AI Research Papers Ranked by Startup Potential

Complex problems, whether in math, logic, or planning, are solved by humans through a sequence of steps where the result of one step informs the next. In this work, we adopt the perspective that the reasoning power of Transformers is fundamentally limited by a fixed maximum number of steps along any latent path of computation. To address this, we introduce Turbo Connection (TurboConn), a novel architecture that overcomes the fixed-depth constraint by routing multiple residual connections from the higher-layer hidden states of each token $t$ to the lower layers of token $t+1$. Fine-tuning pre-trained LLMs with our method not only yields accuracy gains of 0.9% to over 10% on benchmarks like GSM8K, Parity, and multi-step arithmetic, but also demonstrates that the density of these backward connections is critical; our dense interaction significantly outperforms "sparse" alternatives that only pass a single hidden state or vector. Notably, TurboConn can be integrated into pre-trained LLMs to overcome task-specific plateaus: while a fine-tuned Qwen-3-1.7B achieves only 53.78% on Parity, adding our architectural modification enables the model to reach 100% accuracy, all without the necessity to retrain the full model from scratch or sophisticated curriculum learning. Our results provide strong empirical evidence that the depth of the computational path is a key factor in reasoning ability, also offering a new mechanism to enhance LLMs without significantly affecting generation latency.

In jurisdictions like India, where courts face an extensive backlog of cases, artificial intelligence offers transformative potential for legal judgment prediction. A critical subset of this backlog comprises appellate cases, which are formal decisions issued by higher courts reviewing the rulings of lower courts. To this end, we present Vichara, a novel framework tailored to the Indian judicial system that predicts and explains appellate judgments. Vichara processes English-language appellate case proceeding documents and decomposes them into decision points. Decision points are discrete legal determinations that encapsulate the legal issue, deciding authority, outcome, reasoning, and temporal context. The structured representation isolates the core determinations and their context, enabling accurate predictions and interpretable explanations. Vichara's explanations follow a structured format inspired by the IRAC (Issue-Rule-Application-Conclusion) framework and adapted for Indian legal reasoning. This enhances interpretability, allowing legal professionals to assess the soundness of predictions efficiently. We evaluate Vichara on two datasets, PredEx and the expert-annotated subset of the Indian Legal Documents Corpus (ILDC_expert), using four large language models: GPT-4o mini, Llama-3.1-8B, Mistral-7B, and Qwen2.5-7B. Vichara surpasses existing judgment prediction benchmarks on both datasets, with GPT-4o mini achieving the highest performance (F1: 81.5 on PredEx, 80.3 on ILDC_expert), followed by Llama-3.1-8B. Human evaluation of the generated explanations across Clarity, Linking, and Usefulness metrics highlights GPT-4o mini's superior interpretability.

Interactive perception (IP) enables robots to extract hidden information in their workspace and execute manipulation plans by physically interacting with objects and altering the state of the environment -- crucial for resolving occlusions and ambiguity in complex, partially observable scenarios. We present Zero-Shot IP (ZS-IP), a novel framework that couples multi-strategy manipulation (pushing and grasping) with a memory-driven Vision Language Model (VLM) to guide robotic interactions and resolve semantic queries. ZS-IP integrates three key components: (1) an Enhanced Observation (EO) module that augments the VLM's visual perception with both conventional keypoints and our proposed pushlines -- a novel 2D visual augmentation tailored to pushing actions, (2) a memory-guided action module that reinforces semantic reasoning through context lookup, and (3) a robotic controller that executes pushing, pulling, or grasping based on VLM output. Unlike grid-based augmentations optimized for pick-and-place, pushlines capture affordances for contact-rich actions, substantially improving pushing performance. We evaluate ZS-IP on a 7-DOF Franka Panda arm across diverse scenes with varying occlusions and task complexities. Our experiments demonstrate that ZS-IP outperforms passive and viewpoint-based perception techniques such as Mark-Based Visual Prompting (MOKA), particularly in pushing tasks, while preserving the integrity of non-target elements.

Online Multi-Agent Reinforcement Learning (MARL) is a prominent framework for efficient agent coordination. Crucially, enhancing policy expressiveness is pivotal for achieving superior performance. Diffusion-based generative models are well-positioned to meet this demand, having demonstrated remarkable expressiveness and multimodal representation in image generation and offline settings. Yet, their potential in online MARL remains largely under-explored. A major obstacle is that the intractable likelihoods of diffusion models impede entropy-based exploration and coordination. To tackle this challenge, we propose among the first \underline{O}nline off-policy \underline{MA}RL framework using \underline{D}iffusion policies (\textbf{OMAD}) to orchestrate coordination. Our key innovation is a relaxed policy objective that maximizes scaled joint entropy, facilitating effective exploration without relying on tractable likelihood. Complementing this, within the centralized training with decentralized execution (CTDE) paradigm, we employ a joint distributional value function to optimize decentralized diffusion policies. It leverages tractable entropy-augmented targets to guide the simultaneous updates of diffusion policies, thereby ensuring stable coordination. Extensive evaluations on MPE and MAMuJoCo establish our method as the new state-of-the-art across $10$ diverse tasks, demonstrating a remarkable $2.5\times$ to $5\times$ improvement in sample efficiency.

Modeling long sequences of user behaviors has emerged as a critical frontier in generative recommendation. However, existing solutions face a dilemma: linear attention mechanisms achieve efficiency at the cost of retrieval precision due to limited state capacity, while softmax attention suffers from prohibitive computational overhead. To address this challenge, we propose HyTRec, a model featuring a Hybrid Attention architecture that explicitly decouples long-term stable preferences from short-term intent spikes. By assigning massive historical sequences to a linear attention branch and reserving a specialized softmax attention branch for recent interactions, our approach restores precise retrieval capabilities within industrial-scale contexts involving ten thousand interactions. To mitigate the lag in capturing rapid interest drifts within the linear layers, we furthermore design Temporal-Aware Delta Network (TADN) to dynamically upweight fresh behavioral signals while effectively suppressing historical noise. Empirical results on industrial-scale datasets confirm the superiority that our model maintains linear inference speed and outperforms strong baselines, notably delivering over 8% improvement in Hit Rate for users with ultra-long sequences with great efficiency.

Recent work on test-time scaling for large language model (LLM) reasoning typically assumes that allocating more inference-time computation uniformly improves correctness. However, prior studies show that reasoning uncertainty is highly localized: a small subset of low-confidence tokens disproportionately contributes to reasoning errors and unnecessary output expansion. Motivated by this observation, we propose Thinking by Subtraction, a confidence-driven contrastive decoding approach that improves reasoning reliability through targeted token-level intervention. Our method, Confidence-Driven Contrastive Decoding, detects low-confidence tokens during decoding and intervenes selectively at these positions. It constructs a contrastive reference by replacing high-confidence tokens with minimal placeholders, and refines predictions by subtracting this reference distribution at low-confidence locations. Experiments show that CCD significantly improves accuracy across mathematical reasoning benchmarks while substantially reducing output length, with minimal KV-cache overhead. As a training-free method, CCD enhances reasoning reliability through targeted low-confidence intervention without computational redundancy. Our code will be made available at: https://github.com/bolo-web/CCD.

In voice conversion (VC) applications, diffusion and flow-matching models have exhibited exceptional speech quality and speaker similarity performances. However, they are limited by slow conversion owing to their iterative inference. Consequently, we propose MeanVoiceFlow, a novel one-step nonparallel VC model based on mean flows, which can be trained from scratch without requiring pretraining or distillation. Unlike conventional flow matching that uses instantaneous velocity, mean flows employ average velocity to more accurately compute the time integral along the inference path in a single step. However, training the average velocity requires its derivative to compute the target velocity, which can cause instability. Therefore, we introduce a structural margin reconstruction loss as a zero-input constraint, which moderately regularizes the input-output behavior of the model without harmful statistical averaging. Furthermore, we propose conditional diffused-input training in which a mixture of noise and source data is used as input to the model during both training and inference. This enables the model to effectively leverage source information while maintaining consistency between training and inference. Experimental results validate the effectiveness of these techniques and demonstrate that MeanVoiceFlow achieves performance comparable to that of previous multi-step and distillation-based models, even when trained from scratch. Audio samples are available at https://www.kecl.ntt.co.jp/people/kaneko.takuhiro/projects/meanvoiceflow/.

Vision-Language-Action (VLA) models enable instruction-following robotic manipulation, but they are typically pretrained on 2D data and lack 3D spatial understanding. An effective approach is representation alignment, where a strong vision foundation model is used to guide a 2D VLA model. However, existing methods usually apply supervision at only a single layer, failing to fully exploit the rich information distributed across depth; meanwhile, naïve multi-layer alignment can cause gradient interference. We introduce ROCKET, a residual-oriented multi-layer representation alignment framework that formulates multi-layer alignment as aligning one residual stream to another. Concretely, ROCKET employs a shared projector to align multiple layers of the VLA backbone with multiple layers of a powerful 3D vision foundation model via a layer-invariant mapping, which reduces gradient conflicts. We provide both theoretical justification and empirical analyses showing that a shared projector is sufficient and outperforms prior designs, and further propose a Matryoshka-style sparse activation scheme for the shared projector to balance multiple alignment losses. Our experiments show that, combined with a training-free layer selection strategy, ROCKET requires only about 4% of the compute budget while achieving 98.5% state-of-the-art success rate on LIBERO. We further demonstrate the superior performance of ROCKET across LIBERO-Plus and RoboTwin, as well as multiple VLA models. The code and model weights can be found at https://github.com/CASE-Lab-UMD/ROCKET-VLA.

Traditional AI alignment primarily focuses on individual model outputs; however, autonomous agents in long-horizon workflows require sustained reliability across entire interaction trajectories. We introduce APEMO (Affect-aware Peak-End Modulation for Orchestration), a runtime scheduling layer that optimizes computational allocation under fixed budgets by operationalizing temporal-affective signals. Instead of modifying model weights, APEMO detects trajectory instability through behavioral proxies and targets repairs at critical segments, such as peak moments and endings. Evaluation across multi-agent simulations and LLM-based planner--executor flows demonstrates that APEMO consistently enhances trajectory-level quality and reuse probability over structural orchestrators. Our results reframe alignment as a temporal control problem, offering a resilient engineering pathway for the development of long-horizon agentic systems.

This work studies heterogeneous Multi-Objective Reinforcement Learning (MORL), where objectives can differ sharply in temporal frequency. Such heterogeneity allows dense objectives to dominate learning, while sparse long-horizon rewards receive weak credit assignment, leading to poor sample efficiency. We propose a Parallel Reward Integration with Symmetry (PRISM) algorithm that enforces reflectional symmetry as an inductive bias in aligning reward channels. PRISM introduces ReSymNet, a theory-motivated model that reconciles temporal-frequency mismatches across objectives, using residual blocks to learn a scaled opportunity value that accelerates exploration while preserving the optimal policy. We also propose SymReg, a reflectional equivariance regulariser that enforces agent mirroring and constrains policy search to a reflection-equivariant subspace. This restriction provably reduces hypothesis complexity and improves generalisation. Across MuJoCo benchmarks, PRISM consistently outperforms both a sparse-reward baseline and an oracle trained with full dense rewards, improving Pareto coverage and distributional balance: it achieves hypervolume gains exceeding 100\% over the baseline and up to 32\% over the oracle. The code is at \href{https://github.com/EVIEHub/PRISM}{https://github.com/EVIEHub/PRISM}.

Providing scaffolding through educational chatbots built on Large Language Models (LLM) has potential risks and benefits that remain an open area of research. When students navigate impasses, they ask for help by formulating impasse-driven questions. Within interactions with LLM chatbots, such questions shape the user prompts and drive the pedagogical effectiveness of the chatbot's response. This paper focuses on such student questions from two datasets of distinct learning contexts: formative self-study, and summative assessed coursework. We analysed 6,113 messages from both learning contexts, using 11 different LLMs and three human raters to classify student questions using four existing schemas. On the feasibility of using LLMs as raters, results showed moderate-to-good inter-rater reliability, with higher consistency than human raters. The data showed that 'procedural' questions predominated in both learning contexts, but more so when students prepare for summative assessment. These results provide a basis on which to use LLMs for classification of student questions. However, we identify clear limitations in both the ability to classify with schemas and the value of doing so: schemas are limited and thus struggle to accommodate the semantic richness of composite prompts, offering only partial understanding the wider risks and benefits of chatbot integration. In the future, we recommend an analysis approach that captures the nuanced, multi-turn nature of conversation, for example, by applying methods from conversation analysis in discursive psychology.

Chain-of-thought (CoT) monitors are LLM-based systems that analyze reasoning traces to detect when outputs may exhibit attributes of interest, such as test-hacking behavior during code generation. In this paper, we use information-theoretic analysis to show that non-zero mutual information between CoT and output is a necessary but not sufficient condition for CoT monitorability. We identify two sources of approximation error that may undermine the performance of CoT monitors in practice: information gap, which measures the extent to which the monitor can extract the information available in CoT, and elicitation error, which measures the extent to which the monitor approximates the optimal monitoring function. We further demonstrate that CoT monitorability can be systematically improved through targeted training objectives. To this end, we propose two complementary approaches: (a) an oracle-based method that directly rewards the monitored model for producing CoTs that maximize monitor accuracy, and (b) a more practical, label-free approach that maximizes conditional mutual information between outputs and CoTs. Across multiple different environments, we show both methods significantly improve monitor accuracy while preventing CoT degeneration even when training against a monitor, thereby mitigating reward hacking when the task reward is imperfectly specified.

Decoding sits between a language model and everything we do with it, yet it is still treated as a heuristic knob-tuning exercise. We argue decoding should be understood as a principled optimisation layer: at each token, we solve a regularised problem over the probability simplex that trades off model score against structural preferences and constraints. This single template recovers greedy decoding, Softmax sampling, Top-K, Top-P, and Sparsemax-style sparsity as special cases, and explains their common structure through optimality conditions. More importantly, the framework makes it easy to invent new decoders without folklore. We demonstrate this by designing Best-of-K (BoK), a KL-anchored coverage objective aimed at multi-sample pipelines (self-consistency, reranking, verifier selection). BoK targets the probability of covering good alternatives within a fixed K-sample budget and improves empirical performance. We show that such samples can improve accuracy by, for example, +18.6% for Qwen2.5-Math-7B on MATH500 at high sampling temperatures.

While mechanistic interpretability has developed powerful tools to analyze the internal workings of Large Language Models (LLMs), their complexity has created an accessibility gap, limiting their use to specialists. We address this challenge by designing, building, and evaluating ELIA (Explainable Language Interpretability Analysis), an interactive web application that simplifies the outcomes of various language model component analyses for a broader audience. The system integrates three key techniques -- Attribution Analysis, Function Vector Analysis, and Circuit Tracing -- and introduces a novel methodology: using a vision-language model to automatically generate natural language explanations (NLEs) for the complex visualizations produced by these methods. The effectiveness of this approach was empirically validated through a mixed-methods user study, which revealed a clear preference for interactive, explorable interfaces over simpler, static visualizations. A key finding was that the AI-powered explanations helped bridge the knowledge gap for non-experts; a statistical analysis showed no significant correlation between a user's prior LLM experience and their comprehension scores, suggesting that the system reduced barriers to comprehension across experience levels. We conclude that an AI system can indeed simplify complex model analyses, but its true power is unlocked when paired with thoughtful, user-centered design that prioritizes interactivity, specificity, and narrative guidance.

Unsupervised representations are widely assumed to be neutral with respect to sensitive attributes when those attributes are withheld from training. We show that this assumption is false. Using SOMtime, a topology-preserving representation method based on high-capacity Self-Organizing Maps, we demonstrate that sensitive attributes such as age and income emerge as dominant latent axes in purely unsupervised embeddings, even when explicitly excluded from the input. On two large-scale real-world datasets (the World Values Survey across five countries and the Census-Income dataset), SOMtime recovers monotonic orderings aligned with withheld sensitive attributes, achieving Spearman correlations of up to 0.85, whereas PCA and UMAP typically remain below 0.23 (with a single exception reaching 0.31), and against t-SNE and autoencoders which achieve at most 0.34. Furthermore, unsupervised segmentation of SOMtime embeddings produces demographically skewed clusters, demonstrating downstream fairness risks without any supervised task. These findings establish that \textit{fairness through unawareness} fails at the representation level for ordinal sensitive attributes and that fairness auditing must extend to unsupervised components of machine learning pipelines. We have made the code available at~ https://github.com/JosephBingham/SOMtime

Scalable robot policy pre-training has been hindered by the high cost of collecting high-quality demonstrations for each platform. In this study, we address this issue by uniting offline reinforcement learning (offline RL) with cross-embodiment learning. Offline RL leverages both expert and abundant suboptimal data, and cross-embodiment learning aggregates heterogeneous robot trajectories across diverse morphologies to acquire universal control priors. We perform a systematic analysis of this offline RL and cross-embodiment paradigm, providing a principled understanding of its strengths and limitations. To evaluate this offline RL and cross-embodiment paradigm, we construct a suite of locomotion datasets spanning 16 distinct robot platforms. Our experiments confirm that this combined approach excels at pre-training with datasets rich in suboptimal trajectories, outperforming pure behavior cloning. However, as the proportion of suboptimal data and the number of robot types increase, we observe that conflicting gradients across morphologies begin to impede learning. To mitigate this, we introduce an embodiment-based grouping strategy in which robots are clustered by morphological similarity and the model is updated with a group gradient. This simple, static grouping substantially reduces inter-robot conflicts and outperforms existing conflict-resolution methods.

LLM-based systems increasingly generate structured workflows for complex tasks. In practice, automatic evaluation of these workflows is difficult, because metric scores are often not calibrated, and score changes do not directly communicate the severity of workflow degradation. We introduce WorkflowPerturb, a controlled benchmark for studying workflow evaluation metrics. It works by applying realistic, controlled perturbations to golden workflows. WorkflowPerturb contains 4,973 golden workflows and 44,757 perturbed variants across three perturbation types (Missing Steps, Compressed Steps, and Description Changes), each applied at severity levels of 10%, 30%, and 50%. We benchmark multiple metric families and analyze their sensitivity and calibration using expected score trajectories and residuals. Our results characterize systematic differences across metric families and support severity-aware interpretation of workflow evaluation scores. Our dataset will be released upon acceptance.

The paradigm of decision-making has been revolutionised by reinforcement learning and deep learning. Although this has led to significant progress in domains such as robotics, healthcare, and finance, the use of RL in practice is challenging, particularly when learning decision policies in high-stakes applications that may require guarantees. Traditional RL algorithms rely on a large number of online interactions with the environment, which is problematic in scenarios where online interactions are costly, dangerous, or infeasible. However, learning from offline datasets is hindered by the presence of hidden confounders. Such confounders can cause spurious correlations in the dataset and can mislead the agent into taking suboptimal or adversarial actions. Firstly, we address the problem of learning from offline datasets in the presence of hidden confounders. We work with instrumental variables (IVs) to identify the causal effect, which is an instance of a conditional moment restrictions (CMR) problem. Inspired by double/debiased machine learning, we derive a sample-efficient algorithm for solving CMR problems with convergence and optimality guarantees, which outperforms state-of-the-art algorithms. Secondly, we relax the conditions on the hidden confounders in the setting of (offline) imitation learning, and adapt our CMR estimator to derive an algorithm that can learn effective imitator policies with convergence rate guarantees. Finally, we consider the problem of learning high-level objectives expressed in linear temporal logic (LTL) and develop a provably optimal learning algorithm that improves sample efficiency over existing methods. Through evaluation on reinforcement learning benchmarks and synthetic and semi-synthetic datasets, we demonstrate the usefulness of the methods developed in this thesis in real-world decision making.

In active sequential testing, also termed pure exploration, a learner is tasked with the goal to adaptively acquire information so as to identify an unknown ground-truth hypothesis with as few queries as possible. This problem, originally studied by Chernoff in 1959, has several applications: classical formulations include Best-Arm Identification (BAI) in bandits, where actions index hypotheses, and generalized search problems, where strategically chosen queries reveal partial information about a hidden label. In many modern settings, however, the hypothesis space is continuous and naturally coincides with the query/action space: for example, identifying an optimal action in a continuous-armed bandit, localizing an $ε$-ball contained in a target region, or estimating the minimizer of an unknown function from a sequence of observations. In this work, we study pure exploration in such continuous spaces and introduce Continuous In-Context Pure Exploration for this regime. We introduce C-ICPE-TS, an algorithm that meta-trains deep neural policies to map observation histories to (i) the next continuous query action and (ii) a predicted hypothesis, thereby learning transferable sequential testing strategies directly from data. At inference time, C-ICPE-TS actively gathers evidence on previously unseen tasks and infers the true hypothesis without parameter updates or explicit hand-crafted information models. We validate C-ICPE-TS across a range of benchmarks, spanning continuous best-arm identification, region localization, and function minimizer identification.

The increasing deployment of Federated Learning (FL) in Intrusion Detection Systems (IDS) introduces new challenges related to data privacy, centralized coordination, and susceptibility to poisoning attacks. While significant research has focused on protecting traditional FL-IDS with centralized aggregation servers, there remains a notable gap in addressing the unique challenges of decentralized FL-IDS (DFL-IDS). This study aims to address the limitations of traditional centralized FL-IDS by proposing a novel defense framework tailored for the decentralized FL-IDS architecture, with a focus on privacy preservation and robustness against poisoning attacks. We propose PenTiDef, a privacy-preserving and robust defense framework for DFL-IDS, which incorporates Distributed Differential Privacy (DDP) to protect data confidentiality and utilizes latent space representations (LSR) derived from neural networks to detect malicious updates in the decentralized model aggregation context. To eliminate single points of failure and enhance trust without a centralized aggregation server, PenTiDef employs a blockchain-based decentralized coordination mechanism that manages model aggregation, tracks update history, and supports trust enforcement through smart contracts. Experimental results on CIC-IDS2018 and Edge-IIoTSet demonstrate that PenTiDef consistently outperforms existing defenses (e.g., FLARE, FedCC) across various attack scenarios and data distributions. These findings highlight the potential of PenTiDef as a scalable and secure framework for deploying DFL-based IDS in adversarial environments. By leveraging privacy protection, malicious behavior detection in hidden data, and working without a central server, it provides a useful security solution against real-world attacks from untrust participants.

In environments with sparse or delayed rewards, reinforcement learning (RL) incurs high sample complexity due to the large number of interactions needed for learning. This limitation has motivated the use of large language models (LLMs) for subgoal discovery and trajectory guidance. While LLMs can support exploration, frequent reliance on LLM calls raises concerns about scalability and reliability. We address these challenges by constructing a memory graph that encodes subgoals and trajectories from both LLM guidance and the agent's own successful rollouts. From this graph, we derive a utility function that evaluates how closely the agent's trajectories align with prior successful strategies. This utility shapes the advantage function, providing the critic with additional guidance without altering the reward. Our method relies primarily on offline input and only occasional online queries, avoiding dependence on continuous LLM supervision. Preliminary experiments in benchmark environments show improved sample efficiency and faster early learning compared to baseline RL methods, with final returns comparable to methods that require frequent LLM interaction.

Reinforcement learning (RL) agents often suffer from high sample complexity in sparse or delayed reward settings due to limited prior structure. Large language models (LLMs) can provide subgoal decompositions, plausible trajectories, and abstract priors that facilitate early learning. However, heavy reliance on LLM supervision introduces scalability constraints and dependence on potentially unreliable signals. We propose MIRA (Memory-Integrated Reinforcement Learning Agent), which incorporates a structured, evolving memory graph to guide early training. The graph stores decision-relevant information, including trajectory segments and subgoal structures, and is constructed from both the agent's high-return experiences and LLM outputs. This design amortizes LLM queries into a persistent memory rather than requiring continuous real-time supervision. From this memory graph, we derive a utility signal that softly adjusts advantage estimation to influence policy updates without modifying the underlying reward function. As training progresses, the agent's policy gradually surpasses the initial LLM-derived priors, and the utility term decays, preserving standard convergence guarantees. We provide theoretical analysis showing that utility-based shaping improves early-stage learning in sparse-reward environments. Empirically, MIRA outperforms RL baselines and achieves returns comparable to approaches that rely on frequent LLM supervision, while requiring substantially fewer online LLM queries. Project webpage: https://narjesno.github.io/MIRA/

Current biomedical question answering (QA) systems often assume that medical knowledge applies uniformly, yet real-world clinical reasoning is inherently conditional: nearly every decision depends on patient-specific factors such as comorbidities and contraindications. Existing benchmarks do not evaluate such conditional reasoning, and retrieval-augmented or graph-based methods lack explicit mechanisms to ensure that retrieved knowledge is applicable to given context. To address this gap, we propose CondMedQA, the first benchmark for conditional biomedical QA, consisting of multi-hop questions whose answers vary with patient conditions. Furthermore, we propose Condition-Gated Reasoning (CGR), a novel framework that constructs condition-aware knowledge graphs and selectively activates or prunes reasoning paths based on query conditions. Our findings show that CGR more reliably selects condition-appropriate answers while matching or exceeding state-of-the-art performance on biomedical QA benchmarks, highlighting the importance of explicitly modeling conditionality for robust medical reasoning.

Traditional neural topic models are typically optimized by reconstructing the document's Bag-of-Words (BoW) representations, overlooking contextual information and struggling with data sparsity. In this work, we propose a novel approach to construct semantically-grounded soft label targets using Language Models (LMs) by projecting the next token probabilities, conditioned on a specialized prompt, onto a pre-defined vocabulary to obtain contextually enriched supervision signals. By training the topic models to reconstruct the soft labels using the LM hidden states, our method produces higher-quality topics that are more closely aligned with the underlying thematic structure of the corpus. Experiments on three datasets show that our method achieves substantial improvements in topic coherence, purity over existing baselines. Additionally, we also introduce a retrieval-based metric, which shows that our approach significantly outperforms existing methods in identifying semantically similar documents, highlighting its effectiveness for retrieval-oriented applications.

Recent advances in deep learning, exemplified by Hyper-Connections (HC), have expanded the residual connection paradigm by introducing wider residual streams and diverse connectivity patterns. While these innovations yield significant performance gains, they compromise the identity mapping property of residual connections, leading to training instability, limited scalability, and increased memory overhead. To address these challenges, we propose JPmHC (Jacobian-spectrum Preserving manifold-constrained Hyper-Connections), a framework that replaces identity skips with a trainable linear mixer acting on n parallel streams while explicitly controlling gradient conditioning. By constraining the mixer M on operator-norm-bounded manifolds (e.g., bistochastic, Stiefel, Grassmann), JPmHC prevents gradient pathologies and enhances stability. JPmHC introduces three key contributions: (i) a free-probability analysis that predicts Jacobian spectra for structured skips, providing actionable design rules for mixer selection; (ii) memory-efficient implicit differentiation for fixed-point projections, reducing activation memory and synchronization overhead; and (iii) a Stiefel-constrained mixer via Cayley transforms, ensuring orthogonality without post-hoc normalization. Empirical evaluations on ARC-AGI demonstrate that JPmHC achieves faster convergence, higher accuracy, and lower computational cost compared to bistochastic baselines. As a flexible and scalable extension of HC, JPmHC advances spectrum-aware, stable, and efficient deep learning, offering insights into topological architecture design and foundational model evolution.

Large Language Models (LLMs) demonstrate significant potential in multi-agent negotiation tasks, yet evaluation in this domain remains challenging due to a lack of robust and generalizable benchmarks. Abdelnabi et al. (2024) introduce a negotiation benchmark based on Scoreable Games, with the aim of developing a highly complex and realistic evaluation framework for LLMs. Our work investigates the reproducibility of claims in their benchmark, and provides a deeper understanding of its usability and generalizability. We replicate the original experiments on additional models, and introduce additional metrics to verify negotiation quality and evenness of evaluation. Our findings reveal that while the benchmark is indeed complex, model comparison is ambiguous, raising questions about its objectivity. Furthermore, we identify limitations in the experimental setup, particularly in information leakage detection and thoroughness of the ablation study. By examining and analyzing the behavior of a wider range of models on an extended version of the benchmark, we reveal insights that provide additional context to potential users. Our results highlight the importance of context in model-comparative evaluations.

Despite having hundreds of millions of speakers, handwritten Devanagari text remains severely underrepresented in publicly available benchmark datasets. Existing resources are limited in scale, focus primarily on isolated characters or short words, and lack controlled lexical content and writer level diversity, which restricts their utility for modern data driven handwriting analysis. As a result, they fail to capture the continuous, fused, and structurally complex nature of Devanagari handwriting, where characters are connected through a shared shirorekha (horizontal headline) and exhibit rich ligature formations. We introduce DohaScript, a large scale, multi writer dataset of handwritten Hindi text collected from 531 unique contributors. The dataset is designed as a parallel stylistic corpus, in which all writers transcribe the same fixed set of six traditional Hindi dohas (couplets). This controlled design enables systematic analysis of writer specific variation independent of linguistic content, and supports tasks such as handwriting recognition, writer identification, style analysis, and generative modeling. The dataset is accompanied by non identifiable demographic metadata, rigorous quality curation based on objective sharpness and resolution criteria, and page level layout difficulty annotations that facilitate stratified benchmarking. Baseline experiments demonstrate clear quality separation and strong generalization to unseen writers, highlighting the dataset's reliability and practical value. DohaScript is intended to serve as a standardized and reproducible benchmark for advancing research on continuous handwritten Devanagari text in low resource script settings.

In this work, we present HiAER-Spike, a modular, reconfigurable, event-driven neuromorphic computing platform designed to execute large spiking neural networks with up to 160 million neurons and 40 billion synapses - roughly twice the neurons of a mouse brain at faster than real time. This system, assembled at the UC San Diego Supercomputer Center, comprises a co-designed hard- and software stack that is optimized for run-time massively parallel processing and hierarchical address-event routing (HiAER) of spikes while promoting memory-efficient network storage and execution. The architecture efficiently handles both sparse connectivity and sparse activity for robust and low-latency event-driven inference for both edge and cloud computing. A Python programming interface to HiAER-Spike, agnostic to hardware-level detail, shields the user from complexity in the configuration and execution of general spiking neural networks with minimal constraints in topology. The system is made easily available over a web portal for use by the wider community. In the following, we provide an overview of the hard- and software stack, explain the underlying design principles, demonstrate some of the system's capabilities and solicit feedback from the broader neuromorphic community. Examples are shown demonstrating HiAER-Spike's capabilities for event-driven vision on benchmark CIFAR-10, DVS event-based gesture, MNIST, and Pong tasks.

Evaluation is no longer a final checkpoint in the machine learning lifecycle. As AI systems evolve from static models to compound, tool-using agents, evaluation becomes a core control function. The question is no longer "How good is the model?" but "Can we trust the system to behave as intended, under change, at scale?". Yet most evaluation practices remain anchored in assumptions inherited from the model-centric era: static benchmarks, aggregate scores, and one-off success criteria. This paper argues that such approaches are increasingly obscure rather than illuminating system behavior. We examine how evaluation pipelines themselves introduce silent failure modes, why high benchmark scores routinely mislead teams, and how agentic systems fundamentally alter the meaning of performance measurement. Rather than proposing new metrics or harder benchmarks, we aim to clarify the role of evaluation in the AI era, and especially for agents: not as performance theater, but as a measurement discipline that conditions trust, iteration, and governance in non-deterministic systems.

Graph data often contain noisy and spurious correlations that mask the true causal relationships, which are essential for enabling graph models to make predictions based on the underlying causal structure of the data. Dependence on spurious connections makes it challenging for traditional Graph Neural Networks (GNNs) to generalize effectively across different graphs. Furthermore, traditional aggregation methods tend to amplify these spurious patterns, limiting model robustness under distribution shifts. To address these issues, we propose Causal Neighbourhood Learning with Graph Neural Networks (CNL-GNN), a novel framework that performs causal interventions on graph structure. CNL-GNN effectively identifies and preserves causally relevant connections and reduces spurious influences through the generation of counterfactual neighbourhoods and adaptive edge perturbation guided by learnable importance masking and an attention-based mechanism. In addition, by combining structural-level interventions with the disentanglement of causal features from confounding factors, the model learns invariant node representations that are robust and generalize well across different graph structures. Our approach improves causal graph learning beyond traditional feature-based methods, resulting in a robust classification model. Extensive experiments on four publicly available datasets, including multiple domain variants of one dataset, demonstrate that CNL-GNN outperforms state-of-the-art GNN models.

Real-world knowledge representation often requires capturing subjective, continuous attributes -- such as political positions -- that conflict with pairwise validation, the widely accepted gold standard for human evaluation. We address this challenge through a dual-scale validation framework applied to political stance prediction in argumentative discourse, combining pointwise and pairwise human annotation. Using 22 language models, we construct a large-scale knowledge base of political position predictions for 23,228 arguments drawn from 30 debates that appeared on the UK politicial television programme \textit{Question Time}. Pointwise evaluation shows moderate human-model agreement (Krippendorff's $α=0.578$), reflecting intrinsic subjectivity, while pairwise validation reveals substantially stronger alignment between human- and model-derived rankings ($α=0.86$ for the best model). This work contributes: (i) a practical validation methodology for subjective continuous knowledge that balances scalability with reliability; (ii) a validated structured argumentation knowledge base enabling graph-based reasoning and retrieval-augmented generation in political domains; and (iii) evidence that ordinal structure can be extracted from pointwise language models predictions from inherently subjective real-world discourse, advancing knowledge representation capabilities for domains where traditional symbolic or categorical approaches are insufficient.

The expressive power of Graph Neural Networks (GNNs) is often analysed via correspondence to the Weisfeiler-Leman (WL) algorithm and fragments of first-order logic. Standard GNNs are limited to performing aggregation over immediate neighbourhoods or over global read-outs. To increase their expressivity, recent attempts have been made to incorporate substructural information (e.g. cycle counts and subgraph properties). In this paper, we formalize this architectural trend by introducing Template GNNs (T-GNNs), a generalized framework where node features are updated by aggregating over valid template embeddings from a specified set of graph templates. We propose a corresponding logic, Graded template modal logic (GML(T)), and generalized notions of template-based bisimulation and WL algorithm. We establish an equivalence between the expressive power of T-GNNs and GML(T), and provide a unifying approach for analysing GNN expressivity: we show how standard AC-GNNs and its recent variants can be interpreted as instantiations of T-GNNs.

Biological neural networks (like the hippocampus) can internally generate "replay" resembling stimulus-driven activity. Recent computational models of replay use noisy recurrent neural networks (RNNs) trained to path-integrate. Replay in these networks has been described as Langevin sampling, but new modifiers of noisy RNN replay have surpassed this description. We re-examine noisy RNN replay as sampling to understand or improve it in three ways: (1) Under simple assumptions, we prove that the gradients replay activity should follow are time-varying and difficult to estimate, but readily motivate the use of hidden state leakage in RNNs for replay. (2) We confirm that hidden state adaptation (negative feedback) encourages exploration in replay, but show that it incurs non-Markov sampling that also slows replay. (3) We propose the first model of temporally compressed replay in noisy path-integrating RNNs through hidden state momentum, connect it to underdamped Langevin sampling, and show that, together with adaptation, it counters slowness while maintaining exploration. We verify our findings via path-integration of 2D triangular and T-maze paths and of high-dimensional paths of synthetic rat place cell activity.

Discrete image tokenizers encode visual inputs as sequences of tokens from a finite vocabulary and are gaining popularity in multimodal systems, including encoder-only, encoder-decoder, and decoder-only models. However, unlike CLIP encoders, their vulnerability to adversarial attacks has not been explored. Ours being the first work studying this topic, we first formulate attacks that aim to perturb the features extracted by discrete tokenizers, and thus change the extracted tokens. These attacks are computationally efficient, application-agnostic, and effective across classification, multimodal retrieval, and captioning tasks. Second, to defend against this vulnerability, inspired by recent work on robust CLIP encoders, we fine-tune popular tokenizers with unsupervised adversarial training, keeping all other components frozen. While unsupervised and task-agnostic, our approach significantly improves robustness to both unsupervised and end-to-end supervised attacks and generalizes well to unseen tasks and data. Unlike supervised adversarial training, our approach can leverage unlabeled images, making it more versatile. Overall, our work highlights the critical role of tokenizer robustness in downstream tasks and presents an important step in the development of safe multimodal foundation models.

Conversational AI has been proposed as a scalable way to correct public misconceptions and spread misinformation. Yet its effectiveness may depend on perceptions of its political neutrality. As LLMs enter partisan conflict, elites increasingly portray them as ideologically aligned. We test whether these credibility attacks reduce LLM-based persuasion. In a preregistered U.S. survey experiment (N=2144), participants completed a three-round conversation with ChatGPT about a personally held economic policy misconception. Compared to a neutral control, a short message indicating that the LLM was biased against the respondent's party attenuated persuasion by 28%. Transcript analysis indicates that the warnings alter the interaction: respondents push back more and engage less receptively. These findings suggest that the persuasive impact of conversational AI is politically contingent, constrained by perceptions of partisan alignment.

Reinforcement Learning from Human Feedback (RLHF) or Verifiable Rewards (RLVR) are two key steps in the post-training of modern Language Models (LMs). A common problem is reward hacking, where the policy may exploit inaccuracies of the reward and learn an unintended behavior. Most previous works address this by limiting the policy update with a Kullback-Leibler (KL) penalty towards a reference model. We propose a different framing: Train the LM in a way that biases policy updates towards regions in which the reward is more accurate. First, we derive a theoretical connection between the accuracy of a reward model and the flatness of an optimum at convergence. Gradient regularization (GR) can then be used to bias training to flatter regions and thereby maintain reward model accuracy. We confirm these results by showing that the gradient norm and reward accuracy are empirically correlated in RLHF. We then show that Reference Resets of the KL penalty implicitly use GR to find flatter regions with higher reward accuracy. We further improve on this by proposing to use explicit GR with an efficient finite-difference estimate. Empirically, GR performs better than a KL penalty across a diverse set of RL experiments with LMs. GR achieves a higher GPT-judged win-rate in RLHF, avoids overly focusing on the format in rule-based math rewards, and prevents hacking the judge in LLM-as-a-Judge math tasks.