Computer Science (arXiv)

Reliable optimization is central to neural network (NN) training, yet Adam, the default optimizer for modern LLMs, rests on a fragile foundation. This thesis develops a principled grounding for Adam and motivates new designs. First, we revisit Adam's divergence--convergence debate and show the existence of a problem-dependent phase transition: with properly chosen, batch-size-dependent hyperparameters, Adam converges, whereas under small-$β_2$ regimes it can diverge. Second, we investigate why Adam substantially outperforms SGD on Transformers through Hessian structure. We find that the Hessian evolves toward a near-block-diagonal form along training, accompanied by strong block heterogeneity. We prove that this structure makes Adam's diagonal preconditioner effective. We further show that this special Hessian structure originates from consecutive multiplications of large matrix variables, and we provide a rigorous analysis based on random matrix theory. Finally, these insights motivate Adam-mini, a new optimizer that reduces Adam's memory footprint by 50\% while preserving its performance. Our results also have broader implications beyond Adam: they reveal new local structures in matrix-based nonconvex problems, and also help understand and improve recent NN optimizers, such as Muon.

Decades of space-syntax research have established that the topology of the street network conditions movement, co-presence and urban activity. The standard vocabulary for this -- integration, choice and connectivity -- summarises each street's position as a scalar centrality, yet two streets with identical centrality can sit in radically different morphological fabric. We introduce a compact, comparable, machine-learning-ready encoding of that local fabric: the spectral fingerprint, a fixed-dimensional kernel-density representation of the graph-Laplacian eigenvalue distribution of each node's k-hop ego subgraph, computed on the COINS dual graph of the street network. From it we derive two interpretable scalar readouts: the Mesh Index (MI), a normalised spectral entropy, and the Connectivity Resilience Index (CRI), the algebraic connectivity (Fiedler value). Both are corrected for an ego-subgraph-size confound that dominates raw spectral statistics. Applied to the full street network of Poznan, Poland (1,908 continuity-based strokes), the descriptor is robust to its encoding hyperparameters (spectral resolution and kernel bandwidth; Spearman rho >= 0.98) while remaining scale-dependent in its neighbourhood radius. The size adjustment leaves the Mesh Index near-orthogonal to integration (r = 0.06), carrying information classical centrality does not. The fingerprint separates morphological tissue types without supervision, and the Mesh Index is associated with the functional diversity of street-adjacent activity at the neighbourhood scale (r ~ 0.19, on open OpenStreetMap data). We delineate the method's scope honestly: it characterises what kind of activity a street's position affords, not the price that activity commands. It offers an information-theoretic morphological descriptor that complements space syntax and is directly consumable by modern graph-learning pipelines.

RAW video restoration is fundamental to high-quality low-level perception and serves as the basis for a wide range of downstream vision applications. While binary neural networks (BNNs) enable efficient lightweight deployment for image enhancement, their deficiencies in modeling temporal coherence and activation value distributions hinder their effectiveness when applied to video scenarios. In this paper, we propose BinRVR, a binarized RAW video restoration framework that reduces computation and parameters by approximately 96% while incurring only about 4% performance degradation. Specifically, we present a Binarized Information Interaction Module (BIIM) to jointly model spatial and temporal information in an efficient and unified manner. Moreover, we develop a Distribution-Aware Binarized Convolution (DAB-Conv) that leverages the statistics of full-precision activations to mitigate quantization errors. The proposed framework further supports multi-bit quantization, enabling flexible accuracy-efficiency trade-offs across different hardware constraints. Extensive experiments demonstrate that our BinRVR achieves competitive performance compared with state-of-the-art binarized methods on RAW video restoration tasks, including low-light enhancement, denoising, deblurring, and super-resolution. We further explore the potential of our method on downstream video applications, including object detection and monocular depth estimation.

Surgical segmentation networks can fail silently under acquisition degradation: predicted masks may be wrong even when model confidence remains high. Existing deployment-time monitors rely primarily on uncertainty estimates and can therefore miss confident failures. We present TCSR-Monitor (Temporal Conformal Surgical Risk Monitor), a post-hoc failure-monitoring framework that combines confidence with observable shape, temporal-consistency, and image-quality cues. TCSR-Monitor wraps a frozen segmentation model, requires no model internals, and operates without ground truth at deployment. We also introduce a validation protocol to assess whether alarms remain credible under distribution shift. On EndoVis 2017, leave-one-corruption-out evaluation shows that TCSR-Monitor generalizes to unseen acquisition degradations and substantially outperforms confidence-based baselines. A circularity control confirms that it predicts segmentation failure rather than simply detecting corrupted images. Mondrian conformal calibration balances miss-rates across degradation severities, but a single global threshold still produces false alarms on up to 40% of correctly segmented frames at moderate corruption. Zero-shot transfer to SAM2 demonstrates feature portability, although entropy outperforms the transferred monitor at both evaluated thresholds. Overall, reliable monitoring under acquisition degradation benefits from complementary observable signals beyond confidence alone, but substantial false-alarm and transfer limitations remain.

Many areas of AI research, such as language model interpretability and chain of thought faithfulness, seek to explain model behaviors. But what constitutes a "good" explanation? In this work, we evaluate explanations through the lens of counterfactual simulatability-whether the explanation is useful for predicting model behaviors on related counterfactual inputs. To this end, we introduce CHIVE (Counterfactual Hypothesis Investigation Via Edits), a novel agentic pipeline that identifies unexpected model behaviors in the wild and investigates them with counterfactual prompt edits. This yields thousands of high-quality explanations for naturally-occurring model behaviors along with supporting counterfactual evidence. We apply CHIVE in two ways. First, we evaluate whether common LLM interpretability techniques improve an agent's ability to predict counterfactual model behaviors. Surprisingly, we find no uplift from any of the interpretability techniques studied. Second, we use CHIVE to generate training data. We find that training models to predict outcomes of CHIVE-generated counterfactual experiments generalizes to various out-of-distribution settings. Overall, CHIVE automatically discovers explanations of naturally-occurring LLM behaviors, enabling us to evaluate and improve methods for explaining LLM behaviors.

Despite progress in instruction-based video editing, unimodal textual instructions inherently struggle to convey fine-grained textures and complex dynamics. To bridge this perceptual gap, we propose Visual In-context Editing, a new paradigm elevating video editing from textual instructions to multi-modal visual guidance encompassing single image, image pair, and video pair. To facilitate this paradigm, we curate VicEdit-400K, the first large-scale dataset for visual in-context video editing. We develop an automated pipeline to generate 400K high-quality samples across ten task types, ensuring superior visual fidelity and semantic consistency through multi-dimensional filtering. Leveraging this foundation, we introduce VicEdit, a unified framework to bridge visual and textual contexts. To adaptively extract editing semantics from heterogeneous references, we design Modality-Adaptive Semantic Distillation, which produces modality-specific semantic tokens from visual references. These tokens are then synergistically integrated with textual instructions through Dual-Context Injection, enabling the generation process to benefit from both visual and textual signals. Extensive evaluations on VicEditBench demonstrate that VicEdit achieves state-of-the-art performance across both basic instruction editing and visual in-context editing tasks, establishing visual in-context learning as a powerful and controllable paradigm for video editing.

Large Language Models (LLMs) have achieved remarkable progress in code generation, yet ensuring correctness in complex, repository-level tasks remains challenging. Existing approaches often use generated tests as static post-hoc validators, which limits their ability to guide implementation and may introduce misleading feedback when the tests themselves are incomplete or incorrect. In this paper, we introduce TDD-Agent, which operationalizes the test-driven development paradigm for code generation. TDD-Agent first prompts the model to generate executable tests, encouraging it to clarify expected behaviors before implementation, and then performs iterative dual-track refinement over both the generated code and tests using execution feedback. We first isolate the effect of test-first reasoning through a prompt variant TDD-prompt on LiveCodeBench, where it consistently improves upon reasoning-based prompting baselines. Building on this finding, we evaluate the full TDD-Agent framework on RepoEval, a repository-level benchmark, and show that it consistently outperforms retrieval-based and agent-based baselines. Additional analyses show that iterative refinement improves not only code correctness but also the effectiveness of the generated tests, yielding higher pass rates, coverage, and mutation scores, suggesting that tests can serve as evolving reasoning artifacts rather than fixed validators. Our source code is available at https://anonymous.4open.science/r/TDD-Agent-Framework-6370/.

Long-term manipulation planning requires robots to reason not only about immediate task success but also about how current decisions affect future interactions with the environment. In this context, household service robots may need to organize groceries in partially occupied shelves while using limited storage space efficiently and preserving access for subsequent placements. In this paper, we consider an online multi-object shelf-placement setting in which future objects arrivals are unknown. Existing approaches do not jointly address semantic organization, dense space utilization, and manipulator accessibility during sequential shelf filling. To address this gap, we propose Semantic-Dense Placement Planning (SDPP), an accessibility-preserving approach that ranks candidate poses using a semantic-density score combining inter-object semantic similarity with spatial proximity. An Accessibility Map (AM) further filters candidates unlikely to be reachable before motion planning and penalizes placements that reduce the remaining accessible workspace. Simulation experiments show that SDPP significantly improves semantic placement quality over state-of-the-art baselines and achieves the highest average shelf density, while the AM substantially reduces the time required to identify feasible placement poses. A qualitative real-world experiment demonstrates the applicability of our pipeline in a domestic shelf-storage scenario.

Reinforcement learning algorithms for Large Language Models (LLMs) are largely distinguished by their variance reduction strategy. Group-relative methods like GRPO reduce gradient variance by sampling multiple rollouts per prompt, but provide only sequence-level credit. Training is also blocked by straggler rollouts, reducing throughput and increasing off-policyness. Learned value functions theoretically address both problems, providing token-level advantages without requiring large groups. However, additional infrastructure engineering challenges combined with the practical success of critic-free methods have made it difficult to justify their inclusion in RL pipelines. We propose two complementary strategies to improve the performance of value function RL: 1) Privileged Value Functions (PVF) which provide an elegant mechanism to inject additional task-relevant token-level signal without biasing the policy objective; 2) TETHER, a baseline that adaptively interpolates between group-relative and value baselines depending on the value function accuracy. Across several reasoning tasks, both strategies consistently improve over the standard value function baseline, and are competitive with or outperform mean-baseline GRPO.

Liquid democracy permits voters to vote directly or delegate their votes to others. Existing algorithmic analyses assign each voter a single scalar parameter, interpreted as an independent probability of voting for the ground truth. This representation is inadequate when delegation is fixed before public information changes different voters' reliability in different ways. In this work, we study a minimal common-signal model of this phenomenon. Delegation occurs before a public binary signal is realised, while voting occurs afterwards. Conditional on the signal, sink votes are independent, and each voter has a signal-specific competence; marginally, correctness events are correlated through the common signal. We show that delegation based on average competence is not a safe scalarisation: it can violate do-no-harm, and a beneficial delegation rule may send votes to voters with lower average competence. We present and analyse three novel delegation mechanisms for this setting. The first is a conservative intersection mechanism that delegates only to neighbours whose competence exceeds the delegator's by a prescribed margin in every signal state; the conservative intersection mechanism inherits all the guarantees of delegation in the scalar competence setting. The next mechanism is a confounded-set mechanism that enable the delegation to neighbours who are favoured in one state and worse by at most a prescribed tolerance in the other. For this mechanism, we prove expected-margin bounds, and identify the weight-dispersion and mechanism-concentration conditions needed to obtain majority-correctness guarantees. Finally, on bounded-in-degree graphs, a multi-round certified-path mechanism propagates nonnegative two-dimensional path certificates; it is acyclic, yields statewise terminal competence improvement, and gives uniform bounds on path length and terminal voting weight.

Physical design algorithms operate within tightly coupled, multi-stage optimization flows, where stage-local gains may vanish or induce downstream degradation. Existing program-evolution frameworks often rely on stage-local objectives or undifferentiated multi-metric feedback, which neither guarantee better final results nor identify which unmet requirement should guide the next iteration. We present GoalEvolve, a goal-driven framework that makes physical design algorithm evolution accountable for the final quality of results (QoR) of the complete flow. Given a multi-objective QoR target region, GoalEvolve converts unmet requirements into normalized target gaps, identifies the dominant bottleneck, and uses stage-resolved checkpoint evidence to locate the responsible stage. An LLM-based Teacher then narrows the search to a relevant algorithmic decision and source region, while parallel Student agents implement and validate hypotheses through full-flow evaluation. Local effects, optimization debt, and downstream retention are retained as mechanism evidence for subsequent evolution. Across eight ASAP7 designs, GoalEvolve improves post-route TNS by 30.67% on average and reduces leakage and dynamic power by 21.18% and 9.42% versus default OpenROAD. Relative to commercial-tool goals, it closes 62.20% of the normalized power gap on power-dominant designs, surpasses the TNS goals on both timing-dominant designs, and closes 32.48% of the equal-weight timing-power gap on joint designs. Across all three designs evaluated against Codex goal mode under matched budgets, GoalEvolve further improves TNS by 26.46% while reducing leakage and dynamic power by 12.38% and 0.76%, respectively.

Per-pixel rendering difficulty is conventionally measured by the sample variance $\hatσ^2(p)$ of a Monte Carlo estimator, yet this signal is least reliable exactly where difficulty concentrates: under heavy-tailed transport its relative error is governed by the integrand's kurtosis, and the split-half reliability of variance-derived evaluation targets reaches only 0.23-0.29 even at 40,000 samples per pixel. We propose a complementary discrete transport-mechanism descriptor: every contribution event is classified by its end-vertex BSDF lobe, the presence of a delta-specular event, and a single-/multi-bounce distinction, yielding seven mutually exclusive labels whose six named mechanisms receive all observed energy on tested scenes, with continuous side-channels retaining the mechanism mixture. Across seven scenes, the dominant label agrees 87-99.6% between 64 and 4096 samples per pixel -- where quantile-binned variance agrees as little as 21% -- and is robust to restoring the estimator's MIS half. The descriptor exposes cross-scene structure a scalar variance cannot represent, including a geometry-controlled sign reversal of the delta-mediated/glossy correlation. Using the label to correct a noisy pilot variance improves on pilot-variance sample allocation at equal budget on every test-matrix scene with heavy-tailed buckets, while reducing exactly to the incumbent where such buckets are absent, with gains surviving a random-partition placebo and persisting over a robust (median-of-means) pilot baseline. Pre-registered third-party sentinel tests confirm the account out of distribution: coverage and stability transfer, a structural finding survives a blind sign prediction, and on the ajar-door scene, where pilot-variance allocation fails 6.8 dB below uniform sampling, the label identifies from the pilot alone that the failure is not of the kind it repairs, and correctly abstains.

Designing large soft robots capable of generating high forces for physical human-robot interaction remains a significant challenge in soft robotics. Prior work in large soft robots has focused on proof-of-concept prototypes, and no systematic framework exists for determining the suitability of a design paradigm for a desired task. This manuscript introduces a method for optimizing the geometry of a soft robot limb, maximizing its blocking force subject to an anti-bucking constraint under its own gravitational loading. We demonstrate that an explicit solution exists to the proposed optimization problem under certain assumptions. Experiments with three geometries of a large, soft, pneumatically-actuated manipulator demonstrate that the method correctly predicts which designs meet constraints and which produces the largest end-effector forces. This method, with its closed-form solution, can allow designers to determine a-priori if an intended class of soft manipulators is an appropriate choice for physical interaction at large size scales.

Corrupted, inconsistent, or anomalous data silently threatens the safety and reliability of medical AI. Despite growing regulatory recognition of dataset quality assurance (QA) for high-risk medical AI, scalable automated detection remains underdeveloped. We employ unsupervised anomaly detection (AD) and out-of-distribution (OOD) detection as an automated dataset QA mechanism for multi-center dynamic contrast-enhanced breast MRI. We build a controlled AD benchmark of 17 realistic QA-relevant anomaly types from six public datasets (protocol violations, processing errors, incorrect anatomical regions) and propose a taxonomy of radiological image anomalies based on human visual perception, enabling fine-grained analysis of AD failure modes. The benchmark includes near-, medium-far-, far-OOD samples, as well as in-distribution and external normal data. Four methods are evaluated: a projection-based method extended with a domain-specific feature extractor and a novel positional encoding, a reconstruction-based approach extended to full 3D volumes with an augmented training objective, and two unmodified hybrid OOD detection methods. Medium-far- and far-OOD samples are detected reliably, whereas near-OOD samples and external normal data from unseen institutions expose method-specific differences. The 3D reconstruction-based approach best balances detection performance (AUROC: 0.936) and generalization to unseen institutions. The projection-based method with positional encoding achieves the highest overall detection performance (AUROC: 0.954). Both hybrid methods exhibit critical failure modes, confirming that methods validated for one modality or anatomy may not generalize without domain-specific adaptation. Implants and mastectomies remain an open challenge for all methods. Our results establish a foundation and practical guidance on scalable unsupervised QA in medical AI pipelines.

The rapid evolution of text-to-image (T2I) generation models has effectively solved the foundational challenge of raw pixel synthesis, shifting the community's focus toward fulfilling increasingly intricate user requests. While recent agentic image generation workflows enhance static inference with advanced capabilities like external knowledge retrieval and iterative reasoning, they mostly operate in isolated silos with fixed ``one-size-fits-all" topologies. This inevitably leads to severe compute-mismatch, where simple queries are forced through computationally heavy pipelines. To bridge this gap, we present GenRouter, the first unified workflow routing framework for agentic image generation. We first formulate GenCanvas, standardizing diverse agentic pipelines into a universal set of foundational primitives and executable templates. Operating over this unified space, GenRouter adaptively routes heterogeneous prompts to their optimal workflows via (i) demand profiling, (ii) experience matching, and (iii) Pareto filtering. Extensive experiments across diverse benchmarks demonstrate that GenRouter achieves superior visual alignment while reducing execution costs by over 95% and latency by 65% compared to heavyweight static pipelines. Furthermore, the system continuously self-evolves via accumulated experience, enabling robust zero-shot generalization that boosts performance and halves computational overhead.

Identifying cell types directly from routine haematoxylin and eosin (H&E) histology would enable single-cell analysis at scale, but training such models has relied on manual pathologist annotations, which are slow, expensive and unreliable for many cell types. We instead supervise morphology with molecules. Imaging-based spatial transcriptomics profiles individual cells in situ on a section that can afterwards be stained with H&E, so that molecular identity and morphology are observed for the same physical cell. We assembled 81 such paired Xenium sections spanning 16 organs, derived per-cell labels by clustering, marker-gene annotation, organ-wise human review and quality control, and mapped them onto the cell types commonly reported in each organ. This yielded 15.4 million cells, each with a paired H&E image patch and one of 23 cell types, on which we trained CytoFormer, a cell foundation model with a multi-task, per-organ classification head. On spatially held-out tissue CytoFormer reached an accuracy of 0.85 and a macro-F1 of 0.78 across all 16 organs, and its predictions reproduced the tissue architecture of an entire held-out section. The representation also transfers: with the encoder frozen, a linear head on CytoFormer features performed better than six pathology foundation models on four expert-annotated benchmarks, including on organs and cell types that were not part of pretraining. Finally, in an interactive active-learning setting, CytoFormer's embeddings are markedly more label-efficient than existing pathology foundation models, detecting normal epithelium amid look-alike tumour with an F1 of 0.82 from only a few annotations and leading the strongest baseline by 0.13 in F1. CytoFormer turns paired H&E and spatial transcriptomics into a reusable, label-efficient representation for cell-level analysis of routine histology.

Video generation is rapidly evolving from single-shot clips to multi-shot narratives, where the human character serves as the core narrative anchor. However, existing benchmarks mainly assess character appearance or individual-shot quality, without measuring whether physical and emotional states remain coherent across cuts. They also rarely provide criterion-specific evaluation methods, although physical continuity, facial dynamics, and cinematic relations require different visual, temporal, and relational evidence. To address these limitations, we introduce PersonaShot, the first person-centric benchmark for narrative continuity in multi-shot video generation. PersonaShot contains approximately 1,000 multi-shot segments and 16 metrics spanning physical continuity, affective dynamics, and cinematic grammar. \textbf{\textit{1)} Narrative Continuity Benchmark:} We evaluate character coherence across three temporal levels: within-shot states, cross-shot transitions, and sequence-level trajectories. \textbf{\textit{2)} Human-Aligned Specialist Evaluators:} We distill reasoning from a large multimodal teacher into lightweight criterion-specific evaluators, each grounded in the visual, temporal, or relational evidence required by its metric, and align them with expert human judgments. \textbf{\textit{3)} Systematic Evaluation and Insights:} Our evaluation reveals distinct capability profiles across state-of-the-art models and a clear gap between perceptual quality and cross-shot narrative continuity. Even visually compelling videos frequently exhibit physical-state resets, abrupt affective shifts, and broken cinematic relations across shots. Human studies further demonstrate strong agreement between our evaluators and expert judgments.

In-context imitation learning enables few-shot policy generalization but struggles to maintain performance on unseen objects and novel scenarios. To address this, we introduce MatchingPolicy, a correspondence-driven framework that explicitly decouples demonstration-to-scene matching from policy learning. Central to our method is a correspondence-aware diffusion policy that conditions robotic actions directly on dense semantic correspondences. This architectural separation resolves the inherent conflict between correspondence identification and action adaptation, enabling robust out-of-distribution transfer. Our framework integrates vision foundation models with a novel two-stage matching algorithm to dynamically establish reliable correspondences. Extensive evaluations on RLBench and real-world manipulation tasks confirm that MatchingPolicy achieves superior few-shot performance, generalizing reliably across unseen object instances and semantic categories.

Underactuated tendon-driven hands offer compact actuation and passive compliance, but tendon elongation under restoring-spring loading introduces configuration-dependent joint deviations. This paper presents H-PAC, a modular 6-actuator, 15-DoF robotic hand with a control-oriented modeling and implementation framework. A sparse analytical actuator-joint model is derived from the tendon-routing geometry, and a mechanics-based compensation model is developed to account for tendon-elasticity-induced joint errors. The proposed method is implemented in a hierarchical architecture: a host computer performs workspace-constrained posture mapping and compensation, while an ESP32 generates synchronized commands for six position-controlled servos. The same control parameters and execution strategy are used across all tasks without task-specific retuning. Monotonic servo-sweep experiments show that the compensation substantially improves joint-angle prediction. The MAE of the index DIP joint decreases from 1.15 degrees to 0.18 degrees, and all nine evaluated joints achieve an MAE below 0.23 degrees. Representative postures and grasping configurations are further executed using the same control pipeline without external joint or force sensing in the control loop. The results demonstrate a practical approach to improving posture reproducibility in compact underactuated robotic end-effectors.

As autonomous vehicles (AVs) approach Level 4 and Level 5 operational capability [SAE International, 2018], their on- board decision systems must handle not only safety-critical locomotion but also their subsequent moral weight. This paper details the Ethical Decision Head (EDH), a deep re- inforcement learning (RL) framework that encodes ethical reasoning as a differentiable reward signal, enabling a pol- icy gradient agent to learn morally-aligned driving behavior in scenarios whose state representation is aligned with the CARLA simulation environment [Dosovitskiy et al., 2017]. Two normative frameworks are instantiated and evaluated: a Utilitarian framework minimizing total casualties and a Kan- tian framework enforcing course maintenance as a categori- cal imperative. The EDH is trained via Proximal Policy Op- timization (PPO) [Schulman et al., 2017] against a Bradley- Terry reward model [Bradley and Terry, 1952] learned from pairwise human preference annotations over 200 collision- imminent scenarios. Results reveal an asymmetry in the learnability of normative ethical frameworks under human su- pervision. The Kantian condition, which reduces to a con- stant prediction task under the codebook, serves as a pipeline control: it confirms training stability and rules out infrastruc- ture failure as an explanation for the utilitarian result. The Utilitarian agent learned something more unsettling: human raters rewarded self-sacrifice over casualty minimization, and the model learned that preference faithfully. This divergence between what humans prescribe in theory and what they re- ward in practice suggests that RLHF does not learn ethics as philosophers define it, but as humans live it.

A radiologist reading a model's output faces two problems. The model returns a number and no reason, and any system that turns that number into readable prose can quietly add claims the model never made. MIRROR is a research prototype built to separate those failures. It chains a multi-label classifier, a Grad-CAM localizer that turns each positive finding into a named anatomical region, and a report writer that receives the labels, probabilities, and regions but never the image. Because the language layer cannot see pixels, it cannot assert a finding the classifier did not make. We are precise about what that buys: a MIRROR report's findings are auditable against the probability vector, while the sentences framing them are ordinary generated text, and we show one stating a cardiothoracic ratio the system never measured. One registry holds the taxonomy, anatomy, and phrasing for chest X-ray, brain MRI, and head CT, so adding a modality is a data change; all three are routed and tested, one is trained. On ChestMNIST that classifier reaches macro AUROC 0.729 and ranks better than chance on all 14 labels, at 1.6 to 6.8 times the precision a random ranker would get. Yet at the default 0.5 threshold it emits no positive prediction at all for 11 of them, and its excellent-looking Brier score of 0.045 sits beside the 0.047 earned by a predictor that ignores the image. The discrimination is real; the decisions are not. Under the class imbalance normal in radiology, aggregate metrics flatter models that do nothing, and should be reported against that floor.

Large language models (LLMs) are increasingly deployed as decision-making agents in settings that require sophisticated environmental exploration. However, existing work has raised questions about how LLMs actually balance exploration and exploitation. Unlike classical agents, LLM agents engage with tasks through natural language, exposing them to semantic information with no formal counterpart in the task structure. We introduce the semantic bandit, an extension of the multi-armed bandit setting that explicitly considers the textual labels assigned to actions, and use it to study how semantic priors --- inductive biases arising from associations between language and expected reward learned during pre-training, shape LLM exploration behaviour. We find that semantically informative action labels reduce exploration in favour of exploitation, improving performance when aligned with the reward structure and severely degrading it when misaligned. We further find that negative rewards trigger substantially more exploration than equivalent positive rewards, consistent with an expected-scale bias induced by reward conventions common in pre-training data. Overall, we argue that the use of language to define the environment and rewards introduces unavoidable biases derived from the fact that the model is trained on word co-occurence, with implications for the reliability and robustness of LLM agents in real-world decision-making settings.

Recommender systems typically rely on implicit feedback (e.g., clicks) to infer user preferences. However, such data is inherently prone to various biases, including position bias and popularity bias. Position bias occurs when higher-ranked items receive more interactions regardless of true relevance. Popularity bias reinforces frequent exposure of popular items while under-recommending relevant, yet less popular ones. Directly learning from such data fails to capture true user preferences, leading to suboptimal recommendations. This research focuses on mitigating position bias and popularity bias in recommender systems. Specifically, I address position bias in learning-to-rank (LTR) systems and popularity bias in collaborative filtering (CF) models and social recommender systems based on graph neural networks. My work develops methods that overcome the limitations of existing approaches to mitigating position bias and popularity bias, enabling more relevant and personalized recommendations that align with users' preferences.

Various machine unlearning techniques have been developed in response to privacy legislation requirements, enabling individuals to exercise their legal right to have their data $D_f$ removed from a machine learning model. This process is typically accomplished via the use of an unlearning function denoted as $U$. Existing methods focus on designing an intricate $U$ to unlearn $D_f \subset D$ from a previous model $A(D)$, so that the unlearned model performs as closely as possible to the retrained model $A(D \setminus D_f)$. However, these methods often suffer from high computational costs when dealing with massive training data, as the complex structures of $U$ become a bottleneck even for models with fewer parameters. Inspired by Learning to Optimize, we introduce the first learning-based model-agnostic approach, Learning-to-UnLearn (L2UL). Our core insight is to shift from manually designing $U$ to learning the unlearning behaviors from a distribution perspective, thereby acquiring a simple and efficient $U$ via learning. Our experimental results demonstrate that the accuracy achieved by L2UL is comparable to that of retraining while exhibiting impressive efficiency, particularly in data-intensive scenarios. Furthermore, we validate the performance and scalability of our method on larger models ResNet.

Motivated by modern marketplaces, where the platform or the seller routinely gathers detailed user profiles, we study a novel learning theoretic model that simultaneously involves information and mechanism design. Specifically, we consider the economic setting recently introduced by Bergemann et al. (2022), where in addition to the menu of quality-price pairs, the seller offers information on the value of the match between product quality and buyer's taste via a signaling scheme. We relax the assumption that the seller knows the buyers' belief about the distribution of tastes and study the sample requirements of designing a revenue maximizing scheme. We consider both the batch setting where we have access to data from a set of i.i.d. buyers and an online demand query model where we observe the buyers' behaviors to seller's schemes. Despite the apparent non-convexity of the problem, we also give the first FPTAS to compute a scheme that maximizes the revenue within an arbitrarily small additive loss, which was left open by Bergemann et al. (2022). Overall, this brings a new learning perspective in asymmetric economic settings where buyers and sellers know different types of information.