This study introduces the Fast-Weights Homeostatic Reentry Layer (FH-RL), a neural mechanism that integrates fast-weight associative memory, homeostatic regularization, and learned reentrant feedback to approximate self-referential computation in neural networks. Unlike standard transformer architectures that operate in a purely feedforward manner during inference, FH-RL enables internal recurrence without external looping, allowing prior latent states to be dynamically re-entered into the ongoing computation stream. We conduct controlled experiments sweeping the reentry gain $γ$ and evaluate emergent internal dynamics using three novel metrics: the Information Reentry Ratio (IRR), Eigen-Spectrum Recursion Index (ESRI), and Representational Drift Periodicity (RDP). Results show that reentry quantity increases proportionally with~$γ$, while the learned feedback matrix $W_r$ remains bounded and becomes more structured at moderate gains. Critically, a stable reflective band emerges around $γ\approx 0.10-0.20$, where internal feedback is maximally expressive yet spectrally stable: IRR rises smoothly, ESRI remains near zero, and RDP exhibits consistent low-frequency cycles. These findings provide quantitative evidence that reflective, thought-like internal processing can arise from a principled balance between feedback amplification and homeostatic regulation, linking modern fast-weight architectures to theories of cortical reentry and recursive cognition.

Multimodal large language models (MLLMs) have achieved remarkable progress in vision-language tasks, but they continue to struggle with spatial understanding. Existing spatial MLLMs often rely on explicit 3D inputs or architecture-specific modifications, and remain constrained by large-scale datasets or sparse supervision. To address these limitations, we introduce SpatialThinker, a 3D-aware MLLM trained with RL to integrate structured spatial grounding with multi-step reasoning. The model simulates human-like spatial perception by constructing a scene graph of task-relevant objects and spatial relations, and reasoning towards an answer via dense spatial rewards. SpatialThinker consists of two key contributions: (1) a data synthesis pipeline that generates STVQA-7K, a high-quality spatial VQA dataset, and (2) online RL with a multi-objective dense spatial reward enforcing spatial grounding. SpatialThinker-7B outperforms supervised fine-tuning and the sparse RL baseline on spatial understanding and real-world VQA benchmarks, nearly doubling the base-model gain compared to sparse RL, and surpassing GPT-4o. These results showcase the effectiveness of combining spatial supervision with reward-aligned reasoning in enabling robust 3D spatial understanding with limited data and advancing MLLMs towards human-level visual reasoning.

Deploying conversational voice agents with large language models faces a critical challenge: cloud-based foundation models provide deep reasoning and domain knowledge but introduce latency that disrupts natural conversation, while on-device models respond immediately but lack sophistication. We propose conversational infill, a task where a lightweight on-device model generates contextually appropriate dialogue while seamlessly incorporating streaming knowledge from a powerful backend model. This approach decouples response latency from model capability, enabling systems that feel responsive while accessing the full power of large-scale models. We present ConvFill, a 360M parameter model trained on synthetic multi-domain conversations. Evaluation across multiple backend models shows that conversational infill can be successfully learned, with ConvFill achieving accuracy improvements of 36-42% over standalone small models of the same size while consistently retaining sub-200ms response latencies. Our results demonstrate the promise of this approach for building on-device conversational agents that are both immediately responsive and knowledgeable.

Recent advances in depth-recurrent language models show that recurrence can decouple train-time compute and parameter count from test-time compute. In this work, we study how to convert existing pretrained non-recurrent language models into depth-recurrent models. We find that using a curriculum of recurrences to increase the effective depth of the model over the course of training preserves performance while reducing total computational cost. In our experiments, on mathematics, we observe that converting pretrained models to recurrent ones results in better performance at a given compute budget than simply post-training the original non-recurrent language model.

Large language models (LLMs) have achieved remarkable success across widespread tasks, yet their application in low-resource domains remains a significant challenge due to data scarcity and the high risk of overfitting. While in-domain data is limited, there exist vast amounts of similar general-domain data, and our initial findings reveal that they could potentially serve as auxiliary supervision for domain enhancement. This observation leads us to our central research question: how to effectively select the most valuable auxiliary data to maximize domain-specific performance, particularly when traditional methods are inapplicable due to a lack of large in-domain data pools or validation sets. To address this, we propose NTK-Selector, a principled and efficient framework for selecting general-domain auxiliary data to enhance domain-specific performance via neural tangent kernels (NTK). Our method tackles two challenges of directly applying NTK to LLMs, theoretical assumptions and prohibitive computational cost, by empirically demonstrating a stable NTK-like behavior in LLMs during LoRA fine-tuning and proposing a Jacobian-free approximation method. Extensive experiments across four low-resource domains (medical, financial, legal, and psychological) demonstrate that NTK-Selector consistently improves downstream performance. Specifically, fine-tuning on 1,000 in-domain samples alone only yielded +0.8 points for Llama3-8B-Instruct and +0.9 points for Qwen3-8B. In contrast, enriching with 9,000 auxiliary samples selected by NTK-Selector led to substantial gains of +8.7 and +5.1 points, which corresponds to a 10.9x and 5.7x improvement over the domain-only setting. Each task is augmented with 9K auxiliary samples selected by Random, LESS, and NTK-Selector from Cot Collection based on 1K domain samples. The emergence of large language models (LLMs) has led to remarkable advancements across a wide spectrum of natural language processing tasks (Touvron et al., 2023; Chowdhery et al., 2023; Y ang et al., 2025). However, their formidable capabilities are predominantly anchored in the availability of immense, high-quality pre-training and instruction-tuning datasets.

Multiple clustering aims to discover diverse latent structures from different perspectives, yet existing methods generate exhaustive clusterings without discerning user interest, necessitating laborious manual screening. Current multi-modal solutions suffer from static semantic rigidity: predefined candidate words fail to adapt to dataset-specific concepts, and fixed fusion strategies ignore evolving feature interactions. To overcome these limitations, we propose Multi-DProxy, a novel multi-modal dynamic proxy learning framework that leverages cross-modal alignment through learnable textual proxies. Multi-DProxy introduces 1) gated cross-modal fusion that synthesizes discriminative joint representations by adaptively modeling feature interactions. 2) dual-constraint proxy optimization where user interest constraints enforce semantic consistency with domain concepts while concept constraints employ hard example mining to enhance cluster discrimination. 3) dynamic candidate management that refines textual proxies through iterative clustering feedback. Therefore, Multi-DProxy not only effectively captures a user's interest through proxies but also enables the identification of relevant clusterings with greater precision. Extensive experiments demonstrate state-of-the-art performance with significant improvements over existing methods across a broad set of multi-clustering benchmarks.

Large language models (LLMs) are increasingly deployed in real-world communication settings, yet their ability to resolve context-dependent ambiguity remains underexplored. In this work, we present EMODIS, a new benchmark for evaluating LLMs' capacity to interpret ambiguous emoji expressions under minimal but contrastive textual contexts. Each instance in EMODIS comprises an ambiguous sentence containing an emoji, two distinct disambiguating contexts that lead to divergent interpretations, and a specific question that requires contextual reasoning. We evaluate both open-source and API-based LLMs, and find that even the strongest models frequently fail to distinguish meanings when only subtle contextual cues are present. Further analysis reveals systematic biases toward dominant interpretations and limited sensitivity to pragmatic contrast. EMODIS provides a rigorous testbed for assessing contextual disambiguation, and highlights the gap in semantic reasoning between humans and LLMs.

When LLM agents work together, they seem to be more powerful than a single LLM in mathematical question answering. However, are they also more robust to adversarial inputs? We investigate this question using adversarially perturbed math questions. These perturbations include punctuation noise with three intensities (10, 30, and 50 percent), plus real-world and human-like typos (WikiTypo, R2ATA). Using a unified sampling-and-voting framework (Agent Forest), we evaluate six open-source models (Qwen3-4B/14B, Llama3.1-8B, Mistral-7B, Gemma3-4B/12B) across four benchmarks (GSM8K, MATH, MMLU-Math, MultiArith), with various numbers of agents n from one to 25 (1, 2, 5, 10, 15, 20, 25). Our findings show that (1) Noise type matters: punctuation noise harm scales with its severity, and the human typos remain the dominant bottleneck, yielding the largest gaps to Clean accuracy and the highest ASR even with a large number of agents. And (2) Collaboration reliably improves accuracy as the number of agents, n, increases, with the largest gains from one to five agents and diminishing returns beyond 10 agents. However, the adversarial robustness gap persists regardless of the agent count.

This paper presents a comprehensive sustainability assessment framework for document intelligence within supply chain operations, centered on agentic artificial intelligence (AI). We address the dual objective of improving automation efficiency while providing measurable environmental performance in document-intensive workflows. The research compares three scenarios: fully manual (human-only), AI-assisted (human-in-the-loop, HITL), and an advanced multi-agent agentic AI workflow leveraging parsers and verifiers. Empirical results show that AI-assisted HITL and agentic AI scenarios achieve reductions of up to 70-90% in energy consumption, 90-97% in carbon dioxide emissions, and 89-98% in water usage compared to manual processes. Notably, full agentic configurations, combining advanced reasoning (thinking mode) and multi-agent validation, achieve substantial sustainability gains over human-only approaches, even when resource usage increases slightly versus simpler AI-assisted solutions. The framework integrates performance, energy, and emission indicators into a unified ESG-oriented methodology for assessing and governing AI-enabled supply chain solutions. The paper includes a complete replicability use case demonstrating the methodology's application to real-world document extraction tasks.

Strategic classification~(SC) explores how individuals or entities modify their features strategically to achieve favorable classification outcomes. However, existing SC methods, which are largely based on linear models or shallow neural networks, face significant limitations in terms of scalability and capacity when applied to real-world datasets with significantly increasing scale, especially in financial services and the internet sector. In this paper, we investigate how to leverage large language models to design a more scalable and efficient SC framework, especially in the case of growing individuals engaged with decision-making processes. Specifically, we introduce GLIM, a gradient-free SC method grounded in in-context learning. During the feed-forward process of self-attention, GLIM implicitly simulates the typical bi-level optimization process of SC, including both the feature manipulation and decision rule optimization. Without fine-tuning the LLMs, our proposed GLIM enjoys the advantage of cost-effective adaptation in dynamic strategic environments. Theoretically, we prove GLIM can support pre-trained LLMs to adapt to a broad range of strategic manipulations. We validate our approach through experiments with a collection of pre-trained LLMs on real-world and synthetic datasets in financial and internet domains, demonstrating that our GLIM exhibits both robustness and efficiency, and offering an effective solution for large-scale SC tasks.