Differentially private stochastic gradient descent (DP-SGD) has become the standard framework for privacy-preserving machine learning, yet its reliance on a fixed gradient clipping threshold to limit sensitivity remains a significant practical limitation. Adaptive clipping algorithms such as AdaClip shift and scale the gradient prior to clipping and adding noise so that the clipped gradient yields a more informative descent direction. The shift and scaling parameters are selected adaptively based on the empirical mean and variance. However, in existing adaptive clipping algorithms, these empirical estimates have not been also used for momentum to accelerate training itself. On the other hand, DP-Adam is an algorithm that exploits Adam-like momentum updates based on the gradient mean and variance to accelerate training, but does not exploit these estimates for adaptive clipping. In this work, we propose Differentially Private Mechanism with Adaptive Clipping and Adaptive Momentum (DP-MacAdam), a novel algorithm that combines these two approaches so as to use the same mean and variance estimates for both clipping and momentum. We perform an analysis showing that DP-MacAdam estimates the gradient variances in a bias-free manner. In addition, we empirically evaluate the privacy and accuracy of DP-MacAdam, demonstrating that it achieves improved model utility compared to DP-SGD, AdaClip, and DP-Adam baselines, without requiring manual tuning of the clipping threshold.

Group Relative Policy Optimisation (GRPO) has emerged as an effective reinforcement-learning algorithm for aligning language models on reasoning tasks, but it treats every token position and every sampled rollout symmetrically. We introduce two complementary extensions: (i) Adaptive-Horizon GRPO (AH-GRPO), which weights each token's policy gradient using a cumulative entropy-based discount that reduces the effective horizon when the model is uncertain, and (ii) Selective-Advantage AH-GRPO (SA-AH-GRPO), which applies this discounting only to negative-advantage rollouts, leaving positive-advantage, successful trajectories unattenuated. We evaluate standard GRPO with alpha = 0, AH-GRPO with alpha = 0.5, and SA-AH-GRPO with alpha = 0.5 on the GSM8K mathematical reasoning benchmark using both Qwen 2.5-1.5B-Instruct and Qwen 2.5-3B-Instruct fine-tuned with LoRA. On the 3B model, SA-AH-GRPO achieves Pass@1 = 0.858 at its peak at step 30 and maintains 0.846 at 180 steps, with training variance reduced to 0.0246, a 3.6 times reduction relative to GRPO while matching its peak accuracy. On the 1.5B model, SA-AH-GRPO achieves a peak Pass@1 of 0.686, improving over the zero-shot baseline of 0.637. Our analysis shows that asymmetric discounting preserves the full gradient signal on correct solutions, prevents entropy collapse, and substantially stabilises training, suggesting a principled inductive bias for reinforcement learning with verifiable rewards on structured generation tasks.

Frontier AI governance frameworks increasingly use cumulative training compute as the primary criterion for designating high-impact models, but enforcement rests on self-reporting because no technical verification primitive for training exists. Any future international agreement on frontier AI faces the same problem at higher stakes: coordinated regulation of technologies with significant externalities has historically rested on technical verification, without which agreements are declaratory. Recent governance analyses judge zero-knowledge proofs a promising candidate but currently impractical at frontier scale [26, 4]. We argue the impracticality is paradigm-bound rather than fundamental, and propose a verification architecture for frontier dense pre-training combining a pre-committed training specification, inter-node network observations, and on-the-fly Merkle commitments of intermediate computation, verified through a zero-knowledge Virtual Machine (zkVM) with native BF16/FP32 precompiles. The proof checks the actual floating-point computation the GPU performed rather than a fixed-point approximation, and preserves model-architecture confidentiality through a private training specification. The protocol produces three proof types: a genesis proof at initialisation, in-training step proofs across the run, and ex-ante attestations enforcing policy-relevant claims as running invariants, turning the training record into a governance-enforceable artefact. We estimate a deployable proof of concept within approximately 36 months at single-digit-percent training-side overhead, against a six-to-ten-year cycle for verification-grade custom silicon. Thirteen open research and engineering problems are catalogued as a research agenda for external contribution

Post-training quantization (PTQ) is critical for the efficient deployment of large language models (LLMs). Recent ultra-low-bit PTQ methods rely on rigid weight-saliency assumptions or position heuristics, introducing substantial hidden scaling overhead. We propose SAGE-PTQ (Saliency-Aware Graph-guided Efficient PTQ), a novel ultra-low-bit quantization framework for LLMs that minimizes hidden scaling cost. SAGE-PTQ separates salient and unsalient weights using distributional statistics, then models subsampled unsalient weights as a sparse graph to estimate the optimal number of groups per layer. SAGE-PTQ applies dual-mode quantization, assigning multi-bit precision to salient weights and binarizing unsalient weights. To reduce scaling overhead, SAGE-PTQ uses one per-channel scale for salient weights and one scalar per unsalient group. Finally, SAGE-PTQ implements adaptive saliency thresholding to select the optimal saliency ratio per matrix. SAGE-PTQ achieves 1.03 weight bits and only 0.004 scaling bits per matrix on average, outperforming state-of-the-art methods such as BiLLM and PB-LLM. On LLaMA-3-8B, SAGE-PTQ achieves 6.74 WikiText2 perplexity, compared to 55.8 for BiLLM, while using less than 50% of BiLLM's GPU memory. On LLaMA-2-70B, SAGE-PTQ provides 1.5x faster decoding on one NVIDIA L40 GPU, demonstrating practical inference efficiency.

Learning from demonstratins (LfD) is usually performed over Euclidean spaces, while the robot state, e.g. orientation, naturally evolves over curved spaces. Therefore, to ensure natural, complex motion generation, we investigate learning from demonstrations over Riemannian manifolds that are capable of encoding both position and orientation data. Here, geodesic paths provide for natural motion between two arbitrary points within the manifold. We propose to numerically estimate geodesics via neural ordinary differential equations, mitigating large computational overhead of existing approaches. Finally, these geodesics can be decoded back into the original task space before deploying on the robot. In this extended abstract, we discuss the architecture of our framework, provide some initial insights from our simulation experiments, including comparison to other geodesic computation mechanisms, and discuss the challenges and prospects for future work.

When a text is translated, does the translation retain the complexity of the original? We introduce ComplexityMT, a new challenge for assessing how text complexity and machine translation interact with and influence each other, using the Common European Framework of Reference for Languages (CEFR) levels as the measure of text complexity. Across six languages, including Arabic, Dutch, English, French, Hindi, and Russian, we evaluate three open-weight models, one closed model, and a commercial machine translation system on two tasks: i) correlation of CEFR with translation difficulty, and ii) shifts in CEFR levels of the source texts. Our experiments show that higher CEFR levels make texts more difficult to translate, and that machine translation shifts the CEFR level of the target text compared to the original source, for most languages. These findings provide new insights for researchers and practitioners working on multilingual pedagogical content generation and machine translation difficulty estimation.

The rapid proliferation of hyperscale data centers (HDCs) in the US, mainly driven by the adoption of artificial intelligence, has raised concerns about this industry's environmental footprint. We compiled facility-level information on 403 US hyperscale data centers operating between May 2024 and April 2025 and estimated their electricity consumption, electricity sources, and attributable CO2 emissions. Across different facility-load scenarios, these HDCs consumed approximately 68-99 TWh of electricity and were associated with about 37-54 million metric tons of CO2. Under the central scenario, HDC electricity demand corresponded to approximately 1.8% of total US electricity consumption, with roughly 54% of attributed generation supplied by fossil-fuel sources. The HDC electricity-weighted average carbon intensity was approximately 545 gCO2/kWh, about 48% above the contemporaneous US national grid-average carbon intensity of 370 gCO2/kWh. Our approach provides an attributional tool for assessing the environmental footprint of hyperscale data centers using the most recent EPA eGRID plant-level data.

Real-world data spans tables, documents, and semi-structured files with implicit semantics. Querying this data requires integrating evidence across inconsistent schemas and formats, yet existing approaches either demand costly manual engineering or bypass structure entirely. We present a system that automatically discovers an executable schema from raw multi-source data and uses it as a shared contract for knowledge graph construction and query-time retrieval. A closed-world field catalog constrains LLM-based schema discovery to attested fields; deterministic structural analysis infers identity keys, foreign keys, and source hierarchy; and the resulting schema drives extraction, deduplication, and cross-source linking into a provenance-aware knowledge graph. At query time the schema -- optionally extended via a monotonic protocol -- conditions a multi-tool agent routing retrieval across structured lookup, graph traversal, and vector search, returning grounded answers with traceable citations. In controlled zero-shot comparisons using the same LLM, data, and evaluation harness, the system improves over retrieval-only and decomposition-based baselines across four QA benchmarks, with ablations showing that schema-conditioned routing, structural intelligence, and schema-guided construction each contribute to the gains.

Early failure alerting requires deciding, while a dialog or agent trajectory is still unfolding, whether to flag it as likely to fail. This is challenging because supervision is typically available only as a trajectory-level success/failure label while alerts must be raised from partial interactions. Prior early-classification methods often bridge this gap by assigning the terminal label to every prefix, treating every turn as failure evidence. We hypothesize that this prefix-label assumption is poorly matched to multi-turn language interactions, where evidence of eventual failure is sparse and often delayed. In this paper, we introduce a two-stage approach that learns from this sparse evidence structure and uses the resulting risk estimates for controllable early alerting. Specifically, our attention-based failure predictor learns sparse turn-level failure evidence from trajectory labels and uses it to estimate failure risk from partial histories. We then pair this predictor with $α$-STOP, a single preference-conditioned stopping policy that selects an accuracy-earliness operating point at inference time rather than training a separate trigger for each preference. Across five benchmarks spanning customer support, task-oriented dialog, persuasion, tool use, and planning, we first show that high-relevance failure evidence occupies only 4.7-11.3% of turns and first appears after 59.0-83.6\% of trajectories on average. We further show that the attention-based predictor improves Pareto-frontier quality (hypervolume) by 1-10\% over naive prefix supervision, and that the full system improves frontier quality by 3-42\% over state-of-the-art trigger policies while reducing training cost per operating point by 1-3 orders of magnitude.

As urban environments continue to evolve rapidly, accurately modeling the dynamic behaviour of Points of Interest is essential for supporting data-driven urban planning and commercial decision-making. While recent advancements in spatio-temporal graph learning have improved POI forecasting, most methods rely on proximity-based graphs and correlation-driven modeling, which overlook the functional dependencies between POIs and fail to capture the causal effects of urban interventions. In this paper, we introduce a novel research problem -- cold-start POI check-in forecasting, which aims to predict the future check-in pattern of a newly introduced POI, by modeling its temporal evolution and functional interactions with nearby POIs in a structured urban spatial context. To address these challenges, we propose CausalPOI, a spatio-temporal graph-based causal representation learning framework. CausalPOI leverages Spatio-Temporal Functional Interaction Graph to model semantic and spatial relationships between POIs, and constructs structurally aligned treatment and control graphs to simulate factual and counterfactual scenarios. Extensive experiments on real-world SafeGraph datasets demonstrate that CausalPOI significantly outperforms state-of-the-art baselines across the board, validating its effectiveness in spatio-temporal forecasting, semantic interaction modeling, and causal effect estimation, providing a more interpretable and actionable foundation for urban intervention analysis. Source code is available at Github.

Motivational architectures in cognitive AI have largely been designed for physical agents regulating bodily needs. Conversational agents operate in a different regime: their sensorimotor loop is linguistic, their environment is a user's evolving mental state, and their consequential actions are speech acts, tool invocations, and strategic silences. This paper proposes a conversational reinterpretation of the OpenPsi motivational lineage, coupled to MetaMo's higher-level motivational scaffold, for agents built on a modular execution substrate. Homeostasis is recast in dialogue-native terms: the agent regulates competence, uncertainty reduction, affiliation, affinity, legitimacy, nurturing, and aesthetic coherence rather than bodily deficits. We propose three contributions: a ten-stage motivational processing pipeline that architecturally separates cognitive modulation from situational appraisal; a dual decision strategy blending urgency-driven fast response with deliberative multi-goal optimization; and an architecturally useful distinction between pre-action feelings and post-action emotions as functionally different forms of affect. We specialize the framework to two example agents -- CompanionAgent and ResearchAgent -- and sketch its extension to social robotics and domain-generic human-level AGI.

When an LLM repeatedly mutates a program, does it explore new forms or circle back to the same ones? We study this question by analyzing LLM-driven mutation chains in the absence of selection pressure within a domain-specific language, varying prompt design, model family, and stochastic replication. We find that LLM-based mutation consistently converges toward restricted attractor regions in program space. Convergence is especially severe at the structural level: in 87% of chains, over 93% of mutations revisit a previously seen structural form, with most variation confined to terminal substitutions within recurring templates. Cycle analysis reveals short cycles and self-loops dominating the transition structure. The rate of convergence varies with prompt wording and model choice, but the phenomenon is robust across conditions. A classical GP subtree mutation operator does not exhibit comparable convergence, suggesting that the effect is intrinsic to the LLM mutation pipeline. These findings reveal a tension at the heart of LLM-driven program evolution: the same capabilities that enable semantics-aware program transformation also carry a systematic bias toward structural homogeneity that must be accounted for if such systems are to sustain open-ended exploration. Source code is available at https://github.com/can-gurkan/lmca.

This work addresses sequentially grasping multiple objects with a single dexterous hand without releasing those already held. Most dexterous grasping methods commit all of the hand's degrees of freedom to a single object, underutilizing its dexterity and leaving no redundancy for subsequent grasps. The proposed solution, MoDex, is a diffusion policy that predicts the next gripper pose directly from observations, conditioned on an opposition space and point cloud. The opposition space condition specifies which fingers participate in the current grasp, enabling the gripper to use only a subset of its available degrees of freedom while reserving the remaining degrees of freedom for subsequent grasps. To facilitate sim-to-real transfer, MoDex is trained in two stages: first through imitation learning on expert demonstrations, and subsequently through reinforcement learning fine-tuning, which consistently improves success rates over the pre-trained policy. We evaluate MoDex in simulation on a MuJoCo-based Franka Emika Panda robot equipped with an Allegro Hand and on the corresponding real-world hardware platform. Across both simulation and real-world experiments, MoDex achieves higher success rates than the evaluated learning-based baselines, improving performance by 2.92-17.92% and 6.67-17.78%, respectively. Project page: https://modex2026.github.io/.

Time series are often embedded in rich contexts that are essential for holistic modeling. Moreover, real-world practitioners often require end-to-end workflows for analyzing temporal dynamics, where widely studied tasks such as forecasting are only one step in a broader solution loop. While generalist AI agents offer a promising interface for such workflows under complex contexts, they still operate primarily in textual spaces that are not fully aligned with structured temporal signals. In this work, we introduce TimeClaw, an agentic harness framework for time series that equips generalist LLM agents with the time series-native runtime support needed for contextualized temporal reasoning. TimeClaw integrates executable temporal tools for grounded and auditable analysis, experience-driven capability evolution for creating reusable analytical routines, and episodic multimodal memory for retrieving relevant reasoning traces. Together, these components unlock harnessed open-ended temporal reasoning with contextual information. Extensive evaluation on multiple benchmarks covering diverse tasks across energy, finance, weather, traffic, and other real-world domains demonstrates improved performance of TimeClaw. Code is available at https://github.com/iDEA-iSAIL-Lab-UIUC/TimeClaw.

Language models increasingly act as epistemic proxies, synthesizing evidence from multiple sources to inform decisions. Whether they evaluate the quality of that evidence, or merely aggregate it based on surface presentation, remains poorly understood. We show that models possess the capability to detect fabricated statistics (correct identification rates of 0.76-1.00 for methodology in isolation) but do not recruit this capability during multi-source synthesis, producing similar numeric estimates whether the statistics are fabricated or valid. Specifically, source influence is governed by a methodology-register gate that responds to the distributional register of analytical text but not to numeric validity: for example, statistically impossible confidence intervals receive the same weight as valid ones. The behavioral dissociation replicates across five models from three families (Claude, Qwen, OLMo) and three professional domains. Mechanistic analyses, including causal tracing, linear probes, and component-level attribution, converge on the same account: the model encodes and causally uses a methodology-register representation that transfers across domains (probe AUC 0.83-0.92), while numeric-validity signals, decodable in isolation, are suppressed to chance during multi-source synthesis. Prompting-based mitigations, even an oracle checklist naming the exact statistical checks, produce blanket skepticism rather than selective discernment, and the post-training pipelines we examine reinforce the stylistic shortcut without building numeric verification. Unlike sycophancy, which tracks user preference, this failure tracks whether a source presents as analytically credible, not whether its claims are internally consistent. We term this epistemic alignment: like preference and safety alignment, the question is not capability but deployment.

Large reasoning models (LRMs) produce reasoning traces with non-linear structures, such as backtracking and self-correction, that complicate the evaluation and monitoring of the reasoning process. We introduce ReasoningFlow, a framework that captures the discourse structures of LRM reasoning traces into fine-grained directed acyclic graphs (DAGs). We develop and validate our annotation schema through careful manual annotation of 31 traces (2.1k steps), achieving high inter-annotator agreement, then scale to automatic annotation of 1,260 traces (247.7k steps) spanning three tasks (math, science, argumentation) and five models (Qwen2.5-32B-Inst, QwQ-32B, DeepSeek-V3, DeepSeek-R1, GPT-oss-120B). By analyzing ReasoningFlow graphs, we find: (1) LRMs exhibit structurally similar traces, despite being trained from different base models and potentially non-overlapping post-training data. (2) ReasoningFlow reveals diverse fine-grained reasoning behaviors (e.g., local verification, self-reflection, and assumptions) that can be used for better reasoning trace monitorability. (3) In LRMs, most of the erroneous steps are not used to derive final answers. (4) Mechanistic causal dependencies between steps do not reflect the language-level discourse structure. We release the dataset and code in: https://github.com/jinulee-v/reasoningflow.

Long-horizon autoformalization of research mathematics fails not only at hard lemmas, but at scale: statements drift, dependencies tangle, context decays, and local repairs corrupt distant work. We present LeanMarathon, a multi-agent harness for reliable research-level Lean autoformalization. Its core abstraction is an evolving blueprint: a Lean file that serves simultaneously as formal proof skeleton, natural-language proof graph, and shared system of record. Four contract-scoped agents construct, audit, prove, and repair this blueprint. These agents are coordinated by a two-stage orchestrator that first stabilizes target fidelity through adversarial review and then discharges the proof directed acyclic graph (DAG) from its dynamic leaves upward in parallel CI-gated rounds. LeanMarathon turns one brittle multi-hour run into many local, recoverable, parallel transactions. We evaluate LeanMarathon on two recent research papers spanning four Erdős problems (#1051, #1196, #164, #1217). Across three autonomous runs, it formalizes all seven target theorems with no sorry, proving 258 lemmas and theorems. These results show that reliable AI co-mathematics requires not only stronger provers, but durable harnesses that preserve target fidelity across long mathematical developments. The code can be found at https://github.com/YuanheZ/LeanMarathon.

Process reward models (PRMs) rely on high-quality process supervision data, yet existing construction methods often provide limited control over error location, error type, and trajectory consistency. We propose a controllable and verifiable framework for synthesizing process supervision data for PRMs. Our framework first constructs a correct symbolic reasoning chain, injects a template-aware error into an intermediate step, recomputes subsequent steps under the corrupted state, and verifies that the injected step is not derivable from its prefix. The resulting paired trajectories are prefix-invalid at the first error while remaining trajectory-consistent after symbolic recomputation, and are translated into aligned natural-language processes for PRM training and evaluation. Experiments show that the synthesized data improve Best-of-8 reranking on logical reasoning benchmarks and transfer to mathematical reasoning. Step-level evaluation further shows that first-error localization remains substantially more challenging than overall step classification, highlighting the need for fine-grained and verifiable process supervision.

Should a single collision necessarily terminate an entire navigation episode? In most deep reinforcement learning (DRL) frameworks for robot navigation, this remains the standard practice: every collision immediately triggers a global environment reset and is penalized as a complete task failure. While a collision during deployment naturally indicates task failure, applying the same treatment during training prevents the agent from exploring challenging obstacle configurations, which slows learning progress in the early training phase. In this work, we challenge this convention and propose a Multi-Collision reset Budget (MCB) framework that decouples local collision termination from global environment resets, allowing the agent to retry difficult configurations within the same episode. Simulation experiments show that MCB improves early-stage learning efficiency by reaching target success-rate levels with fewer interactions, with small collision budgets producing the most consistent gains. Real-world experiments on heterogeneous robot platforms further validate the deployability of the learned policies in cluttered environments.

Steering is a widely used technique for controlling large language models, yet its effects are often unstable and hard to predict. Existing theoretical accounts are largely based on the Linear Representation Hypothesis (LRH). While LRH assumes that concepts can be orthogonalized for lossless control, this idealized mapping fails in real representations and cannot account for the observed unpredictability of steering. By relaxing LRH's orthogonality assumption while preserving linear representations, we show that overlapping concept contributions naturally yield a sample-specific axis-orthogonal structure. We formalize this as the Cylindrical Representation Hypothesis (CRH). In CRH, a central axis captures the main difference between concept absence and presence and drives concept generation. A surrounding normal plane controls steering sensitivity by determining how easily the axis can activate the target concept. Within this plane, only specific sensitive sectors strongly facilitate concept activation, while other sectors can suppress or delay it. While the surrounding normal plane can be reliably identified from difference vectors, the sensitive sector cannot, introducing intrinsic uncertainty at the sector level. This uncertainty provides a principled explanation for why steering outcomes often fluctuate even when using well-aligned directions. Our experiments verify the existence of the cylindrical structure and demonstrate that CRH provides a valid and practical way to interpret model steering behavior in real settings: https://github.com/mbzuai-nlp/CRH.

Reusable robot skills are becoming the basic units through which embodied agents turn open-ended instructions into long-horizon physical behavior. We argue that, while foundation models have collapsed the cost of creating these skills, the cost of trusting them has not. Existing skill-evolution loops refine skills through execution feedback, unit tests, environment reward, or LLM self-critique, but these signals provide only trace-level evidence: they show that a skill worked on sampled executions, not that skill-induced plans satisfy temporal safety contracts under untested conditions. We introduce VASO, a framework for verification-guided self-evolution of LLM-generated robot skill contracts. In VASO, each skill is represented as a semantic contract with two coupled interfaces: a formal interface that aligns robot states, observations, and control commands with logical propositions for model checking, and a planner-facing interface that guides executable behavior generation. A model checker first filters logically inconsistent skill contracts, then verifies plans induced by the skill against global and local temporal specifications. When verification fails, VASO translates the counterexample trace into a textual gradient that updates the reusable skill contract while keeping foundation-model weights frozen. On Clearpath Jackal and PX4 quadcopter tasks, VASO reaches 97.2% formal-specification compliance using fewer than 100 optimization samples, outperforming execution-feedback, prompt-optimization, and fine-tuning baselines. To our knowledge, VASO is the first framework that closes the loop between formal verification and self-evolving LLM-generated skills for physical AI agents: formal counterexamples become optimization feedback for reusable robot skill contracts, rather than merely verifying one-off plans, tuning planner prompts, or fine-tuning model weights.

Lossy compression is essential for massive spatiotemporal data from scientific simulations. Learned compressors can achieve high compression ratios at moderate accuracy targets, but their aggregate reconstruction losses do not guarantee accuracy for each block. Existing Guaranteed Autoencoder (GAE) methods add a per-block residual correction by retaining SVD/PCA-style coefficients until the target is met. This works at moderate tolerances, but in the high-fidelity regime with block-level NRMSE from 10^-6 to 10^-4, the number of retained coefficients grows quickly and the correction stream dominates the total rate. We propose a residual-centric view: the learned residual is structurally different from the original scientific field and should be coded with a representation designed for that residual. We introduce two residual coders. LBRC is a deterministic, training-free pipeline that adaptively quantizes the learned residual to the target NRMSE and losslessly encodes the resulting integer residual using 3D Lorenzo differencing, zigzag mapping, bit-plane coding, and entropy coding. NGLR adds a causal neural predictor that outputs a normalized bias for an integer-rounded Lorenzo prediction in the same deterministic integer pipeline, reducing the entropy of the remaining residual code while preserving deterministic decoding. The predictor weights are serialized and counted in the bitstream. Across E3SM, JHTDB, and ERA5 at block-level NRMSE targets from 10^-6 to 10^-4, LBRC improves compression ratio over GAE by 30-60% and is broadly competitive with SZ. NGLR adds a further 10-40% over LBRC and outperforms SZ in the evaluated high-fidelity regime. These results show that residual representations tailored to learned-compressor residuals can preserve the advantage of learned compression when global residual correction becomes rate-dominant.

LLM-as-judge evaluation is widely used in benchmarking pipelines, where model outputs are compared and ranked using automated evaluators. These pipelines typically assume that judgments are stable properties of fixed inputs. We show that this assumption does not hold under interaction. We study post-decision manipulability: the extent to which an evaluation outcome can be altered through subsequent conversation with the judge after an initial decision has been made. Across controlled experiments on MT-Bench and AlpacaEval, we find that LLM judges are highly stable under repeated and neutral reevaluation, yet become substantially reversible under targeted post-decision challenge. An anti-baseline challenge protocol shows that stable judgments can be overturned through motivated interaction, while a counterbalanced target-validation protocol separates this reversibility from net target-directed steering. These reversals have practical consequences: they can degrade agreement with human preferences, shift benchmark rankings, and produce harmful evaluation changes despite high self-reported confidence. Authority framing is especially destabilizing, and revised judgments are often accompanied by low-overlap justifications, suggesting post hoc rationalization rather than reliable error correction. We introduce the Evaluation Robustness Score (ERS) to quantify interactional robustness by combining reversal susceptibility with counterbalanced directional effects. Our findings identify post-decision interaction as a distinct failure mode for LLM-as-judge evaluation and motivate evaluation protocols that measure not only static agreement, but robustness under challenge.

Multi-table question answering requires models to retrieve relevant evidence, link schemas, and perform compositional reasoning across relational tables. Existing multi-table Q&A resources typically provide questions and final answers but lack reasoning supervision that explains how answers are derived. To address this gap, we construct a synthetic contrastive reasoning-trace dataset for MMQA by generating validated positive traces and plausible negative traces with heterogeneous LLMs. We then use the resulting preference pairs to fine-tune open-weight LLMs with Contrastive Preference Optimization (CPO). Across Qwen3-14B, Mistral-8B, and Llama-3.1-8B, CPO achieves absolute average improvements over Q&A supervised fine-tuning ranging from 9.7%-16.3%, with gains up to 21 percentage points on MMQA. Ablations show that heterogeneous positive and negative trace generators strengthen the contrastive signal, and automated as well as human evaluations indicate that the generated pairs are largely faithful, coherent, and meaningfully contrastive.

We propose an extended family of structured spatial priors that incorporates the total variation (TV) function with $\ell_p$ norms. The prior is proven to be proper and incorporated into a Bayesian regression framework to enable uncertainty quantification in $T_1$ mapping, with posterior inference performed using the No-U-Turn Sampler (NUTS). This TV--$\ell_p$ construction is proven to constitute a well-defined family of prior distributions, and it naturally enforces spatial consistency and smooth variations in the estimated parameter maps. The method was evaluated in comparison to maximum-likelihood estimation and several Bayesian alternative priors based on the uniform, Gamma, and bounded TV priors. The evaluation includes experiments on synthetic brain and cardiac $T_1$ mapping datasets, as well as a real in-vivo breast $T_1$ mapping dataset. The results show that the TV--$\ell_p$ prior yields more concentrated posterior densities, indicating reduced uncertainty. It also consistently achieves lower variance and smaller (negative) bias, leading to more reliable estimates. Overall, embedding a TV-based structured penalty along with $\ell_p$ norms in a prior in a Bayesian model improves spatial coherence in $T_1$ maps and enhances uncertainty quantification, offering a robust approach for $T_1$ mapping with uncertainties.

To develop and evaluate deep learning models for automated grading of age-related macular degeneration (AMD) severity using optical coherence tomography (OCT) and OCT angiography (OCTA) data. Two hundred seventy-one participants aged >= 50 years with varying AMD severities. Central macular 6 x 6 mm OCT/OCTA volumes were acquired using a swept-source OCTA system (SOLIX; Visionix/Optovue Inc., CA). AMD severity was graded into four stages (No AMD, Early AMD, Intermediate AMD, and Advanced AMD) according to the AREDS simplified severity scale. Three deep learning models were developed using different input modalities: (1) biomarker maps derived from segmented pathological features, including retinal fluid, drusen, geographic atrophy (GA), and macular neovascularization (MNV); (2) two-dimensional (2D) en face OCT and OCTA projections; and (3) three-dimensional (3D) OCT/OCTA volumes. EfficientNet-based architectures were trained using normalized inputs, data augmentation, and five-fold cross-validation. A total of 2,030 OCT/OCTA volumes from 351 eyes of 271 participants were analyzed. All models demonstrated strong AMD staging performance with substantial agreement with the reference standard (QWK >= 0.83). The biomarker-based model achieved the highest overall performance (QWK = 0.85 +/- 0.03, mean +/- standard deviation) and the best detection of early AMD (F1-score = 0.59 +/- 0.14). The 3D model achieved performance comparable to the 2D OCT/OCTA model (QWK = 0.83 +/- 0.04 vs. 0.83 +/- 0.09), while the 2D OCT/OCTA model showed the highest precision (0.79 +/- 0.06) and most accurately identified eyes without AMD. Deep learning models using OCT/OCTA data can accurately and automatically grade AMD severity. Among the evaluated approaches, the biomarker-based model provided the most balanced performance and showed particular value for early AMD detection.

We test whether a single screen-and-ablate recipe -- identify attention-head circuits by task-pattern selectivity, then verify by causal ablation against a matched-random null -- produces consistent mechanistic claims across model families. The recipe ports across pipelines; the specific circuit it identifies does not. Across four composed tasks (indirect-object identification, greater-than, successor sequences, variable binding) and three 1B-class language models from distinct training pipelines (Pythia 1B / Pile / dense; OLMo 1B / DCLM / dense; OLMoE 1B-7B / DCLM / mixture-of-experts), we run a unified protocol with the matched-random null sampled across ten seeds per cell. The resulting 12 (task, model) cells contain no two that share the same primary causal screen at comparable effect size: the same task, with the same behavioral capability, is implemented through different attention-pattern types across models. We introduce a five-category screen-outcome taxonomy -- primary cause, secondary cause, correlate, interferer, null -- with quantitative thresholds, and show that all five outcomes appear in the panel. We propose a falsifiable hypothesis: the MoE model in our panel builds composed-task circuits on top of a foundational previous-token positional substrate (the prev-token-circuit ablation is the strongest causal screen on 3 of 4 tasks for OLMoE 1B-7B), with the IOI exception consistent with IOI being a final-position name-copying task whose structure directly probes a different pattern. The hypothesis comes with explicit predictions for other MoE language models. We frame the methodology honestly: the spectral participation-ratio signal from the companion methodology paper is a general indicator of specialized computation; what makes a finding task-specific is the task-pattern screen plus a per-model causal verification.

Many human-centered tasks, including natural language inference (NLI) and emotion recognition (ER), have multiple plausible interpretations, leading to label ambiguity and challenging disagreements across human annotators. As LLMs are increasingly deployed in real-world settings, faithfully modeling such ambiguity is essential to identify contested inputs, preserve variability in ambiguous cases, and capture the full distribution of human judgments. Yet, existing LLM alignment approaches have predominantly assumed a single correct label, excluding annotator disagreement during optimization. Instead of treating this ambiguity as noise, we show how to treat it as information that improves model behavior through a new algorithm called SMARTLY HANDLING AMBIGUOUS LABELS IN ALIGNING LLMS (SHALA-LLM). This reinforcement learning framework provides a new way for LLMs to learn directly from annotator distributions while dynamically prioritizing highly ambiguous samples during optimization. Experiments on ambiguity-sensitive NLI and ER benchmarks, including ChaosNLI, GoEmotions, and MSP-Podcast, demonstrate that SHALA-LLM improves agreement with annotator label distributions, e.g. on ChaosNLI, it reduces Jensen-Shannon Distance by up to 62.1%. At the same time, SHALA-LLM improves F1 by up to 16.7%, showing that modeling annotator disagreement can also strengthen classification performance.

Optical coherence tomographic angiography (OCTA) is a powerful technique for imaging retinal microvasculature. However, acquiring reliable quantification of retinal blood flow and areas of retinal nonperfusion is challenging because of imaging artifacts. Existing methods primarily focus on noise suppression, projection artifact removal, or signal enhancement to improve the image quality of OCTA in cross-sectional or two-dimensional (2D) en face projections, while neglecting the intrinsic three-dimensional vascular architecture. In this study, we propose a deep learning-based algorithm for restoring capillary anatomical vasculature from a single OCTA volume. The network consists of an EfficientNet-B5 encoder and a decoder incorporating concurrent spatial and channel squeeze-and-excitation modules, connected via skip connections to preserve spatial resolution. Three adjacent B-frames are used as input to predict the restored middle B-frame. We evaluated the performance of the model using the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) against ground truth generated from averaging multiple scans. The results show that the proposed model significantly (both p < 0.001) improved image quality compared with the original single OCTA volume, with a PSNR of 26.16 +/- 1.26 vs. 22.23 +/- 0.78 and an SSIM of 0.91 +/- 0.02 vs. 0.72 +/- 0.03. The proposed model also significantly (p < 0.001) improved microvascular fidelity, measured by the Dice coefficient overlap between the model output and ground truth, in both 2D and 3D by at least 3.8% and 51.2%, respectively, across several different vascular slabs.

Reliable neural architecture selection is an open challenge in time-series forecasting under limited, noisy, and heterogeneous data, where standard heuristic architecture design and validation approaches fail to ensure accurate and reliable prediction and generalization. We propose EVIDENT (EVidence-based IDEntification of Neural archiTectures), a framework for architecture selection that integrates Bayesian training, evidence-based ranking, and task-specific validation under uncertainty. The framework explores the candidate architecture pool and identifies the lowest-capacity model that satisfies a prescribed validation criterion. We demonstrate this method using temporal convolutional networks (TCNs) for individualized blood glucose forecasting in type 1 diabetes patients. The results show that EVIDENT systematically rejects both under- and over-parameterized TCN architectures on population-level diabetes data, while identifying models that generalize reliably to unseen patients. When multiple architectures are competitive, the framework further supports plausibility-weighted ensemble predictions that enhance predictive performance. Compared with a random-search baseline, EVIDENT identified smaller architectures with more consistent forecasting performance on unseen patients. These findings establish EVIDENT as a strategy to neural architecture discovery, enabling reliable model selection for high-consequence forecasting in data-limited and heterogeneous settings.