensibility, elasticity, bending plasticity, and complex interactions with other objects-providing a robust foundation for learning and evaluating manipulation skills. Building on this simulator, we propose a benchmark suite of representative tasks that highlight the unique challenges of DLO manipulation. The successful execution of these tasks is often hindered by the topological complexity and grasp sensitivity inherent to DLOs. Therefore, we introduce a specialized DLO agent that explicitly manages these challenges by proposing strategic grasping points and decomposing long-horizon tasks to maximize control authority. Finally, we evaluate various policy-learning algorithms using our framework, alongside sim-to-real transfer experiments, demonstrating our platform's potential to advance DLO manipulation.

As LLMs become more widely deployed, they are increasingly expected to work alongside other AI agents rather than operating in isolation. Effective coordination in these settings requires agents to communicate, share information and make decisions under uncertainty. We introduce SMAC-Talk, a natural language extension of the StarCraft Multi-Agent Challenge for evaluating LLM-based agents in cooperative multi-agent environments. The environment has several key features such as decentralized control, partial observability and long-horizon decision making. SMAC-Talk includes a natural language communication channel which is used to probe agent coordination and trust. We use this communication channel to construct different evaluation scenarios, including settings with an embedded deceptive communicator that tries to disrupt and deceive allies through communication alone. We provide three agents for benchmarking using 4 models from the Qwen3.5 family and study how reasoning structure, memory and model scale affect coordination between agents. We release SMAC-Talk as an open benchmark to support the research community in developing and evaluating LLM agents in cooperative multi-agent settings.

The increasing use of large language models has raised concerns about the spread of AI-generated fake news, particularly under varying prompting strategies. Most existing detection models are trained and evaluated under a single generation setting, leaving their ability to generalize across unseen prompts unclear. In this study, we investigate cross-prompt generalization in fake news detection using three datasets of AI-generated articles produced under distinct prompts, combined with real news articles. We extract interpretable linguistic features capturing lexical diversity, readability, and emotion-based characteristics and evaluate a random forest classifier under a cross-prompt framework, where models trained on one prompt are tested on another. Across all six train-test combinations, performance remains consistently high, with AUC values ranging from 0.988 to 1.000. Analysis of feature distributions shows that AI-generated text exhibits increased lexical diversity, reduced readability, and substantially lower emotional intensity compared to the overall dataset, with variations across prompts. Despite these distributional shifts, the classifier maintains strong performance, indicating that these features capture stable properties of AI-generated text that generalize across prompting strategies. These findings suggest that feature-based approaches can provide robust detection of AI-generated fake news under prompt variability.

Remote photoplethysmography (rPPG) achieves low heart-rate error on uncompressed benchmarks yet is deployed over compressed video channels in telehealth, neonatal ICU, and driver fatigue applications. No prior work identifies the physical quantity determining when spatial decomposition outperforms global-projection methods under codec compression. We propose Spatial Artifact Coherence (SAC), defined as the ratio of off-diagonal to diagonal energy in the 4x4 inter-patch Green-channel covariance matrix (bandpass 0.75-2.5 Hz), and the PatchPCA algorithm family (four codec-aware rPPG algorithms). We evaluate 280 subjects across three public datasets, 11 codec degradation variants (MPEG-4, H.265, H.264, JPEG, chroma subsampling), and 13 algorithms via Wilcoxon tests (BH-FDR, q < 0.05, 904 tests). SAC explains 93.8% of between-variant variance in PCA advantage (r = +0.969), with zero overlap between codec families: non-MPEG-4 variants cluster at SAC 0.10-0.18 with 84-90% PCA win rates, while MPEG-4 variants cluster at SAC 0.48-0.59 with 61% win rate and a 5.8x reduction in mean improvement. Within subjects, 78% confirm the expected pattern (p < 10^-22, dz = 0.73). Within-variant subject-level SAC correlation is r = +0.099, confirming SAC classifies codec families rather than predicting individual outcomes. MPEG-4's effect is structural (macroblock DCT geometry, not noise amplitude), governed by source codec state, not resolution. P-Hybrid is identified as the most deployment-robust algorithm. Two necessary operating conditions for PatchPCA advantage are established: SAC < 0.30 and low-to-moderate motion, directly ruling out raw-to-MPEG-4 transcoding pipelines. SAC provides a physically grounded metric for codec-aware rPPG algorithm selection in clinical remote monitoring systems.

Retrieving the few past turns that answer a new query across long multi-session histories is the retrieval bottleneck behind long-term conversational memory (LoCoMo, LongMemEval). Recent concurrent work, Nano-Memory, shows that scoring a session by the maximum query-turn similarity (late interaction, "Turn Isolation Retrieval") beats mean-pooled session embeddings. We do not claim that effect; we replicate it and ask what a training-free, CPU-only retrieval stage should add around it. We report four findings. (1) Fuse: score-level fusion of the late-interaction dense score with BM25, under a single leave-one-conversation-out weight, adds +8.8 to +17.2 points of LoCoMo Hit@1 over late interaction alone across six encoders (all p<1e-4), reaching Hit@1 0.752 / NDCG@5 0.829 (e5-large-v2), +11.2 pp over BM25. (2) An off-the-shelf web-search cross-encoder reranker over the fused top-10 hurts here, degrading Hit@1 by 6.9 pp (one reranker, one configuration). (3) A pooling-operator ablation shows top-k late interaction matches max-similarity, but a naive smooth-max (log-sum-exp) collapses for half the encoders. (4) The late-minus-early gap is large for all six encoders and tends to be larger for larger ones, while the marginal fusion gain shrinks; on LongMemEval-S, a lexical regime where BM25 saturates, the net fusion gain over BM25 is small and not significant. A per-category analysis frames the gain as a division of labor: dense late interaction helps most on multi-hop and temporal questions but trails BM25 on adversarial ones. The contribution is a controlled, reproducible account of a strong training-free retrieval recipe, not the late-interaction retriever itself (Nano-Memory's). We make no claim to a complete memory architecture; this is a retrieval-stage study.

We describe our approach to the CTF4Science Lorenz challenge, a benchmark that mixes short-horizon forecasting, long-time distribution matching, and trajectory reconstruction across nine task pairs. The key discovery is that no single model family dominated all metrics. Instead, we built a metric-aware hybrid system that assigned a different predictor to each metric family: (1) synthetic-pretrained denoisers for full-trajectory reconstruction, (2) Lorenz ODE fitting and trajectory shooting for the first 20 forecast steps, and (3) histogram-tail substitution using synthetic Lorenz libraries for long-time evaluation. A representative mature submission from this system family scored 83.83551 on the public leaderboard, and a small follow-up stack of the same ideas reached 83.85529. We focus on the cleaner intermediate system because it captures the full method while remaining simple enough to reproduce and analyze, while the final submission can be understood as a conservative extension of the same backbone.

Aspect-Based Sentiment Analysis (ABSA) requires high-quality datasets to train reliable models. However, existing annotation tools treat output as flat files, leaving researchers to manually consolidate multi-annotator data, reconstruct relational structures, and compute reliability metrics through custom scripts. This paper introduces ACAT (Aspect-based sentiment analysis Collaborative Annotation Tool), a web-based platform natively supporting four ABSA workflows: (1) Aspect-Category Sentiment Analysis, (2) Clause-Level Segmentation, (3) Aspect-Term Sentiment Analysis with character-level position tracking, and (4) Aspect Sentiment Triplet Extraction with dual span offset preservation. Its core contribution is an automated Extract, Transform, Load (ETL) pipeline that aligns collaborative annotations and computes Inter-Annotator Agreement (IAA) metrics directly at export, yielding training-ready datasets. In a preliminary validation on 1,002 restaurant reviews with two annotators of differing expertise, ACAT achieves a median annotation time of 31.58 seconds and a raw IAA ranging from 0.78 to 0.86 across all tasks.

Offline goal-conditioned reinforcement learning requires both long-horizon reachability estimates and local action comparisons. Dual goal representations provide value fields that capture global goal reachability, but they do not directly specify which action should be preferred at a given state. We propose Dual Advantage Fields, a policy-extraction method that turns a bilinear dual value model into a local advantage signal. Under bilinear dual parameterization, the goal embedding is the gradient of the value field with respect to the state representation. DAF learns an action-effect model that predicts the discounted feature displacement induced by an action and scores actions by the alignment between this displacement and the goal direction. In the realizable case, this score equals the goal-conditioned Bellman advantage, yielding a standard local policy-improvement guarantee. On OGBench locomotion, manipulation, and puzzle tasks, DAF improves aggregate RLiable metrics and performs strongly in settings where locally correct actions differ from direct movement toward the final goal.

Safe navigation in dynamic and uncertain environments often relies on accurate estimation of, or assumptions about, the true underlying uncertainty. However, accurately characterizing the true uncertainty distribution is often difficult due to limited data or imperfect information. An incorrect understanding of the uncertainty and its associated risk may lead to dangerous decisions even under high levels of risk aversion. To address this issue, we propose a risk-aware model predictive control (RA-MPC) framework that incorporates prediction sets to guarantee risk control below a user-specified threshold without requiring assumptions about the underlying uncertainty distribution. To generate the prediction sets, we develop a distribution-free risk quantification framework that extends conformal risk control (CRC) to general spectral risk measures. We then show that incorporating the prediction sets into the MPC framework provides statistical safety guarantees in terms of spectral risk constraint satisfaction even under uncertainty misspecification. We validate the proposed framework in simulated vehicle obstacle avoidance scenarios, demonstrating improved safety and reduced solve time compared to a baseline RA-MPC framework.

We formulate the problem of \emph{exact unlearning} in reinforcement learning, where the goal is to design an efficient framework that enables the removal of any user's data upon deletion request, i.e., the online learner's output after unlearning is \emph{indistinguishable} from what would have been produced had the deleted user never interacted with the learner. For any $ρ>0$, we show that there exists a reinforcement learning (RL) algorithm that is $ρ$-TV-stable and supports an exact unlearning procedure whose expected computational cost is only a $ρ\sqrt{\ln T}$ fraction of the computational cost of retraining from scratch. We construct such a $ρ$-TV-stable RL algorithm for tabular Markov decision processes (MDPs), which achieves a regret bound of $\mathcal{O}(H^2 \sqrt{SAT} + H^3 S^2 A + {H^{2.5} S^2 A}/ρ)$, where $S, A, H$, and $T$ denote the number of states, the number of actions, the episode horizon, and the number of episodes, respectively. We also establish a lower bound of $Ω(H\sqrt{\!SAT}\! +\! {SAH}/ρ)$ for $ρ$-TV-stable RL algorithms, showing that our algorithm is nearly minimax optimal.

Vision-language foundation models such as CLIP and SigLIP provide widely used representations for multimodal learning systems. While these models are typically compared through downstream performance, such evaluations often do not explain how their representations differ structurally. In this work, we study this problem through the task of Contrastive Embedding Clustering: identifying sample subsets that are weakly clustered under one representation but strongly clustered under another. We propose \emph{Kernel Optimization for Discrepancy Analysis (KODA)}, a kernel-based framework for contrastive representation comparison and alignment. KODA constructs unified multimodal kernels through modality-wise kernel composition and formulates discrepancy discovery as a constrained optimization problem that searches for coherent structures in one representation while suppressing coherence in a reference representation. This yields interpretable discrepancy directions associated with specific sample subsets and modality interactions. To scale KODA to large vision-language datasets, we develop randomized low-dimensional approximations of joint kernels using random projections, including Random Fourier Features for shift-invariant kernels. Empirically, KODA identifies consistent and interpretable discrepancy structures across vision-language representations and provides sample subsets for representation alignment. The code is available at https://github.com/yokiwuuu/KODA.

Interpretable linguistic features offer a promising approach for explaining why a given text appears machine-generated, particularly for non-expert users. However, existing findings on which features reliably indicate LLM-generated text remain fragmented across feature sets, models, and text domains. To address this gap, we conduct a large-scale empirical study assessing the robustness of linguistic signals for characterizing AI-generated text. Our analysis covers 284 interpretable linguistic features across outputs from 27 LLMs and ten text domains under cross-model and cross-domain generalization settings. We show that classifiers based solely on linguistic features can reliably distinguish AI-generated from human-written text. However, many previously proposed indicators prove strongly context-dependent, with the exception of measures of lexical richness, which remain robust signals across model families and text domains. These results demonstrate which linguistic signals generalize across contexts and provide a foundation for more reliable, interpretable analyses of AI-generated language.

We study a distributional generalization of the matrix completion problem in which each entry of the target matrix is a probability distribution rather than a scalar. In this setting, only a subset of matrix entries is observed, and even for observed entries, the underlying distributions are not directly accessible; instead, we observe finitely many samples drawn from them. To represent distributional entries, we employ kernel mean embeddings and introduce a notion of Tucker rank for distribution-valued matrices to capture their low-rank structure. The infinite-dimensional nature of kernel embeddings poses significant methodological challenges. To address this, we introduce functional unfolding operators that link the proposed distributional low-rank structure to the classical Tucker rank for finite-dimensional tensors. Based on this framework, we propose a novel estimator for distributional matrix completion. We establish non-asymptotic error bounds that characterize the statistical performance of the estimator. Extensive experiments on synthetic data and a real-world application demonstrate the effectiveness of the proposed method.

Task-conditioned manipulation requires grounding instructions to task-relevant functional parts rather than object categories. This setting is scene-dependent and often one-to-many in cluttered scenes: the same object may afford different interactions across tasks, while a single task may correspond to either one functional region or multiple valid functional regions, depending on the scene layout. Existing affordance datasets and benchmarks remain misaligned with this setting, as they typically focus on grasping or object-level affordances, rely on synthetic scenes, or assume a single instruction-region correspondence. We present Affordance2Action (A2A), a benchmark-centered learning framework for scene-level, task-conditioned part affordance grounding. At its core is A2A-Bench, a manipulation-oriented benchmark that covers both single-region and multi-region instruction correspondences in everyday scenes, with the latter highlighting the ambiguity and diversity of affordance grounding in realistic multi-object environments. To construct it at scale, we build A2A-AffordGen, an agent-assisted annotation pipeline that combines language-model filtering, interactive part segmentation, instance-level mask-out refinement, task-reasoning instruction generation, and human verification. A2A-Bench's supervision further supports diverse downstream applications, with real-time affordance grounding and affordance-conditioned manipulation policies as two representative examples. Experiments show that A2A exposes substantial gaps in generic segmentation, VLM-based grounding, and affordance distillation baselines, while improving task-level localization and providing useful spatial priors for downstream manipulation. All datasets and code will be publicly released to promote open research.

Safety alignment in large language models (LLMs) is fragile in part because it is often shallow: fine-tuning mainly reshapes the model's behavior near the first few output tokens. We argue that this phenomenon can be understood through autoregressive consistency, the tendency of next-token prediction to preserve and extend the current response trajectory consistently. By analyzing the learning dynamics of safety alignment, we show that autoregressive consistency can concentrate alignment updates on early tokens, offering a mechanistic explanation for shallow safety alignment. The same mechanism also predicts a broader class of attacks on LLMs: attacks that induce harmful continuation states at arbitrary positions in the output trajectory. As a concrete example, we introduce random insertion attack, which inserts a short harmful span into an otherwise safe refusal trajectory and exploits autoregressive consistency to sustain the resulting harmful branch, thereby bypassing safety alignment. Notably, a short harmful span can redirect the generation to be harmful even after a long refusal prefix, highlighting autoregressive consistency as a potential broader failure mechanism. This suggests that safety alignment should also break harmful autoregressive consistency throughout the output trajectory. We therefore propose adversarial safety alignment, an initial framework based on worst-case harmful continuation states, and instantiate it with random worst-insertion training. Overall, our results suggest that autoregressive consistency should be treated as a central consideration in both safety alignment and attack design.

We tackle the Metro Network Expansion Problem (MNEP), a subset of the Transport Network Design Problem (TNDP), which focuses on expanding metro systems to satisfy travel demand. Traditional methods rely on exact and heuristic approaches that require expert-defined constraints to reduce the search space. Recently, deep reinforcement learning (Deep RL) has emerged due to its effectiveness in complex sequential decision-making processes-it remains, however, computationally expensive, environmentally costly, and requires additional engineering to interpret. We show that MNEP problems are small enough to not require Deep RL methods. Reformulating the MNEP as a Non-Markovian Rewards Decision Process (NMRDP), we use tabular RL to achieve similar performance with significantly fewer training episodes, additionally offering greater interpretability. Additionally, we incorporate social equity criteria into the reward functions, focusing on efficiency and fairness, highlighting the versatility of our method. Evaluated in real-world settings-Xi'an and Amsterdam-our method reduces total episodes by a factor of 18 and total carbon emissions by a factor of 12 on average, while remaining competitive with Deep RL. This approach offers a replicable, modular, interpretable, and resource-efficient solution with potential applications to other combinatorial optimization problems.

Practical text-recognition pipelines for historical documents typically decompose layout analysis into line detection followed by a separate reading-order step, with the latter most often handled by a hand-coded geometric heuristic that struggles with marginalia, multiple columns, tables, and source-specific editorial conventions. This article introduces Orli (Ordered Regression of Lines), an end-to-end model that casts both sub-tasks as a single image-to-sequence problem: from a page image, Orli autoregressively generates text-line baselines directly in reading order. Baselines are represented in a chord-frame parameterization that anchors a line's position, orientation, and extent while encoding local geometry through perpendicular offsets; an iterative refinement head and a local visual refiner produce the final curve. Trained on a heterogeneous corpus of 196,691 pages spanning ten writing systems, Orli marginally exceeds the previously reported state of the art for cBAD line detection without dataset-specific training, reaches near perfect coverage and ordering on multiple reading-order benchmarks zero-shot, and adapts to more specialized out-of-domain layouts with limited fine-tuning. The method's source code and model weights are available under an open license at https://github.com/mittagessen/orli.

Data samples used for training often differ from those encountered during fine-tuning and deployment, and while ML models show promise, their performance remains limited when only small annotated datasets are available. Performance often degrades under distribution shifts caused by diverse sensors, populations, and application settings. Although pre-training helps, models frequently encounter out-of-distribution (OOD) data in real-world settings, leading to reduced robustness. Existing adaptation methods usually assume fixed distribution shifts and struggle when multiple types or severities occur. In particular, they overlook shift severity, for example treating adaptation to a large familiar dataset the same as adaptation to a small dataset with a new task, which limits generalisation. To address this, we propose ADAPTOOD, a novel framework that leverages data uncertainty to quantify distribution shift severity and guide fine-tuning for time series. This uncertainty measures how strongly samples from the target deployment distribution deviate from the pre-training distribution, providing a direct signal of OOD severity. Our framework combines this uncertainty with low-rank model updates and adaptive hyperparameter optimisation to improve adaptation. We show that ADAPTOOD achieves up to 7% higher accuracy and 12.9% higher precision than existing methods in OOD tasks, maintaining strong performance as distribution shift severity increases.

Different predictors often excel on different inputs, so picking the best one per instance promises higher accuracy than committing to a single model. In practice, selectors trained from logged data routinely fail to beat the strongest single predictor. Three causes typically go unseparated before more tuning is applied: a mismatched learner, a state that does not predict which model wins, or buffer-to-deployment label shift. A three-stage diagnostic rules them out on a shared buffer. Stage~1 estimates a local ceiling on oracle recovery from $k$-NN label consistency. Stage~2 asks whether paired BC and offline-RL learners (BC, DQN, and CQL across penalty weights) reach that ceiling. Stage~3 ablates the selector state to test whether richer features would raise it. The combined verdict points to the most promising next step: tuning the learner, redesigning the state, or collecting new data. We apply it to selecting among five dropout-prediction models on edX clickstream data. Across 16 windows, the oracle beats the strongest single base model by 9.7 accuracy points on average, yet BC, DQN, and CQL land in the same test-accuracy band below it (robust to a tenfold buffer sweep and $N{=}2{,}000$ held-out examples). The bottleneck is local representational ambiguity: CQL closes the imitation gap without a deployment gain (not conservatism), regret clusters tightly across learners (not tie-breaking), and the three learners converge on test accuracy (not shift). The next iteration should change the state or collect new data, not tune the offline learner further.

Safety alignment in instruction-tuned large language models (LLMs) depends on a model's ability to reliably refuse to respond to harmful or disallowed requests. Recent work has shown that a steering vector can be applied to a dense LLM during inference to effectively suppress refusal behavior, inducing response to harmful requests. We extend this refusal steering method to three open-source Mixture-of-Experts (MoE) LLMs and find that steering performance is uninhibited by the complex routing patterns inherent to the MoE architecture. We then propose two expert-aware refusal steering methods that leverage refusal-specific expert routing patterns and expert-specific steering directions to suppress normal refusal behavior. We find that refusal behavior can be effectively steered based on the output of a single expert. Our results show that refusal signals captured by steering methods differ from expert routing behavior, suggesting a substantial role for attention in MoE refusal behavior.

Multi-agent Next-Best-View (NBV) selection for safe path planning in uncertain and unknown environments requires informative, safety-aware, and efficient coordination. Centralized approaches rely on sharing raw sensor data or significant communication overhead, resulting in limited scalability. We propose a distributed, risk-aware multi-agent NBV framework in which each robot maintains a private local 3D Gaussian Splatting map and the team jointly maximizes expected information gain (EIG) restricted to masked zones along planned trajectories. The resulting distributed objective is solved by Consensus ADMM (C-ADMM) over a communication graph, with each robot exchanging only candidate viewpoints, planned trajectory descriptors, and scalar EIG contributions. Collision risk along each trajectory is modeled via Average Value-at-Risk (AV@R) over the local 3DGS map and used both to shape the masking radius and to score planned paths. Experiments in Gibson environments at multiple team sizes show that the distributed formulation approaches the centralized baseline in mapping quality and trajectory safety while reducing communication by orders of magnitude.