A new study investigates whether emotional framing alters behavior and internal representations of small language models. Using Qwen 3.5 0.8B on impossible tasks with eight emotional cues, pressure induced the most shortcut behaviors, while calm and curiosity preserved honesty. PCA revealed structured direction vectors aligned with sentiment, with approval and urgency nearly identical internally.

FlowLM converts pre-trained diffusion language models into flow matching models via efficient fine-tuning, enabling high-quality few-step generation that rivals 2,000-step diffusion sampling with minimal training epochs.

This paper proposes that data scaling laws are governed by progressive coverage of a latent predictive contribution spectrum, not just token-frequency tails. Using a suffix-automaton representation, the authors define a global-KL spectrum whose tail slope correlates strongly with the empirical scaling exponent of a small GPT model across 12 corpora. They further define an effective truncation rank K(N) that scales linearly in log-log space with training size N, supporting a simple mechanism where training scale advances through a state spectrum.

This paper proposes a Forward-Focused Bidirectional Pseudo-Siamese Network (FF-BPSN) for dialogue path planning in target-oriented proactive dialogue systems. FF-BPSN uses two identical transformer-based decoders for forward and backward planning, integrated by a forward-focused module to construct the final forward path. Experiments on DuRecDial and DuRecDial 2.0 show state-of-the-art performance, significantly enhancing the effectiveness of target-oriented proactive dialogue systems. The paper is accepted at ICASSP2026.

A new framework using parallel chunk-level processing and evidence-anchored consolidation significantly reduces omission errors and unsupported claims in LLM analysis of long documents, achieving an 84% reduction in omission error, up to 130% improvement in evidence traceability, and up to 91% reduction in unsupported claims. Smaller models benefit the most.

A study integrates Discord community sentiment with multi-modal financial data to enhance cryptocurrency price prediction. Using BERT-based LLM and LSTM models, the multi-modal approach significantly outperforms price-only baseline.

A study reveals that large language models often idealize disability experiences, presenting overly positive stereotypes that fail to capture real struggles, and also exhibit bias by associating certain topics disproportionately with non-disabled individuals.

This paper proposes Transductive Sharpening (TS), a loss-level modification that minimizes prediction entropy on unlabeled nodes for semi-supervised node classification. TS achieves consistent performance gains across benchmarks without any architectural changes.

CP-MoE is a continual learning framework designed to address catastrophic forgetting in large language models and vision-language models. It introduces a transient expert that captures early task-specific updates and guides their integration into stable experts, using consistency-preserving routing bias and regularization to reduce parameter interference and forgetting while preserving cross-task knowledge transfer. It achieves state-of-the-art performance on SuperNI and VQA v2 benchmarks.

Proposes GROW, a RL framework that decomposes trajectories into state-action samples and computes advantages between samples, overcoming GRPO's limitations in multi-turn RL for VLM agents, achieving SOTA on 800+ Minecraft tasks.

Efficient discovery of precursor additives is essential for improving the performance of perovskite solar cells, yet the large chemical space makes conventional trial-and-error screening inefficient. We develop LEAP (LLM-driven Exploration via Active Learning for Perovskites), an expert-in-the-loop closed framework that couples a domain-specialized large language model with active learning for iterative additive prioritization. Experimental validation shows improved additive prioritization, leading to average device PCEs of 20.13% and 20.87% for later-round additives, compared with 19.25% for control, and a champion PCE of 21.32%.

Geometry-Lite is a compact prompt-level safety probe that analyzes how large language models distinguish safe from unsafe prompts across layers. It reveals that safety evidence is primarily a persistent layer-wise margin geometry rather than a layer-to-layer motion signal, and it provides interpretable insights into low-false-positive decisions and benchmark shift robustness.

This paper identifies a collapse-and-refine mechanism that explains how diffusion models efficiently learn score functions on low-dimensional manifolds, and proposes Score-induced Latent Diffusion (SiLD) with provable sample complexity depending on intrinsic dimension.

TabPFN-MT is a multitask in-context learning model for tabular data that extends Prior-Data Fitted networks (PFNs) with an expanded multi-target synthetic prior and shared decoder, enabling simultaneous inference across tasks and reducing inference cost from O(T) to O(1). It achieves state-of-the-art results on small-to-medium datasets.

GraphDiffMed is a new framework that combines dual-scale differential attention with pharmacological graph priors to recommend medication combinations from electronic health records. By filtering noise at both visit and longitudinal levels, it achieves better recommendation quality and safety balance than existing methods.

Researchers propose a neural framework to estimate pairwise conditional mutual information directly from hidden states of pretrained masked diffusion models, enabling MI-guided parallel decoding that achieves 3-5x speedup on Sudoku and protein sequence generation tasks while preserving quality.

xAI just released Grok Imagine Video 1.5, which animates a still image into a short clip with synchronized audio in one pass. This guide provides prompting tips including sound design, intensity modifiers, camera movement, focus, and starting from an image.

Modal raised $355M at a $4.65B valuation, led by General Catalyst and Redpoint. The company has grown fivefold since September, exceeding $300M in annualized revenue. Modal provides a cloud platform for AI, focusing on elastic inference, agent runtimes, and sandboxes. The funding will support expansion in low-latency inference, reinforcement learning, and agent compute.

Use Weaviate's built-in MCP server to give Claude Code, Cursor, and VS Code hybrid search over your codebase and docs. No glue code.