In Audio-Visual Navigation (AVN), agents must locate sound sources in unseen 3D environments using visual and auditory cues. However, existing methods often struggle with generalization in unseen scenarios, as they tend to overfit to semantic sound features and specific training environments. To address these challenges, we propose the \textbf{Binaural Difference Attention with Action Transition Prediction (BDATP)} framework, which jointly optimizes perception and policy. Specifically, the \textbf{Binaural Difference Attention (BDA)} module explicitly models interaural differences to enhance spatial orientation, reducing reliance on semantic categories. Simultaneously, the \textbf{Action Transition Prediction (ATP)} task introduces an auxiliary action prediction objective as a regularization term, mitigating environment-specific overfitting. Extensive experiments on the Replica and Matterport3D datasets demonstrate that BDATP can be seamlessly integrated into various mainstream baselines, yielding consistent and significant performance gains. Notably, our framework achieves state-of-the-art Success Rates across most settings, with a remarkable absolute improvement of up to 21.6 percentage points in Replica dataset for unheard sounds. These results underscore BDATP's superior generalization capability and its robustness across diverse navigation architectures.

This survey examines recent sensor-based planning and control methods for Unmanned Underwater Vehicles (UUVs). In complex, uncertain underwater environments, UUVs require advanced planning and control strategies for effective navigation. These vehicles face significant challenges including drifting and noisy sensor measurements, absence of Global Navigation Satellite System (GNSS) signals, and low-bandwidth, high-latency underwater acoustic communications. The focus is on reactive local planning layers that adapt to real-time sensor inputs such as SONAR and Inertial Measurement Units (IMU) to improve localization accuracy and autonomy in dynamic ocean conditions, enabling dynamic obstacle avoidance and on-the-fly re-planning. The survey categorizes the existing literature into decoupled and coupled architectures for sensor-based planning and control. The decoupled architecture sequentially addresses planning and control stages, whereas coupled architectures offer tighter feedback loops for more immediate responsiveness. A comparative analysis of coupled planning and control methods reveals that while PID controllers are simple, they lack predictive capability for complex maneuvers. Model Predictive Control (MPC) offers superior path optimization but can be computationally intensive, and invariant-set controllers provide strong safety guarantees at the potential cost of agility in confined environments. Key contributions include a taxonomy of architectures combining planning and control, a focus on adaptive local planning, and an analysis of controller roles in integrated planning frameworks for autonomous navigation of UUVs.

Multimodal self-supervised pretraining offers a promising route to cancer prognosis by integrating histopathology whole-slide images, gene expression, and pathology reports, yet most existing approaches require fully paired and complete inputs. In practice, clinical cohorts are fragmented and often miss one or more modalities, limiting both supervised fusion and scalable multimodal pretraining. We propose PRIME, a missing-aware multimodal self-supervised pretraining framework that learns robust and transferable representations from partially observed cohorts. PRIME maps heterogeneous modality embeddings into a unified token space and introduces a shared prototype memory bank for latent-space semantic imputation via patient-level consensus retrieval, producing structurally aligned tokens without reconstructing raw signals. Two complementary pretraining objectives: inter-modality alignment and post-fusion consistency under structured missingness augmentation, jointly learn representations that remain predictive under arbitrary modality subsets. We evaluate PRIME on The Cancer Genome Atlas with label-free pretraining on 32 cancer types and downstream 5-fold evaluation on five cohorts across overall survival prediction, 3-year mortality classification, and 3-year recurrence classification. PRIME achieves the best macro-average performance among all compared methods, reaching 0.653 C-index, 0.689 AUROC, and 0.637 AUROC on the three tasks, respectively, while improving robustness under test-time missingness and supporting parameter-efficient and label-efficient adaptation. These results support missing-aware multimodal pretraining as a practical strategy for prognosis modeling in fragmented clinical data settings.

This paper proposes a novel framework for enhancing the prediction accuracy and lead time of El Niño events, crucial for mitigating their global climatic, economic, and societal impacts. Traditional prediction models often rely on oceanic and atmospheric indices, which may lack the granularity or dynamic interplay captured by comprehensive meteorological and geographical datasets. Our framework integrates real-time global weather forecast data with anomalies, subsurface ocean heat content, and atmospheric pressure across various temporal and spatial resolutions. Leveraging a hybrid deep learning architecture that combines a Convolutional Neural Network (CNN) for spatial feature extraction and a Long Short-Term Memory (LSTM) network for temporal dependency modeling, the framework aims to identify complex precursors and evolving patterns of El Niño events.

Strategically locating a sawmill is vital for enhancing the efficiency, profitability, and sustainability of timber supply chains. Our study proposes a Learning-Based Multi-Criteria Decision-Making (LB-MCDM) framework that integrates machine learning (ML) with GIS-based spatial location analysis via MCDM. The proposed framework provides a data-driven, unbiased, and replicable approach to assessing site suitability. We demonstrate the utility of the proposed model through a case study in Mississippi (MS). We apply five ML algorithms (Random Forest Classifier, Support Vector Classifier, XGBoost Classifier, Logistic Regression, and K-Nearest Neighbors Classifier) to identify the most suitable sawmill locations in Mississippi. Among these models, the Random Forest Classifier achieved the highest performance. We use the SHAP (SHapley Additive exPlanations) technique to determine the relative importance of each criterion, revealing the Supply-Demand Ratio, a composite feature that reflects local market competition dynamics, as the most influential factor, followed by Road, Rail Line and Urban Area Distance. The validation of suitability maps generated by our LB-MCDM model suggests that 10-11% of the MS landscape is highly suitable for sawmill location.

Speculative sampling (SpS) has been successful in accelerating the decoding throughput of auto-regressive large language models by leveraging smaller draft models. SpS strictly enforces the generated distribution to match that of the verifier LLM. This is unnecessarily restrictive as slight variations of the verifier's distribution, such as sampling with top-$k$ or temperature, would also be acceptable. Typical acceptance sampling (TAS) alleviates this issue by accepting more tokens using entropy-based heuristics. However, this approach distorts the verifier distribution, potentially degrading output quality when the verifier encodes critical information. In this work, we formalize the speculative sampling algorithm through the lens of constrained optimization. Based on this formulation, we propose Cactus (constrained acceptance speculative sampling), a method that guarantees controlled divergence from the verifier distribution and increasing acceptance rates. Empirical results across a wide range of benchmarks confirm the effectiveness of our approach.

Model-free deep reinforcement learning (DRL) methods suffer from poor sample efficiency. To overcome this limitation, this work introduces an adaptive reduced-order-model (ROM)-based reinforcement learning framework for active flow control. In contrast to conventional actor--critic architectures, the proposed approach leverages a ROM to estimate the gradient information required for controller optimization. The design of the ROM structure incorporates physical insights. The ROM integrates a linear dynamical system and a neural ordinary differential equation (NODE) for estimating the nonlinearity in the flow. The parameters of the linear component are identified via operator inference, while the NODE is trained in a data-driven manner using gradient-based optimization. During controller--environment interactions, the ROM is continuously updated with newly collected data, enabling adaptive refinement of the model. The controller is then optimized through differentiable simulation of the ROM. The proposed ROM-based DRL framework is validated on two canonical flow control problems: Blasius boundary layer flow and flow past a square cylinder. For the Blasius boundary layer, the proposed method effectively reduces to a single-episode system identification and controller optimization process, yet it yields controllers that outperform traditional linear designs and achieve performance comparable to DRL approaches with minimal data. For the flow past a square cylinder, the proposed method achieves superior drag reduction with significantly fewer exploration data compared with DRL approaches. The work addresses a key component of model-free DRL control algorithms and lays the foundation for designing more sample-efficient DRL-based active flow controllers.

We present Territory Paint Wars, a minimal competitive multi-agent reinforcement learning environment implemented in Unity, and use it to systematically investigate failure modes of Proximal Policy Optimisation (PPO) under self-play. A first agent trained for $84{,}000$ episodes achieves only $26.8\%$ win rate against a uniformly-random opponent in a symmetric zero-sum game. Through controlled ablations we identify five implementation-level failure modes -- reward-scale imbalance, missing terminal signal, ineffective long-horizon credit assignment, unnormalised observations, and incorrect win detection -- each of which contributes critically to this failure in this setting. After correcting these issues, we uncover a distinct emergent pathology: competitive overfitting, where co-adapting agents maintain stable self-play performance while generalisation win rate collapses from $73.5\%$ to $21.6\%$. Critically, this failure is undetectable via standard self-play metrics: both agents co-adapt equally, so the self-play win rate remains near $50\%$ throughout the collapse. We propose a minimal intervention -- opponent mixing, where $20\%$ of training episodes substitute a fixed uniformly-random policy for the co-adaptive opponent -- which mitigates competitive overfitting and restores generalisation to $77.1\%$ ($\pm 12.6\%$, $10$ seeds) without population-based training or additional infrastructure. We open-source Territory Paint Wars to provide a reproducible benchmark for studying competitive MARL failure modes.

Experimental access to real honeybee colonies requires robotic systems capable of operating within limited spatial constraints, tolerating hive-specific fouling and environmental conditions, and supporting both sensing and localized actuation without frequent hardware redesign. This paper introduces COMB, a compact, open-source, modular mechatronic platform designed for in-hive experiments within standard observation-hive frames. The platform integrates a XY positioning stage, a Movable Access Window (MAW) for sealed tool access through the hive boundary, interchangeable payload modules, and an embedded control architecture that enables repeatable trajectory execution and signal generation. The platform's capabilities are demonstrated through three representative modules: a biomimetic dance-and-signaling payload, a close-range comb scanner, and an electromagnetic wing actuator for localized oscillatory stimulation. This paper details the hardware and software design of COMB, outlines its operational capabilities, and describes the supporting infrastructure for conducting real-world in-hive experiments. The platform is characterized in engineering terms through tracking waggle-trajectory executions, performing multi-image stitching for repeated comb mosaics, and conducting video-based spectral analysis of the wing actuator. These results position COMB as a reusable experimental robotics platform for controlled in-hive sensing and actuation, and as a compact, generalized successor to earlier task-specific honeybee robotic systems.

Large Vision-Language Models (LVLMs) suffer from prohibitive inference costs due to the massive number of visual tokens processed by the language decoder. Existing pruning methods often lead to significant performance degradation because the irreversible removal of visual tokens causes a distribution shift in the hidden states that deviates from the pre-trained full-token regime. To address this, we propose Representation Consistency Pruner, which we refer to as RCP, as a novel framework that integrates cumulative visual token pruning with a delayed repair mechanism. Specifically, we introduce a cross-attention pruner that leverages the intrinsic attention of the LLM as a baseline to predict cumulative masks, ensuring consistent and monotonic token reduction across layers. To compensate for the resulting information loss, we design a delayed repair adapter denoted as DRA, which caches the essence of pruned tokens and applies FiLM-based modulation specifically to the answer generation tokens. We employ a repair loss to match the first and second-order statistics of the pruned representations with a full-token teacher. RCP is highly efficient because it trains only lightweight plug-in modules while allowing for physical token discarding at inference. Extensive experiments on LVLM benchmarks demonstrate that RCP removes up to 88.9\% of visual tokens and reduces FLOPs by up to 85.7\% with only a marginal average accuracy drop, and outperforms prior methods that avoid fine-tuning the original model on several widely used benchmarks.

While Physics-Informed Neural Networks offer a promising framework for solving partial differential equations, the standard $L^2$ loss formulation is fundamentally insufficient when applied to the Bhatnagar-Gross-Krook (BGK) model. Specifically, simply minimizing the standard loss does not guarantee accurate predictions of the macroscopic moments, causing the approximate solutions to fail in capturing the true physical solution. To overcome this limitation, we introduce a velocity-weighted $L^2$ loss function designed to effectively penalize errors in the high-velocity regions. By establishing a stability estimate for the proposed approach, we shows that minimizing the proposed weighted loss guarantees the convergence of the approximate solution. Also, numerical experiments demonstrate that employing this weighted PINN loss leads to superior accuracy and robustness across various benchmarks compared to the standard approach.

Robots operating in shared workspaces must maintain safe coordination with other agents whose behavior may change during task execution. When a collaborating agent switches strategy mid-episode, continuing under outdated assumptions can lead to unsafe actions and increased collision risk. Reliable detection of such behavioral regime changes is therefore critical. We study regime-switch detection under controlled non-stationarity in ManiSkill shared-workspace manipulation tasks. Across ten detection methods and five random seeds, enabling detection reduces post-switch collisions by 52%. However, average performance hides significant reliability differences: under a realistic tolerance of +-3 steps, detection ranges from 86% to 30%, while under +-5 steps all methods achieve 100%. We introduce UA-TOM, a lightweight belief-tracking module that augments frozen vision-language-action (VLA) control backbones using selective state-space dynamics, causal attention, and prediction-error signals. Across five seeds and 1200 episodes, UA-TOM achieves the highest detection rate among unassisted methods (85.7% at +-3) and the lowest close-range time (4.8 steps), outperforming an Oracle (5.3 steps). Analysis shows hidden-state update magnitude increases by 17x at regime switches and decays over roughly 10 timesteps, while the discretization step converges to a near-constant value (Delta_t approx 0.78), indicating sensitivity driven by learned dynamics rather than input-dependent gating. Cross-domain experiments in Overcooked show complementary roles of causal attention and prediction-error signals. UA-TOM introduces 7.4 ms inference overhead (14.8% of a 50 ms control budget), enabling reliable regime-switch detection without modifying the base policy.

The domain of automatic video trailer generation is currently undergoing a profound paradigm shift, transitioning from heuristic-based extraction methods to deep generative synthesis. While early methodologies relied heavily on low-level feature engineering, visual saliency, and rule-based heuristics to select representative shots, recent advancements in Large Language Models (LLMs), Multimodal Large Language Models (MLLMs), and diffusion-based video synthesis have enabled systems that not only identify key moments but also construct coherent, emotionally resonant narratives. This survey provides a comprehensive technical review of this evolution, with a specific focus on generative techniques including autoregressive Transformers, LLM-orchestrated pipelines, and text-to-video foundation models like OpenAI's Sora and Google's Veo. We analyze the architectural progression from Graph Convolutional Networks (GCNs) to Trailer Generation Transformers (TGT), evaluate the economic implications of automated content velocity on User-Generated Content (UGC) platforms, and discuss the ethical challenges posed by high-fidelity neural synthesis. By synthesizing insights from recent literature, this report establishes a new taxonomy for AI-driven trailer generation in the era of foundation models, suggesting that future promotional video systems will move beyond extractive selection toward controllable generative editing and semantic reconstruction of trailers.

The reversal curse describes a failure of autoregressive language models to retrieve a fact in reverse order (e.g., training on ``$A > B$'' but failing on ``$B < A$''). Recent work shows that objectives with bidirectional supervision (e.g., bidirectional attention or masking-based reconstruction for decoder-only models) can mitigate the reversal curse. We extend this evaluation to include a vanilla masked language modeling (MLM) objective and compare it to decoder-only masking-based training across four reversal benchmarks and then provide a minimal mechanistic study of \emph{how} these objectives succeed. We show that reversal accuracy requires training signal that explicitly makes the source entity a prediction target, and we find little evidence that success corresponds to a single direction-agnostic representation of a fact. Instead, representation distances and linear probes are consistent with storing forward and reverse directions as distinct entries, with different indexing geometry for MLM versus decoder-only masking-based training. Our results caution that objective-level ``fixes'' can improve reversal behavior without necessarily inducing the kind of latent generalization one might expect from a unified concept.

Enhancing the reasoning capability of large language models (LLMs) remains a core challenge in natural language processing. The Chain-of-Thought (CoT) paradigm dominates practical applications for its single-round efficiency, yet its reasoning chains often exhibit logical gaps. While multi-round paradigms like Graph-of-Thoughts (GoT), Tree-of-Thoughts (ToT), and Atom of Thought (AoT) achieve strong performance and reveal effective reasoning structures, their high cost limits practical use. To address this problem, this paper proposes a topology-based method for optimizing reasoning chains. The framework embeds essential topological patterns of effective reasoning into the lightweight CoT paradigm. Using persistent homology, we map CoT, ToT, and GoT into a unified topological space to quantify their structural features. On this basis, we design a unified optimization system: a Topological Optimization Agent diagnoses deviations in CoT chains from desirable topological characteristics and simultaneously generates targeted strategies to repair these structural deficiencies. Compared with multi-round reasoning methods like ToT and GoT, experiments on multiple datasets show that our approach offers a superior balance between reasoning accuracy and efficiency, showcasing a practical solution to ``single-round generation with multi-round intelligence''.

Many combinatorial optimisation problems hide algebraic structures that, once exposed, shrink the search space and improve the chance of finding the global optimal solution. We present a general framework that (i) identifies algebraic structure, (ii) formalises operations, (iii) constructs quotient spaces that collapse redundant representations, and (iv) optimises directly over these reduced spaces. Across a broad family of rule-combination tasks (e.g., patient subgroup discovery and rule-based molecular screening), conjunctive rules form a monoid. Via a characteristic-vector encoding, we prove an isomorphism to the Boolean hypercube $\{0,1\}^n$ with bitwise OR, so logical AND in rules becomes bitwise OR in the encoding. This yields a principled quotient-space formulation that groups functionally equivalent rules and guides structure-aware search. On real clinical data and synthetic benchmarks, quotient-space-aware genetic algorithms recover the global optimum in 48% to 77% of runs versus 35% to 37% for standard approaches, while maintaining diversity across equivalence classes. These results show that exposing and exploiting algebraic structure offers a simple, general route to more efficient combinatorial optimisation.

The paper considers a new quantitative-qualitative proximity measure for the features of information objects, where data enters a common information resource from several sources independently. The goal is to determine the possibility of their relation to the same physical object (observation object). The proposed measure accounts for the possibility of differences in individual feature values - both quantitative and qualitative - caused by existing determination errors. To analyze the proximity of quantitative feature values, the author employs a probabilistic measure; for qualitative features, a measure of possibility is used. The paper demonstrates the feasibility of the proposed measure by checking its compliance with the axioms required of any measure. Unlike many known measures, the proposed approach does not require feature value transformation to ensure comparability. The work also proposes several variants of measures to determine the proximity of information objects (IO) based on a group of diverse features.

Metacognition, understood as the monitoring and regulation of one's own cognitive processes, is inherently sequential: an agent evaluates an internal state, updates it, and may then re-evaluate under modified criteria. Order effects in cognition are well documented, yet it remains unclear whether such effects reflect classical state changes or reveal a deeper structural non-commutativity. We develop an operational framework that makes this distinction explicit. In our formulation, metacognitive evaluations are modeled as state-transforming operations acting on an internal state space with probabilistic readouts, thereby separating evaluation back-action from observable output. We show that order dependence prevents any faithful Boolean-commutative representation. We then address a stronger question: can observed order effects always be explained by enlarging the state space with classical latent variables? To formalize this issue, we introduce two assumptions, counterfactual definiteness and evaluation non-invasiveness, under which the existence of a joint distribution over all sequential readouts implies a family of testable constraints on pairwise sequential correlations. Violation of these constraints rules out any classical non-invasive account and certifies what we call genuine non-commutativity. We provide an explicit three-dimensional rotation model with fully worked numerical examples that exhibits such violations. We also outline a behavioral paradigm involving sequential confidence, error-likelihood, and feeling-of-knowing judgments following a perceptual decision, together with the corresponding empirical test. No claim is made regarding quantum physical substrates; the framework is purely operational and algebraic.

Recent advancements in Vision-Language Models (VLMs) have significantly pushed the boundaries of Visual Question Answering (VQA).However,high-resolution details can sometimes become noise that leads to hallucinations or reasoning errors. In this paper,we propose Degradation-Driven Prompting (DDP), a novel framework that improves VQA performance by strategically reducing image fidelity to force models to focus on essential structural information. We evaluate DDP across two distinct tasks. Physical attributes targets images prone to human misjudgment, where DDP employs a combination of 80p downsampling, structural visual aids (white background masks and orthometric lines), and In-Context Learning (ICL) to calibrate the model's focus. Perceptual phenomena addresses various machine-susceptible visual anomalies and illusions, including Visual Anomaly (VA), Color (CI), Motion(MI),Gestalt (GI), Geometric (GSI), and Visual Illusions (VI).For this task, DDP integrates a task-classification stage with specialized tools such as blur masks and contrast enhancement alongside downsampling. Our experimental results demonstrate that less is more: by intentionally degrading visual inputs and providing targeted structural prompts, DDP enables VLMs to bypass distracting textures and achieve superior reasoning accuracy on challenging visual benchmarks.

People often optimize for long-term goals in collaboration: A mentor or companion doesn't just answer questions, but also scaffolds learning, tracks progress, and prioritizes the other person's growth over immediate results. In contrast, current AI systems are fundamentally short-sighted collaborators - optimized for providing instant and complete responses, without ever saying no (unless for safety reasons). What are the consequences of this dynamic? Here, through a series of randomized controlled trials on human-AI interactions (N = 1,222), we provide causal evidence for two key consequences of AI assistance: reduced persistence and impairment of unassisted performance. Across a variety of tasks, including mathematical reasoning and reading comprehension, we find that although AI assistance improves performance in the short-term, people perform significantly worse without AI and are more likely to give up. Notably, these effects emerge after only brief interactions with AI (approximately 10 minutes). These findings are particularly concerning because persistence is foundational to skill acquisition and is one of the strongest predictors of long-term learning. We posit that persistence is reduced because AI conditions people to expect immediate answers, thereby denying them the experience of working through challenges on their own. These results suggest the need for AI model development to prioritize scaffolding long-term competence alongside immediate task completion.

Recent advances in multimodal large language models (MLLMs) have enabled impressive progress in vision-language understanding, yet their high computational cost limits deployment in resource-constrained scenarios such as personal assistants, document understanding, and smart cameras. Most existing methods rely on Transformer-based cross-attention, whose quadratic complexity hinders efficiency. Moreover, small vision-language models often struggle to precisely capture fine-grained, task-relevant visual regions, leading to degraded performance on fine-grained reasoning tasks that limit their effectiveness in the real world. To address these issues, we introduce Firebolt-VL, an efficient vision-language model that replaces the Transformer-based decoder with a Liquid Foundation Model (LFM) decoder. To further enhance visual grounding, we propose a Token-Grid Correlation Module, which computes lightweight correlations between text tokens and image patches and modulates via the state-space model with FiLM conditioning. This enables the model to selectively emphasize visual regions relevant to the textual prompt while maintaining linear-time inference. Experimental results across multiple benchmarks demonstrate that Firebolt-VL achieves accurate, fine-grained understanding with significantly improved efficiency. Our model and code are available at: https://fireboltvl.github.io

Diffusion models are promising for sparse-view novel view synthesis (NVS), as they can generate pseudo-ground-truth views to aid 3D reconstruction pipelines like 3D Gaussian Splatting (3DGS). However, these synthesized images often contain photometric and geometric inconsistencies, and their direct use for supervision can impair reconstruction. To address this, we propose Partial-Reference Image Quality Assessment (PR-IQA), a framework that evaluates diffusion-generated views using reference images from different poses, eliminating the need for ground truth. PR-IQA first computes a geometrically consistent partial quality map in overlapping regions. It then performs quality completion to inpaint this partial map into a dense, full-image map. This completion is achieved via a cross-attention mechanism that incorporates reference-view context, ensuring cross-view consistency and enabling thorough quality assessment. When integrated into a diffusion-augmented 3DGS pipeline, PR-IQA restricts supervision to high-confidence regions identified by its quality maps. Experiments demonstrate that PR-IQA outperforms existing IQA methods, achieving full-reference-level accuracy without ground-truth supervision. Thus, our quality-aware 3DGS approach more effectively filters inconsistencies, producing superior 3D reconstructions and NVS results. The project page is available at https://kakaomacao.github.io/pr-iqa-project-page/.

Developers are publishing AI agent skills that replicate a colleague's communication style, encode a supervisor's mentoring heuristics, or preserve a person's behavioral repertoire beyond biological death. To explain why, we propose Gradual Cognitive Externalization (GCE), a framework arguing that ambient AI systems, through sustained causal coupling with users, transition from modeling cognitive functions to constituting part of users' cognitive architectures. GCE adopts an explicit functionalist commitment: cognitive functions are individuated by their causal-functional roles, not by substrate. The framework rests on the behavioral manifold hypothesis and a central falsifiable assumption, the no behaviorally invisible residual (NBIR) hypothesis: for any cognitive function whose behavioral output lies on a learnable manifold, no behaviorally invisible component is necessary for that function's operation. We document evidence from deployed AI systems showing that externalization preconditions are already observable, formalize three criteria separating cognitive integration from tool use (bidirectional adaptation, functional equivalence, causal coupling), and derive five testable predictions with theory-constrained thresholds.

The evaluation of general-purpose artificial agents, particularly those based on large language models, presents a significant challenge due to the non-transitive nature of their interactions. When agent A defeats B, B defeats C, and C defeats A, traditional ranking methods that force a linear ordering can be misleading and unstable. We argue that for such cyclic domains, the fundamental object of evaluation should not be a ranking but a set-valued core, as conceptualized in classical tournament theory. This paper introduces Soft Tournament Equilibrium (STE), a differentiable framework for learning and computing set-valued tournament solutions directly from pairwise comparison data. STE first learns a probabilistic tournament model, potentially conditioned on rich contextual information. It then employs novel, differentiable operators for soft reachability and soft covering to compute continuous analogues of two seminal tournament solutions: the Top Cycle and the Uncovered Set. The output is a set of core agents, each with a calibrated membership score, providing a nuanced and robust assessment of agent capabilities. We develop the theoretical foundation for STE to prove its consistency with classical solutions in the zero-temperature limit, which establishes its Condorcet-inclusion properties, and analyzing its stability and sample complexity. We specify an experimental protocol for validating STE on both synthetic and real-world benchmarks. This work aims to provide a complete, standalone treatise that re-centers general-agent evaluation on a more appropriate and robust theoretical foundation, moving from unstable rankings to stable, set-valued equilibria.

Electronic Patient Record (EPR) systems contain valuable clinical information, but much of it is trapped in unstructured text, limiting its use for research and decision-making. Large language models can extract such information but require substantial computational resources to run locally, and sending sensitive clinical data to cloud-based services, even when deidentified, raises significant patient privacy concerns. In this study, we develop a resource-efficient semi-automated annotation workflow using small language models (SLMs) to extract structured information from unstructured EPR data, focusing on paediatric histopathology reports. As a proof-of-concept, we apply the workflow to paediatric renal biopsy reports, a domain chosen for its constrained diagnostic scope and well-defined underlying biology. We develop the workflow iteratively with clinical oversight across three meetings, manually annotating 400 reports from a dataset of 2,111 at Great Ormond Street Hospital as a gold standard, while developing an automated information extraction approach using SLMs. We frame extraction as a Question-Answering task grounded by clinician-guided entity guidelines and few-shot examples, evaluating five instruction-tuned SLMs with a disagreement modelling framework to prioritise reports for clinical review. Gemma 2 2B achieves the highest accuracy at 84.3%, outperforming off-the-shelf models including spaCy (74.3%), BioBERT-SQuAD (62.3%), RoBERTa-SQuAD (59.7%), and GLiNER (60.2%). Entity guidelines improved performance by 7-19% over the zero-shot baseline, and few-shot examples by 6-38%, though their benefits do not compound when combined. These results demonstrate that SLMs can extract structured information from specialised clinical domains on CPU-only infrastructure with minimal clinician involvement. Our code is available at https://github.com/gosh-dre/nlp_renal_biopsy.

Knowledge distillation compresses large teachers into smaller students, but performance saturates at a loss floor that persists across training methods and objectives. We argue this floor is geometric: neural networks represent far more features than dimensions through superposition, and a student of width $d_S$ can encode at most $d_S \cdot g(α)$ features, where $g(α) = 1/((1-α)\ln\frac{1}{1-α})$ is a sparsity-dependent capacity function. Features beyond this budget are permanently lost, yielding an importance-weighted loss floor. We validate on a toy model (48 configurations, median accuracy >93%) and on Pythia-410M, where sparse autoencoders measure $F \approx 28{,}700$ features at $α\approx 0.992$ (critical width $d_S^* \approx 1{,}065$). Distillation into five student widths confirms the predicted monotonic floor ordering. The observed floor decomposes into a geometric component and a width-independent architectural baseline ($R^2 = 0.993$). Linear probing shows coarse concepts survive even 88% feature loss, revealing the floor arises from aggregate loss of fine-grained features in the importance distribution's long tail. Our results connect representation geometry to distillation limits and provide a practical tool for predicting distillation performance from SAE measurements alone.

Context-dependent sequential decision making is commonly addressed either by providing context explicitly as an input or by increasing recurrent memory so that contextual information can be represented internally. We study a third alternative: realizing contextual dependence by intervening on a shared recurrent latent state, without enlarging recurrent dimensionality. To this end, we introduce an intervention-based recurrent architecture in which a recurrent core first constructs a shared pre-intervention latent state, and context then acts through an additive, context-indexed operator. We evaluate this idea on a context-switching sequential decision task under partial observability. We compare three model families: a label-assisted baseline with direct context access, a memory baseline with enlarged recurrent state, and the proposed intervention model, which uses no direct context input to the recurrent core and no memory growth. On the main benchmark, the intervention model performs strongly without additional recurrent dimensions. We also evaluate the models using the conditional mutual information (I(C;O | S)) as a theorem-motivated operational probe of contextual dependence at fixed latent state. For task-relevant phase-1 outcomes, the intervention model exhibits positive conditional contextual information. Together, these results suggest that intervention on a shared recurrent state provides a viable alternative to recurrent memory growth for contextual control in this setting.

Forecasting startup success is notoriously difficult, partly because meaningful outcomes, such as exits, large funding rounds, and sustained revenue growth, are rare and can take years to materialize. As a result, signals are sparse and evaluation cycles are slow. Y Combinator batches offer a unique mitigation: each batch comprises around 200 startups, funded simultaneously, with evaluation at Demo Day only three months later. We introduce YC Bench, a live benchmark for forecasting early outperformance within YC batches. Using the YC W26 batch as a case study (196 startups), we measure outperformance with a Pre-Demo Day Score, a KPI combining publicly available traction signals and web visibility. This short-term metric enables rapid evaluation of forecasting models. As a baseline, we take Google mentions prior to the YC W26 application deadline, a simple proxy for prior brand recognition, recovering 6 of 11 top performers at YC Demo Day (55% recall). YC Bench provides a live benchmark for studying startup success forecasting, with iteration cycles measured in months rather than years. Code and Data are available on GitHub: https://github.com/benstaf/ycbench

While textual frequency has been validated as relevant to human cognition in reading speed, its relatedness to Large Language Models (LLMs) is seldom studied. We propose a novel research direction in terms of textual data frequency, which is an understudied topic, to the best of our knowledge. Our framework is composed of three units. First, this paper proposes Textual Frequency Law (TFL), which indicates that frequent textual data should be preferred for LLMs for both prompting and fine-tuning. Since many LLMs are closed-source in their training data, we propose using online resources to estimate the sentence-level frequency. We then utilize an input paraphraser to paraphrase the input into a more frequent textual expression. Next, we propose Textual Frequency Distillation (TFD) by querying LLMs to conduct story completion by further extending the sentences in the datasets, and the resulting corpora are used to adjust the initial estimation. Finally, we propose Curriculum Textual Frequency Training (CTFT) that fine-tunes LLMs in an increasing order of sentence-level frequency. Experiments are conducted on our curated dataset Textual Frequency Paired Dataset (TFPD) on math reasoning, machine translation, commonsense reasoning and agentic tool calling. Results show the effectiveness of our framework.

We introduce ThinkTwice, a simple two-phase framework that jointly optimizes LLMs to solve reasoning problems and refine the answers, based on Group Relative Policy Optimization (GRPO). In each pair of training steps, ThinkTwice first optimizes the model on solving reasoning problems, then optimizes it on refining its own solutions to the same problems, using the same binary correctness reward in both phases without correctness signals or critique annotations. Across five mathematical reasoning benchmarks and two model families including Qwen3-4B and Olmo3-7B, ThinkTwice substantially improves both reasoning and refinement performance over competitive online policy optimization baselines. Specifically, on Qwen3-4B, ThinkTwice outperforms GRPO on AIME by 5 percentage points before refinement and by 11.5 points after one self-refinement step, measured by pass@4. Analysis of the training dynamics of ThinkTwice reveals an implicit rectify-then-fortify curriculum: refinement predominantly corrects errors early in training and naturally shifts toward preserving already-correct solutions as the model improves, yielding a more rectified reward signal. Our work establishes joint training of reasoning and self-refinement as a principled and effective methodology for RLVR.