Out-Of-Domain Generalization

Mixture-of-Experts (MoE) pretraining relies on an auxiliary load-balancing loss (LBL) to drive per-expert utilization toward uniformity. Post-training inherits a different situation: the base router already encodes non-uniform expert co-activation structure, which a re-imposed uniformity objective flattens away. We show that downstream performance depends instead on holding this inherited routing softly, a principle we term soft router anchoring, and instantiate it as Router Prior Bias (RPB), a training-time bias that pulls the router logits toward a prior read off the frozen base router while leaving the router itself trainable. On math post-training of Moonlight-16B-A3B, RPB attains 45.77 in-domain accuracy against 31.91 under re-applied LBL and 29.44 under unanchored fine-tuning, and retains more out-of-domain capability than either. The ordering against LBL reproduces on a second model family (Qwen3-30B-A3B-Base), and the advantage over LBL is resolvable on an independently sourced corpus. Anchors defined on the router weights, on its logits, or on its output distribution perform comparably with no consistent ordering, which places the effect in the softness of the constraint rather than in the particular prior RPB supplies. Retained community structure in the expert co-activation graph tracks these gains wherever the base router is non-uniform enough for communities to form, yet enforcing the same prior as a hard assignment preserves that structure while performance falls sharply. Community structure is therefore a footprint of soft anchoring rather than its source, and the practical lesson is that inherited routing should be held softly during post-training, since both flattening it toward uniformity and enforcing it absolutely carry a downstream cost. Our code will be released at https://github.com/naver-ai/rpb.

Conversational agents like chatbots and voice assistants are trained to understand and respond to user intents. On encountering an utterance with an intent different from the ones they have been trained on, these agents are expected to classify the intent as unknown' or out of domain'. This problem is known as out of domain (OOD) intent detection. Podolskiy et al. (2021), showed that Mahalanobis distance can be used effectively for identifying OOD intents, outperforming competing approaches. However, their method fails to outperform the baselines in the practically important few-shot setting. In this paper we analyze the reason for low performance and propose a covariance corrected Mahalanobis distance for detecting out-of-domain intents.

We present a lightweight approach to foundation modeling (\textbf{NEXUS}) that leverages pre-trained learning from collider physics data towards out-of-domain tasks in other scientific datasets, using a fully connected autoencoder model with approximately 3 million parameters. The model pre-trains with no supervision over a large-scale collision dataset from the Large Hadron Collider modeled by charged particle track features. Downstream tasks for collider analyses, such as kinematic regression and event classification, are developed on pre-trained model weights and achieve improved accuracy with only small labeled datasets when compared to equivalent architectures trained from scratch. The benefits of pre-training are additionally investigated through latent space interpretation and application to other domains, including gravitational waves, flood forecasting, and neural activity. Furthermore, the relative computational simplicity of NEXUS is demonstrated compared to transformer approaches at comparable scale, opening the door to power-efficient inference and real-time or edge applications of foundation models in scientific experiments.

Background/Objectives: Dermoscopic skin-lesion classifiers lose accuracy when images arrive from a new clinic or a new device. We asked which data augmentations reduce that loss, and measured the effect under a protocol that keeps policy selection separate from policy evaluation. Methods: A ConvNeXt-Large binary malignant-versus-non-malignant classifier was trained on six dermoscopic sources (25,903 images); HAM10000 and ISIC 2016-2020 were held out of training entirely. Single augmentations, photometric combinations and eleven composite policies were ranked on a development split of 1511 held-out images. The winning policy was then evaluated on a confirmation set of 8073 held-out images that took no part in that ranking and from which we removed every image sharing a lesion identifier with the training data and every image contributed by an institution represented in training. Both policies were retrained with four random seeds each and compared with an exact permutation test. Results: The mix policy raised confirmation-set ROC-AUC from 0.787 to 0.826 (+0.039; per-seed ranges 0.772-0.797 and 0.815-0.840, non-overlapping; exact permutation p=0.029), with the same direction on each contributing source. At matched sensitivity the gain is larger in clinical terms: specificity rose from 0.612 to 0.713 at a sensitivity of 0.80, and from 0.284 to 0.397 at a sensitivity of 0.95. In-domain ROC-AUC was preserved (0.938 to 0.941). On an independent clinical cohort acquired with a different device at a different institution (472 images, 22 malignant), performance was maintained (0.934 versus 0.930). Conclusions: Augmentations that model the physical causes of domain shift improve cross-source transfer at no cost to in-domain accuracy, and the improvement survives a selection-disjoint, contamination-free evaluation.

We present MagicSelector, a joint optimization framework integrating Counterfactual task decomposition, Progressive reranking, and Dynamic Top-K, designed to address the fundamental challenges of tool retrieval in agents. MagicSelector is a specialized framework capable of translating ambiguous user instructions into executable atomic subtasks and guiding high-precision tool retrieval, effectively mitigating redundant noise and severe context distraction in out-of-domain (OOD) scenarios. We empower MagicSelector with these capabilities through three key contributions: (1) a preference-guided counterfactual task decomposition mechanism that utilizes a counterfactual reward to quantify the marginal causal gain of decomposition on retrieval ranking, effectively imposing fine-grained structural supervision on logical coherence; (2) a progressive tool reranking method driven by self-distillation hard negative mining, which optimizes both point-wise and list-wise relevance to enhance fine-grained discrimination among highly similar tools; and (3) a dual semantic boundary-aware dynamic Top-K strategy that adaptively monitors reranking score cliffs and inter-tool semantic shifts to dynamically truncate the candidate list, maximizing relevant tool recall while filtering long-tail noise. Evaluated on MTDTool, the first task decomposition benchmark we constructed tailored for mobile multi-turn interactions with process-level annotations, MagicSelector yields promising performance. Extensive experiments demonstrate that MagicSelector significantly outperforms state-of-the-art methods in terms of tool retrieval accuracy, OOD generalization capability, and overall token efficiency, thereby demonstrating the effectiveness of our proposed framework.

Intent detection is a critical task that bridges human intents and system actions in human-machine interaction systems. However, there still exist challenges for detecting out-of-scope (OOS) intents. (i) The traditional methods view the OOS intent detection as a multi-class classification, then the detection accuracy decreases as the class number of the known intents increases; (ii) LLM-embedding methods require large parameters, that makes them difficult to train and practically deploy. Thus, this work proposes a multi-cluster boundary learning method to detect OOS intents via MiniLM embedding (i.e., all-MiniLM-L6-v2) in an one-class classification workflow. The method learns the boundaries of multi-cluster embeddings generated by MiniLM from the training utterances, and then rejects the out-of-domain utterances as OOS intents. Experiments are conducted on public CLINC150, StackOverflow and Banking77 datasets. The results show that the method achieves the state-of-the-art OOS intent detection performance compared the other baselines. Ablation studies are also conducted and the results show that the used MiniLM can better adapt to the workflow and utterance embedding requirements. The code is available at supplementary materials.

Predicting the behavior of dynamical systems (DS) beyond the dynamical and parameter regimes observed in training is a pivotal and essentially unresolved problem in scientific ML. It is central to any good scientific theory, which we expect to be able to make predictions about regimes not covered by currently available data. Recent hierarchical and hyper-network guided approaches for DS reconstruction (DSR) enable training on many DS simultaneously, and revealed that extracted latent features are often related to crucial control parameters of the underlying DS that varied across the training corpus. However, true out-of-domain forecasting abilities of these models, e.g., across tipping points, remain limited, and fine-tuning, or even full model retraining, on time series from the new dynamical regime is usually required. Here, we mathematically analyze the root of these limitations in previous model formulations and identify three core shortcomings rooted in a mismatch between structural assumptions of the reconstruction model and typical properties of physical systems. We propose a combination of remedies for these shortcomings, most importantly feature splitting, and furthermore derive a closed-form bound on the reliable extrapolation range. We demonstrate empirically that our techniques allow for accurate zero-shot prediction into new dynamical regimes, outside the observed training regime, as, e.g., encountered across tipping points.

Scientific reasoning models for biology combine language models with foundation models trained on multimodal biological data, including DNA, RNA, and proteins. These models are built through post-training, yet how each stage shapes reasoning and generalization remains poorly understood. We study when post-training improves performance and when it induces over-specialization. Across genomics, transcriptomics, and proteins, we train and evaluate more than 100 biological reasoning models under controlled variation in backbone, continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL), measuring both in-domain (ID) and out-of-domain (OOD) performance. We find that each post-training stage reshapes generalization in a distinct way rather than contributing uniform gains. CPT improves downstream performance by aligning models with biological language. SFT consistently increases ID performance but causes OOD performance to peak early and decline as models fit the training distribution. RL, when applied to strong SFT checkpoints with aligned rewards, improves OOD performance and partially recovers generalization. These results show that biological reasoning does not improve monotonically with additional supervision or compute. Instead, performance depends on how training stages are composed. Under fixed post-training budgets, the strongest ID-OOD trade-off comes from brief SFT, larger RL allocations, and asymmetric adaptation capacity across stages.

We present SpAArSIST, a deployment-oriented refinement of the widely used AASIST graph pooling backend for self-supervised learning (SSL) based anti-spoofing. Motivated by redundant operations in public implementations, we replace learned pooling and stack-node attention with explicit, lightweight choices: separate train and inference graph pooling ratios $(k_{\mathrm{tr}},k_{\mathrm{inf}})$, magnitude-based node scoring, and mean aggregation of graph nodes. The best overall configuration (rank 1) cuts backend compute by 20.7% (195.045M $\rightarrow$ 154.706M MACs) and model size by 4.1% (611.8k $\rightarrow$ 586.4k params), while improving out-of-domain robustness on In-the-Wild to 2.82% EER and 0.078 minDCF (from 4.64% and 0.133) and remaining competitive on ASVspoof5. We further provide a composite selection score that summarizes accuracy, calibration, and compute to support balanced deployment-oriented model choice.

Equation discovery aims to automate the discovery of scientific models in the form of mathematical equations from data. Technically, equation discovery is implemented by symbolic regression algorithms. Performance of symbolic regression for equation discovery is measured along two dimensions: Prediction accuracy on test data, and recovery of known groundtruth formulas. For standard regression, accuracy is typically measured on in-domain test data, for instance, by splitting a data set randomly into training and test data. While this makes sense for in-domain interpolation, which is the common goal in ordinary regression, it can be a misleading proxy for true model discovery and generalization. The obvious alternative is to measure out-of-domain accuracy. However, obtaining challenging out-of-domain test data is a non-trivial problem. Therefore, we focus on equation recovery for evaluating symbolic regression algorithms for equation discovery. The rationale is that symbolic regression algorithms that perform well in recovering known groundtruth formulas are good candidates to perform well in unknown equation discovery. Existing benchmarks for symbolic regression include equation recovery tasks, however, with only a small number of groundtruth formulas that are publicly known. Moreover, these benchmarks place less emphasis on evaluating the robustness of algorithms in terms of their behavior under changing dimensionality, sampling size, sampling distribution and sampling domain. This, however, is of central importance to practitioners wanting to discover equations for modeling natural phenomena, since data is almost certainly noisy and comes from diverse domains, distributions, and sample sizes. To fill this gap, we introduce the Equation Recovery Benchmark (ERBench), a new evaluation framework designed to rigorously assess algorithms explicitly targeting the task of equation discovery.

Large language models are now adapted through chains of post-training stages rather than through a single instruction-tuning pass. This paper studies whether such sequential post-training gradually compresses internal representations into low-rank, anisotropic, and homogeneous feature spaces. We define a measurement suite for hidden states, logits, token trajectories, and LoRA updates, and we use it to analyze supervised fine-tuning, preference optimization, safety/refusal tuning, math and code specialization, and long chain-of-thought tuning under controlled stage orderings. The central hypothesis is that excessive representation concentration is not merely a geometric curiosity: it predicts reduced plasticity during later adaptation, weaker out-of-domain generalization, and poorer calibration. We further evaluate lightweight interventions, including mixed-domain replay, feature refresh, representation diversity regularization, and LoRA update decorrelation, as ways to preserve future learnability without giving up the behavioral gains of post-training.

Modern learning systems excel at interpolation but struggle to generalize to unseen tasks outside the training distribution's support. This failure occurs even in simple settings, such as handling task parameters beyond the training range, and persists despite advances in foundation models. To this end, we develop the Relational Task Extrapolator (RTE), an algorithm designed to enable systematic extrapolation to novel tasks. The key observation is that extrapolation is inherently relational: extrapolating to unseen tasks requires learning how tasks transform into one another. If a model learns the transformation between tasks A and B during training, it can apply that same transformation to relate known tasks to unseen ones at test time. RTE operationalizes this idea by decomposing each target task into a known anchor task and a transformation linking the anchor and target. It then learns a relational operator, mapping an anchor-transformation pair to predictions for the target task. We instantiate RTE across multiple task extrapolation regimes in function prediction, e.g. where target tasks use out-of-range parameters (parameter extrapolation), have greater compositional depth (length extrapolation), and/or recombine function primitives in unseen ways (compositional extrapolation). We further extend RTE to sequence prediction, integrating it into fine-tuning algorithms for foundation models. Across empirical studies, we find that RTE substantially outperforms existing approaches on extrapolation to novel, unseen tasks.

Public datasets such as DLC-2021, SynID, and KID34K have significantly contributed to research on presentation attack detection for identity documents, including screen replay attacks. However, evaluation of out-of-domain (OOD) robustness remains insufficiently explored, especially under realistic domain shifts. In this work, we introduce Receipt Replay OOD, a small out-of-domain benchmark for screen replay detection. Receipts share several characteristics with identity documents, including planar geometry, curved corners, wear-and-tear artifacts, and text or logo patterns, while avoiding personally identifiable information constraints commonly associated with identity documents. We evaluate document replay detection models under cross-domain conditions and demonstrate the impact of domain shift on generalization performance. The dataset is publicly available.

Self-distillation bootstraps large language models (LLMs) by training on their own generations. However, existing methods either rely on external signals to curate self-generated outputs (e.g., correctness filtering, execution feedback, and reward search), which are costly and unavailable for the best-performing frontier models, or skip curation entirely and train on all raw outputs, an approach that is often domain-specific and hard to generalize. Both also share a deeper weakness that self-generated outputs entangle task-relevant capability with others, such as stylistic patterns, formatting artifacts, and model-specific errors, diluting the signal for the specific capability one aims to improve. In this paper, we propose Self-Policy Distillation (SPD), which achieves generalizable, capability selective without any external signal. Specifically, SPD extracts a low-rank capability subspace from the model's own gradients on correctness-defining tokens, projects key-value (KV) activations into this subspace during self-generation, and fine-tunes on the resulting raw outputs with standard next-token prediction loss. Through extensive experiments across code generation, mathematical reasoning, and multiple-choice QA, we show that SPD achieves up to 13% improvement over state-of-the-art self-distillation methods without external signals and up to 16% improvement over pre-trained baselines. Notably, SPD demonstrates superior generalizability, achieving 15% better performance under out-of-domain generalization settings.

Large language models (LLMs) have enabled web agents that follow natural language goals through multi-step browser interactions. However, agents fine-tuned on specific trajectories and domain often struggle to generalize out of domain, and offline training can be compute-inefficient due to noisy, redundant trajectories and long accessibility-tree (AXTree) states. To address both issues, we propose Weasel, a trajectory selection method for offline training of web agents. Weasel selects a fixed-budget subset of trajectory steps by optimizing an objective that balances unary importance with pairwise diversity over states, websites, and interaction patterns, solving efficiently with a greedy algorithm. We further improve efficiency with target-centered AXTree pruning that keeps only content around the ground-truth action target, and we mitigate style mismatch for reasoning-native models by replacing expert traces with model-generated, style-consistent rationales. Across AgentTrek and NNetNav training datasets, evaluations in WebArena, WorkArena, and MiniWob, and experiments with Qwen2.5-7B, Gemma3-4B, and Qwen3-8B, Weasel improves out-of-domain performance while reducing training cost, producing roughly 9.7-12.5$\times$ training speedups over standard fine-tuning. We make the code available at https://github.com/fatemehpesaran310/weasel.

Elections represent a crucial milestone in a nation's ongoing development. To better understand the political rhetoric from various movements, ranging from left to right, we propose a transformer-based model capable of projecting the political orientation of a text on a continuous left-to-right spectrum, represented by a normalized scalar d between -1 and 1. This approach enables analysts to focus on specific segments of the political landscape, such as conservatives, while excluding liberal and far-right movements. Such a task can only be achieved with multiclass classifiers, provided that the desired orientation is incorporated within one of their predefined classes. To determine the most suitable foundation model among 13 candidate transformers for this task, we constructed four distinct corpora. One corpus comprised annotated plenary notes from the German Bundestag, while another was based on an official online decision-making tool, Wahl-O-Mat. The third corpus consisted of articles from 33 newspapers, each identified by its political orientation, and the fourth included 535,200 tweets from 597 members of the 20th and 21st German Bundestag. To mitigate overfitting, we used two distinct corpora for training and two for testing, respectively. For in-domain performance, DeBERTa-large achieved the highest F1 score F1=0.844 as well as for the X (Twitter) out-of-domain test ACC=0.864. Regarding the newspaper out-of-domain test, Gemma2-2B excelled (MAE = 0.172). This study demonstrates that transformer models can recognize political framing in German news at the level of public opinion polls. Our findings suggest that both the model architecture and the availability of domain-specific training data can be as influential as model size for estimating political bias. We discuss methodological limitations and outline directions for improving the robustness of bias measurement.

As LLM-based agents increasingly rely on external tools, it is important to evaluate their ability to sustain tool-grounded reasoning beyond familiar workflows and short-range interactions. We introduce AgentEscapeBench, an escape-room-style benchmark that tests whether agents can infer, execute, and revise novel tool-use procedures under explicit long-range dependency constraints. Each task defines a directed acyclic dependency graph over tools and items, requiring agents to invoke real external functions, track hidden state revealed incrementally, propagate intermediate results, and submit a deterministically verifiable final answer. AgentEscapeBench includes 270 instances across five difficulty tiers and supports fully automated evaluation. Experiments with sixteen LLM agents and human participants show that performance drops sharply as dependency depth increases: humans decline from 98.3% success at difficulty-5 to 80.0% at difficulty-25, while the best model drops from 90.0% to 60.0%. Trajectory analysis attributes model failures mainly to breakdowns in long-range state tracking, clue adherence, and intermediate-result propagation. These findings suggest that current agents can often handle local tool use but still struggle with deep contextual dependencies. We hope AgentEscapeBench can serve as a diagnostic testbed for measuring current agent capabilities and informing future training efforts toward more robust general-purpose reasoning, action, and adaptation.

In computer vision and machine learning for geographic data, out-of-domain generalization is a pervasive challenge, arising from uneven global data coverage and distribution shifts across geographic regions. Though models are frequently trained in one region and deployed in another, there is no principled method for determining when this cross-region adaptation will be successful. A well-defined notion of distance between distributions can effectively quantify how different a new target domain is compared to the domains used for model training, which in turn could support model training and deployment decisions. In this paper, we propose a strategy for computing distances between geospatial domains that leverages geographic information with Optimal Transport methods (GeoSpOT). In our experiments, GeoSpOT distances emerge as effective predictors of cross-domain transfer difficulty. We further demonstrate that embeddings from pretrained location encoders provide information comparable to image/text embeddings, despite relying solely on longitude-latitude pairs as input. This allows users to get an approximation of out-of-domain performance for geospatial models, even when the exact downstream task is unknown, or no task-specific data is available. Building on these findings, we show that GeoSpOT distances can preemptively guide data selection and enable predictive tools to analyze regions where a model is likely to underperform.

In the paradigm of computer vision-based precise real-world scene understanding, joint reasoning in terms of contextual understanding about the objects present in a scene, their inter-object relations, and the action being performed is an essential prerequisite. However, prior works have not addressed all three jointly, and no large-scale dataset provides grounded annotations at all three levels across diverse visual scenarios. Hence, this work introduces the SCLARO (Scene-Contextual Localisation of Actions, Relations & Objects) dataset, consisting of 615,805 images spanning indoor, outdoor, and driving scenarios, annotated with global action captions, object bounding boxes, and relation triplets that supply structured scene context beyond a free-text caption. To benchmark the dataset, we propose ScenarioCLIP, a tri-level reference model that jointly encodes global scene context, objects, and inter-object relations using disentangled encoders and EMA-based knowledge distillation. We benchmark across a comprehensive suite of tasks on the SCLARO Dataset, namely zero-shot retrieval, linear probe, object detection, predicate classification, scene-graph classification, and out-of-domain generalisation. ScenarioCLIP's disentangled encoders improve over the previous works, such as PyramidCLIP's shared encoder, most notably at the object and relation levels and on out-of-domain generalisation. Code for the data generation pipeline and ScenarioCLIP is available at https://github.com/scenario-clip/SCLARO-ScenarioCLIP

Retrieval-Augmented Generation (RAG) systems are increasingly deployed in high-stakes domains, where safety depends not only on how a system answers, but also on whether a query should be answered given a knowledge base (KB). Out-of-domain (OOD) queries can cause dense retrieval to surface weakly related context and lead the generator to produce fluent but unjustified responses. We study lightweight, KB-aligned OOD detection as an always-on gate for RAG systems. Our approach applies PCA to KB embeddings and scores queries in a compact subspace selected either by explained-variance retention (EVR) or by a separability-driven t-test ranking. We evaluate geometric semantic-search rules and lightweight classifiers across 16 domains, including high-stakes COVID-19 and Substance Use KBs, and stress-test robustness using both LLM-generated attacks and an in-the-wild 4chan attack. We find that low-dimensional detectors achieve competitive OOD performance while being faster, cheaper, and more interpretable than prompted LLM-based judges. Finally, human and LLM-based evaluations show that OOD queries primarily degrade the relevance of RAG outputs, showing the need for efficient external OOD detection to maintain safe, in-scope behavior.