PyTorchKR Papers

최신 논문

총 524,001편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2601.04153v3cs.CV

Diffusion-DRF: Free, Rich, and Differentiable Reward for Video Diffusion Fine-Tuning

Video diffusion alignment has been heavily relied on scalar rewards. These rewards are typically derived from learned reward models in human preference datasets, requiring additional training and extensive collection. Moreover, scalar rewards provide coarse, global supervision, offering limited prompt-generation mismatch credit assignment and making models prone to reward exploitation and unstable optimization. We propose Diffusion-DRF, a free, rich, and differentiable reward framework for video diffusion fine-tuning. Diffusion-DRF employs a frozen, off-the-shelf Vision-Language Model (VLM) as the critic, eliminating the need for reward model training. Instead of relying on a single scalar reward, it decomposes each user prompt into multi-dimensional questions with freeform dense VQA explanation queries, yielding information-rich feedback. By direct differentiable optimization over this rich feedback, Diffusion-DRF achieves stable reward-based tuning without preference datasets collection. Diffusion-DRF achieves significant gains both quantitatively and qualitatively, outperforming state-of-the-art Flow-GRPO by 4.74% in overall performance on unseen VBench-2.0.

arXiv:2601.02867v3cs.CL

Training with Subword Variants Delays Overfitting in Repeated-Data Language Modeling

Subword tokenization introduces a computational layer in language models, where many distinct token sequences decode to the same surface form while preserving meaning, yet induce different internal computations. Despite this non-uniqueness, language models are typically trained using a single canonical deterministic tokenization. In this work, we use non-canonical subword variants as a strictly meaning-preserving form of data augmentation. We introduce a lightweight training architecture that conditions canonical next-token prediction on sampled subword variants via an auxiliary causal encoder and block-causal cross-attention, without modifying the training objective or token interface. In data-constrained pretraining, augmentations with subword variants consistently delay overfitting under repeated data exposure and reduce held-out next-token loss across evaluation datasets. In multilingual fine-tuning, we find that the effectiveness of training with auxiliary subwords depends on tokenizer quality: gains are strongest when canonical tokens are highly compressed and diminish when the tokenizer already over-fragments the input.

arXiv:2601.02694v2cs.NI

Which Deep Learner? A Systematic Evaluation of Advanced Deep Forecasting Models Accuracy and Efficiency for Network Traffic Prediction

Network traffic prediction is an essential tool for automating modern network management. It is a difficult time series forecasting (TSF) problem that has been addressed by Deep Learning (DL) models due to their ability to capture complex patterns. Recent years have witnessed advances in deep forecasting models, from sophisticated transformer architectures to simple linear models, improving performance across diverse prediction tasks. However, given the diversity of these models, the variability of network traffic across environments and timescales, and the inherent challenges of network traffic prediction, it is essential to identify effective deployment choices and modeling directions. Henceforth, this study systematically identifies 12 deep forecasting models, comprising 8 advanced TSF models and 4 traditional deep learning architectures -- each with unique advantages for network traffic prediction -- against 3 statistical baselines, for a total of 15 forecasting methods and evaluate them on four real traffic datasets, across multiple time scales and horizons, assessing performance, robustness to anomalies, data gaps, external factors, data efficiency, and resource efficiency in terms of time, memory, and energy. Our results provide insights into model behaviour and guidelines for network management, highlighting performance regimes, efficiency thresholds, and promising architectures that balance accuracy, efficiency, and robustness to traffic challenges.

arXiv:2601.01075v3cs.LG

Flow Equivariant World Models: Structured Memory for Dynamic Environments

Embodied systems experience the world as 'a symphony of flows': a combination of many continuous streams of sensory input coupled to self-motion, interwoven with the dynamics of external objects. These sensory streams and the underlying dynamics of the world obey smooth, time-parameterized symmetries which existing world models ignore. Without a memory that respects this structure, partial observability presents a major obstacle to existing methods: each observation reveals only a fraction of the world, while unobserved regions continue to evolve. In this work, we introduce Flow Equivariant World Modeling, a framework that leverages time-parameterized symmetries within a latent memory for stable and accurate dynamics prediction over long horizons. The latent memory shifts and transforms equivariantly with self-motion and inferred external object motion, keeping information about out-of-view regions aligned as time progresses. We demonstrate the advantage of this framework over state-of-the-art diffusion, memory-augmented, and recurrent world model architectures on 2D and 3D partially observed video world modeling benchmarks. More broadly, our results suggest that predictive representations become more powerful when they are organized in line with the temporal and dynamical structure of the world they model. Project page: https://flowequivariantworldmodels.github.io/

arXiv:2512.17220v3cs.CL

Mindscape-Aware Retrieval Augmented Generation for Improved Long Context Understanding

Humans understand long and complex texts by relying on a holistic semantic representation of the content. This global view helps organize prior knowledge, interpret new information, and integrate evidence dispersed across a document, as revealed by the Mindscape-Aware Capability of humans in psychology. Current Retrieval-Augmented Generation (RAG) systems lack such guidance and therefore struggle with long-context tasks. In this paper, we propose Mindscape-Aware RAG (MiA-RAG), the first framework to formulate mindscape-aware retrieval and generation as a unified conditioning paradigm for LLM-based RAG. MiA-RAG builds a mindscape through hierarchical summarization and conditions both retrieval and generation on this global semantic representation. This enables the retriever to form enriched query embeddings and the generator to reason over retrieved evidence within a coherent global context. We evaluate MiA-RAG across diverse long-context and bilingual benchmarks for evidence-based understanding and global sense-making. It consistently surpasses baselines, and further analysis shows that it aligns local details with a coherent global representation, enabling more human-like long-context retrieval and reasoning.

arXiv:2512.15211v5cs.CV

WLC: Weber-Inspired Local Contrast Metric for Low-Altitude Image Fusion

Infrared and visible image fusion is a pivotal technology in low-altitude Unmanned Aerial Vehicle (UAV) reconnaissance missions, enabling robust target detection and tracking by integrating thermal saliency with environmental textures. However, the advancement of fusion algorithms is hindered by a critical evaluation bottleneck. In this paper, we identify a systematic failure in traditional no-reference metrics (specifically Statistics-based and Gradient-based metrics) within complex low-light environments, termed as ``Noise Trap''. It is mathematically proven that these metrics are positively correlated with high-frequency sensor noise, paradoxically assigning higher scores to degraded images and misguiding algorithm optimization. To resolve this dilemma, this paper proposes the Weber-inspired Local Contrast (WLC) metric. Grounded in the psychophysical principle of Weber's Law, WLC shifts the evaluation paradigm from global statistical distribution to local semantic contrast. By leveraging infrared priors, it effectively decouples target saliency from global background noise. Extensive experiments on the DroneVehicle dataset demonstrate that WLC exhibits high ``Semantic Discriminability'' in distinguishing thermal targets from background clutter. Furthermore, it achieves remarkable computational efficiency, thus, establishing itself as a reliable and real-time standard for intelligent UAV systems.

arXiv:2512.07988v4cs.LG

HOLE: Homological Observation of Latent Embeddings for Neural Network Interpretability

Deep learning models have achieved remarkable success across various domains, yet their learned representations and decision-making processes remain largely opaque and hard to interpret. This work introduces HOLE (Homological Observation of Latent Embeddings), a method for analyzing and interpreting discriminative neural networks through persistent homology. HOLE extracts topological features from intermediate activations and presents them using a suite of visualization techniques, including cluster flow diagrams, blob graphs, and heatmap dendrograms. These tools facilitate the examination of representation structure and quality across layers. We evaluate HOLE using a range of discriminative models, focusing on representation quality, interpretability across layers, and robustness to input perturbations and model compression. The results indicate that topological analysis reveals patterns associated with class separation, feature disentanglement, and model robustness, providing a complementary perspective for understanding and improving deep learning systems.

arXiv:2511.20836v4cs.CL

Structured Prompts Improve Evaluation of Language Models

As language models (LMs) are increasingly adopted across domains, high-quality benchmarking frameworks are essential for guiding deployment decisions. However, frameworks such as Holistic Evaluation of Language Models (HELM) typically evaluate models under a single static prompt configuration, even though model behavior depends strongly on prompt choice. As a result, reported scores can reflect prompt choice as much as model capability. Declarative prompting frameworks, such as DSPy, offer a scalable way to operationalize this evaluation protocol by evaluating models under structured prompting strategies rather than static prompts. Here, we conduct a reproducible DSPy+HELM benchmarking study and quantify how prompt choice impacts benchmark outcomes. Using five prompting methods, we evaluate four frontier and two open-source LMs across seven benchmarks against HELM baseline scores and find that prompt choice can materially impact leaderboards. Structured prompting improves performance (by 6% on average), alters comparisons (leaderboard rankings shift on $5/7$ benchmarks); chain-of-thought provides the strongest cost-performance tradeoff while more advanced optimizers offer smaller additional gains. Beyond the known intuition that prompts affect outputs, our benchmarking study integrates structured prompting into an established evaluation framework and quantifies how prompt choice alone impacts leaderboards. We open-source (i) DSPy+HELM Evaluation (https://github.com/stanford-crfm/helm/pull/3893) and (ii) Prompt Optimization Pipeline (https://github.com/StanfordMIMI/dspy-helm).

arXiv:2511.18571v2cs.LG

SAMBA: Toward an EEG Foundation Model via Differential Mamba and Coordinate-Driven Spatial Embedding

Modeling long-context electroencephalography (EEG) signals is critical for learning generalizable EEG representations, as EEG recordings are acquired at high sampling rates and often span extended durations that capture evolving neural states beyond short task- specific windows. Transformer-based models have shown promise in modeling short sequences of a few seconds; however, their quadratic complexity limits scalability to longer contexts. Moreover, substantial variability in electrode montages across datasets and recording devices poses a major challenge for building EEG models that generalize across domains. We propose SAMBA, a self-supervised learning framework with a Mamba- based U-shaped encoder-decoder architecture, which effectively captures long-range temporal dependencies and montage variability in EEG data. Leveraging the inherent ability of Mamba in processing long context sizes, we introduce: (1) Temporal Semantic Random Masking for semantic-level sequence reconstruction, (2) a Multi-Head Differential Mamba module to suppress redundancy and emphasize salient temporal structures, and (3) a Coordinate-Driven Spatial Embedding that learns unified embeddings in a three- dimensional Euclidean space, enabling robustness across devices. Experiments on thirteen EEG datasets across diverse tasks, electrode configurations, and sequence durations demonstrate that SAMBA achieves competitive or improved performance compared to recent state-of-the-art methods, while maintaining low memory consumption and efficient inference. The learned spatial weight maps exhibit close alignment with task-relevant neurophysiological regions, highlighting the interpretability of SAMBA. The code is available at: https://github.com/Jiazhen-Hong/SAMBA

arXiv:2511.16145v3cs.LG

Rethinking the Unsupervised Paradigm in Time Series Anomaly Detection: An Empirical Study and Minimalist Baseline

Time series anomaly detection (TSAD) research predominantly focuses on unsupervised architectures to model normal data dynamics under label scarcity. However, practical deployments frequently archive historical operational incidents, questioning whether continuous unsupervised architectural scaling remains the most cost-effective pursuit. In this paper, we empirically quantify the trade-offs between supervisory signals and model complexity. We introduce STAND, a deliberately minimalist supervised baseline based on bidirectional recurrent encoding. Across five benchmark datasets evaluated under a strictly disjoint out-of-sample temporal protocol, minimalist supervised models achieve competitive or superior accuracy relative to complex unsupervised alternatives. Furthermore, controlled label-scaling and few-shot evaluations confirm that these supervisory advantages persist under constrained historical event budgets, while STAND exhibits higher prediction certainty and sharper event delineation than reconstruction-based baselines. These findings encourage the TSAD community to complement algorithmic complexity with data-centric label utilization when annotations are accessible. The source code, benchmark datasets, and an extended Supplementary Document are publicly available at https://emorzz1g.github.io/STAND.

arXiv:2511.13025v2stat.ML

Recovering Geodesic Distances from Noisy Observations

Let $(M,d,\mu)$ be a nice geodesic probability space (such as a manifold) and $X=\{x_i\}_{i\in [N]} \subset M$ a finite $\mu$-i.i.d. sample. Suppose we observe noisy distance random variables $d'(x_i,x_j)$. When can the true distances be recovered to small additive error? This question has attracted interest in two disparate fields: latent-space graph models and boundary-value inverse problems. We first show that, taking $N \asymp \epsilon^{-2d-2}|\log\epsilon|$ and using runtime $O(N^{2.5+o_d (1)})$, we can recover all distances in $X$ to additive error $O(\epsilon|\log\epsilon|)$ with high probability, assuming $\mathbb{E} d'(x_i, x_j) = F(d(x_i,x_j))$ for an unknown, increasing bilipschitz $F$ and with the $d'(x_i,x_j)$ allowed to be arbitrarily heterogeneous in their distribution as long as they are bounded in subgaussian norm. This generalizes existing work, which required $F$ to be known and made strict assumptions on the distribution of $d'(x_i, x_j)$. Achieving this generalization required a novel ``ruler-building'' procedure to learn $F$. We extend our results to a missing data case where we randomly lose access to noisy distances. Our second result uses ideas from the first to tackle a partial observability case: $X = Y \cup Z$ with $Z$ sampled from $M$ and $Y$ sampled from an observer subset satisfying visibility conditions that model the assumptions made on the boundary in boundary-value inverse problems. In this setting, we recover all pairwise distances in $Z$ to additive error $O(\epsilon)$ with $|Z| \asymp \epsilon^{-2d-2}|\log\epsilon|$, $|Y| \asymp \epsilon^{-\max\{d_\nu,4\}}$ with $d_\nu$ determined by the volume-growth properties of the probability measure on the observer set, and runtime $O(|Z|^3)$. This is, to our knowledge, the first quantitative guarantee for this observer/observed split under noise, which complements existing stability results for boundary data.

arXiv:2510.26441v4cs.CV

A-TPT: Angular Diversity Calibration Properties for Test-Time Prompt Tuning of Vision-Language Models

Test-time prompt tuning (TPT) has emerged as a promising technique for adapting large vision-language models (VLMs) to unseen tasks without relying on labeled data. However, the lack of dispersion between textual features can hurt calibration performance, which raises concerns about VLMs' reliability, trustworthiness, and safety. Current TPT approaches primarily focus on improving prompt calibration by either maximizing average textual feature dispersion or enforcing orthogonality constraints to encourage angular separation. However, these methods may not always have optimal angular separation between class-wise textual features, which implies overlooking the critical role of angular diversity. To address this, we propose A-TPT, a novel TPT framework that introduces angular diversity to encourage uniformity in the distribution of normalized textual features induced by corresponding learnable prompts. This uniformity is achieved by maximizing the minimum pairwise angular distance between features on the unit hypersphere. We show that our approach consistently surpasses state-of-the-art TPT methods in reducing the aggregate average calibration error while maintaining comparable accuracy through extensive experiments with various backbones on different datasets. Notably, our approach exhibits superior zero-shot calibration performance on natural distribution shifts and generalizes well to medical datasets. We provide extensive analyses, including theoretical aspects, to establish the grounding of A-TPT. These results highlight the potency of promoting angular diversity to achieve well-dispersed textual features, significantly improving VLM calibration during test-time adaptation. Our code will be made publicly available.

arXiv:2510.25126v4cs.LG

Why Freeze the Encoder? A Controlled Study of End-to-End Sequence-Graph Learning for User Behavior

Modeling user behavior couples two distinct signals: a per-user sequence of events and a persistent user-to-user relation graph. Standard pipelines preserve these native formats by combining them in two stages: a sequence encoder is pretrained, its per-user embeddings are frozen, and a graph model is trained on top to allow for cheap serving and cross-task reuse. We ask a simple question: does training the sequence and graph models end-to-end help, and if so, when? To isolate the impact of the training strategy, we hold a standard sequence-encoder-to-GNN pipeline fixed and vary only the optimization method; on Brightkite, a matched control that starts both pipelines from the same pretrained encoder and changes only whether it is updated attributes a 10.6-point MRR gain to end-to-end updating. Across relationship prediction and fraud detection on five real-world datasets, we find that the benefit of end-to-end training is conditional: it improves clearly over frozen two-stage pipelines on relationship prediction, where the relation graph appears to carry much of the signal, whereas gains shrink where the predictive signal is largely node-local (fraud detection), a regime in which a text-fed relational graph transformer achieves competitive results. Furthermore, recasting the data as a temporal graph yields inconsistent results compared to this native sequence-graph formulation. Ultimately, this work provides a controlled empirical account of when jointly training the two structures helps.

arXiv:2510.25023v2cs.LG

SPIRE: Shared-Private Representation Learning for Continuous Multi-Region LFP under Perturbation

Learning shared and region-specific representations from continuous multi-region local field potentials (LFPs) is challenging because neural signals are nonlinear, autocorrelated, spectrally structured, and temporally misaligned. We introduce SPIRE (Shared-Private Inter-Regional Encoder), a nonlinear, time-resolved framework that separates multi-region LFP into shared and region-specific latent trajectories using region-specific sequence encoders, constrained temporal alignment, shared-only reconstruction, and variance regularization. On synthetic benchmarks with known ground-truth factors, SPIRE preferentially recovers shared and private information despite nonlinear mixing, time-varying delays, and spectral confounding. On held-out human GPi-STN LFP, shared representations strongly predict activity across regions, while cross-region prediction from private representations remains near a structured temporal null, supporting functional rather than strict statistical separation. We then apply representations trained on off-stimulation recordings to deep brain stimulation (DBS) data. Stimulation induces larger changes and more strongly encodes stimulation condition in shared than private representations. A matched CTAE analysis reproduces the shared-over-private effect, while retaining substantially less stimulation information. These findings suggest that DBS-related changes are preferentially expressed in neural representations shared across recorded regions.

arXiv:2510.18410v3cs.LG

MAGDrop: Momentum Adaptive Gradient Dropout With Provable Generalization Guarantees

Standard dropout applies a fixed, uniform noise rate to all units, ignoring available gradient information. We propose momentum adaptive gradient dropout (MAGDrop), a Bayesian linear layer that adapts per-unit weight noise by tracking an exponential moving average of gradient magnitudes. Units with large recent gradient magnitudes receive less noise to preserve learned features; units with below-average gradient magnitudes receive more noise to promote exploration. This adaptation transforms the classifier head into a data-dependent Bayesian predictor whose mean predictor has no additional inference cost. Our primary contribution is a certificate-aware low-rank posterior, MAGDrop-Subspace, that addresses the tension between strong accuracy and tight bounds from probably approximately correct (PAC)-Bayes theory. A full-head stochastic posterior incurs a Kullback--Leibler (KL) divergence of approximately 8,625 nats. By confining the posterior to an $r$-dimensional subspace constructed from an independent prior-construction sample $S_0$ and certifying on an independent sample $S$ with $S_0\perp S$, MAGDrop-Subspace reduces KL by approximately a factor of 800 while keeping accuracy close to the baseline. On CIFAR-10 with a 13-layer convolutional neural network and $r=50$, MAGDrop-Subspace achieves a PAC-Bayes Gibbs-risk bound of 0.1882 at 90\% confidence using binary KL inversion, compared with 0.4439 for the full-head baseline. Mean-predictor accuracy is 82.31\%, compared with 83.74\% for the baseline. Complementary accuracy experiments on ImageNet-1K show that MAGDrop reaches top-1 accuracy similar to the strong baseline: 78.57\% compared with 78.51\% on the base Vision Transformer with $16\times16$ patches, and 82.14\% compared with 82.23\% on ConvNeXt-Base.

arXiv:2510.15786v3cs.RO

DexCanvas: Bridging Human Demonstrations and Robot Learning for Dexterous Manipulation

A dexterous hand succeeds or fails through contact, yet every scalable record of human manipulation (video, vision-based hand tracking, motion capture) observes how the hand moves and misses how it pushes. We present DexCanvas, a hybrid real-synthetic dataset of 7,000 hours of dexterous hand-object interaction with per-frame contact points and contact forces, grown from 70 hours of optical motion capture that spans 21 manipulation types, from power and precision grasps to in-hand manipulation. The forces come from a real-to-sim pipeline: one reinforcement learning (RL) policy, conditioned on object geometry and manipulation type, drives an actuated MANO hand in physics simulation to reproduce each captured object trajectory, and the simulator measures the contacts that produce the reproduced motion. The policy reproduces 87.5% of the demonstrations and still succeeds on 80.9% of rollouts from perturbed initial object poses; these perturbed rollouts expand the 70 captured hours a hundredfold. On six YCB objects absent from the capture set it succeeds 72.5% of the time. The contact channels make better policies: a Diffusion Policy that observes DexCanvas contact features succeeds 82.2% of the time, against 67.5% without them, over 18,522 trials. Contact-driven retargeting carries the same demonstrations to two dexterous robot hands, where the RL stage again produces stable rollouts. A first release of 111,857 force-annotated trajectories is publicly available on Hugging Face.

arXiv:2510.14542v3eess.SY

A Deep State-Space Model Compression Method using Upper Bound on Output Error

We study deep state-space models (Deep SSMs) that contain linear quadratic-output (LQO) systems as internal blocks and present a compression method with a provable output error guarantee. We first derive an upper bound on the output error between two Deep SSMs and show that the bound can be expressed in terms of the $h^2$-error norms between the layerwise LQO systems. In particular, we show that reducing the $h^2$ approximation errors of the LQO systems placed in shallow layers is effective in reducing the derived upper bound on the output error. Next, we formulate an optimization problem for the derived upper bound and develop a gradient-based MOR method. In the numerical experiments, using the IMDb task from the LRA benchmark, we demonstrate the effectiveness of the proposed upper-bound-based compression method. In particular, we show that the number of trainable parameters can be reduced by approximately 60\% without retraining while maintaining the performance of the original model.

arXiv:2510.10035v2cs.AI

Failure-Driven Workflow Refinement

Optimizing LLM-based workflows is typically formulated as a global search, where candidate workflows are evaluated based on a scalar metric. This paradigm, however, suffers from a critical flaw: information collapse. By reducing rich, multi-step execution traces to simple success/failure signals, existing methods are rendered blind to the underlying structure of failures, fundamentally preventing them from modeling the workflow's failure distribution. We reconceptualize this challenge as a distributional problem. We propose a new paradigm where the optimization goal is not to maximize a scalar score, but to directly minimize a workflow's Expected Failure Mass, i.e., the integral of its failure probability density function defined over a high-dimensional Failure Signature Space (FSS). This distributional lens allows us to move from inefficient, zero-order optimization to a principled, gradient-like descent on the failure landscape itself. We introduce CE-Graph, a framework that operationalizes this paradigm through a novel, failure-driven refinement process. CE-Graph approximates the failure distribution from a pool of counterexamples, identifies its densest regions as recurring failure modes, and applies targeted, operator-constrained graph edits via a Propose-and-Verify mechanism to greedily reduce the failure mass. On math, code, and QA benchmarks, our CE-Graph achieves higher robustness at a significantly lower cost than strong baselines. This suggests that a system's reliability emerges not from avoiding failures, but from systematically learning and reshaping the geometric structure of its failure distributions.

arXiv:2510.01137v4cs.LG

Re-examining low-rank adaptation for private LLM fine-tuning

Privacy is a central concern when fine-tuning large language models (LLMs) on sensitive data, and differentially private stochastic gradient descent (DP-SGD) provides formal privacy guarantees. When performance at a fixed privacy level largely saturates on established benchmarks such as GLUE, we ask whether the same performance can be reached with less computation. Both theory and practice show that lower-rank models are better suited to DP training, a property especially relevant for LLMs, whose fine-tuning gradients exhibit a strong low-rank structure. Methods such as DP-LoRA exploit this by expressing each layer's weight update as a sum of only a few non-zero rank-one components. However, we argue that while having few non-zero components is important, the isotropic noise injected by DP-SGD inflates the singular values of the gradient matrix, disrupting their naturally fast decay. We investigate whether this noise-induced spectral distortion slows optimization. Experiments on language classification (GLUE with RoBERTa) and table-to-text generation (E2E and DART with Qwen and Llama models up to 4B parameters) show that partially restoring the original singular-value profile significantly accelerates DP-SGD convergence without compromising privacy guarantees. Our code is available at https://github.com/alidadsetan/re-exam-private-low-rank.

arXiv:2510.00922v2cs.AI

On Discovering Algorithms for Adversarial Imitation Learning

Adversarial Imitation Learning (AIL) methods, while effective in settings with limited expert demonstrations, are often considered unstable. These approaches typically decompose into two components: Density Ratio (DR) estimation $\frac{\rho_E}{\rho_{\pi}}$, where a discriminator estimates the relative occupancy of state-action pairs under the policy versus the expert; and Reward Assignment (RA), where this ratio is transformed into a reward signal used to train the policy. While significant research has focused on improving density estimation, the role of reward assignment in influencing training dynamics and final policy performance has been largely overlooked. RA functions in AIL are typically derived from divergence minimization objectives, relying heavily on human design and ingenuity. In this work, we take a different approach: we investigate the discovery of data-driven RA functions, i.e, based directly on the performance of the resulting imitation policy. To this end, we leverage an LLM-guided evolutionary framework that efficiently explores the space of RA functions, yielding \emph{Discovered Adversarial Imitation Learning} (DAIL), the first meta-learnt AIL algorithm. Remarkably, DAIL generalises across unseen environments and policy optimization algorithms, outperforming the current state-of-the-art of \emph{human-designed} baselines. Finally, we analyse why DAIL leads to more stable training, offering novel insights into the role of RA functions in the stability of AIL. Code is publicly available: https://github.com/shshnkreddy/DAIL.