Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
The paper introduces OSR, a method for removing labels from classification models by redistributing output space statistics to approximate retrained confidence vectors. It functions as a modular output filter, avoiding feature‑space adjustments and heavy retraining. Experiments show it matches full retraining performance while improving computational efficiency and privacy.
The paper introduces SWIM, a task that frames student writing simulation as proficiency‑conditioned essay generation. It evaluates prompting, supervised fine‑tuning, and reinforcement learning for aligning generated essays with student proficiency profiles, using automated essay scoring as a metric. Results show that prompting alone offers limited control, while supervised fine‑tuning and reinforcement learning significantly improve alignment across content, lexical, grammatical, and organizational traits, though low‑proficiency writing remains difficult to replicate.
The paper investigates how Arabic dialects are represented in large language models and whether they can be steered at inference time. By analyzing neuron-level sparsity and vector steering, the authors find that only a small fraction of neurons encode dialect-specific features, while distributed activation directions are more effective for steering. Vector steering can induce dialectal output from both dialectal and MSA prompts, whereas neuron steering works only when the prompt is already dialectal.
MudraGen is a conditional diffusion framework that generates realistic RGB images of Samyukta Hasta mudras, the two‑hand gestures used in Bharatanatyam. It introduces geometry‑aware supervision through three objectives—Keypoint Loss, Joint Offset Loss, and Shape Consistency—to ensure anatomical validity, inter‑hand coordination, and cultural nuance. The model outperforms existing generative approaches in visual realism, anatomical correctness, and fine hand‑pose structure, enabling faithful reproduction of complex mudras.
DrainSinkhorn is a verifier‑gated active‑packing layer that improves batched entropic optimal transport (EOT) by eliminating finished problems from subsequent Sinkhorn updates. It combines candidate‑axis packing, a one‑sided screen, verifier‑gated retirement, and physical compaction, while keeping the EOT objective, per‑instance map, and stopping rule unchanged. The method achieves state‑of‑the‑art execution speedups—up to 4.11× faster on MetroPT‑3 and 3.80× on ImageNet‑32 feature couplings—across multiple backends and tolerance settings. whyItMatters":"The technique delivers significant runtime reductions for heterogeneous batched‑EOT workloads, enabling faster and more efficient optimal transport computations in practical machine‑learning pipelines."
The paper introduces a neural video compression approach that combines deformable temporal alignment with difference-aware spatial selective fusion. A Context-aware Temporal Alignment Module generates complementary temporal context, while a Difference-aware Spatial Selective Fusion module adaptively selects reliable temporal information and suppresses misalignment. Experiments demonstrate that this method improves rate-distortion performance compared to DCVC-DC.
The paper introduces CAPA, a Collaborative Agent Predictive Architecture designed to give large language model (LLM) agents situational awareness in online meetings. CAPA uses a Perceiver to update meeting state, a Predictor to forecast conversation flow, a Controller to decide speaking actions, and a Generator to phrase contributions. Evaluated on 137 AMI meetings, CAPA reduces the silence rate from 51.4% to 2.5%, doubles credited recovery, and maintains low hallucination, demonstrating that structured state tracking is key to effective delegation.
SimCast‑S2S is a generative latent‑diffusion framework designed for probabilistic subseasonal‑to‑seasonal precipitation forecasting. It tackles three key challenges: it uses a diffusion‑based generative pipeline for uncertainty quantification, operates in a compact latent space learned by VAEs for efficient large‑ensemble generation, and employs transfer learning with LoRA to overcome limited training data. On reanalysis data, it outperforms deep‑learning baselines and competes with or surpasses state‑of‑the‑art operational systems such as ECMWF‑S2S.
ParaBridge is a self‑distillation method that uses a temporary paralinguistic instruction scaffold during training to teach a speech language model when non‑lexical cues should influence dialogue responses. By providing dense, full‑vocabulary next‑token targets from the scaffolded view while the scaffold‑free model generates its own replies, ParaBridge stabilizes inference‑time behavior without requiring curated dialogues or external reward models. Experiments on Qwen3‑Omni show significant gains on safety and empathy benchmarks while preserving general performance across multiple tests.
Gaokerena is a family of compact Persian medical language models designed for consumer‑grade hardware. The first variant, Gaokerena‑V, was trained on a 90‑million‑token Persian medical corpus and 20,000 physician Q&A pairs, raising a translated medical MMLU benchmark score from 46.28% to 49.31%. A second variant, Gaokerena‑R, adds a Chain‑of‑Thought approach and two Reinforcement Learning with AI Feedback frameworks, achieving a higher benchmark score of 52.98% while also providing uncertainty estimates based on internal hidden states.
LeanStream introduces a speculate‑and‑refine streaming framework that enables efficient on‑device inference of large language models by progressively refining computation, loading, and cache‑retention priorities using partial GPU results. This approach allows fine‑grained overlap between GPU execution and storage I/O, avoiding the trade‑offs of existing systems that serialize execution or incur redundant weight fetches. Implemented on mobile and embedded platforms, LeanStream reduces memory usage by 4.8× to 7.5× compared to prior work while improving token generation throughput by 1.6× to 2.1×.
The paper introduces STAIR, a retrieval system that uses a document’s Table of Contents to guide large language models in accessing global structure, thereby reducing hallucinations in Retrieval Augmented Generation. Experiments with a fine‑tuned Differentiable Search Index show that ToC‑based retrieval yields a low hallucination rate (<0.05%) and improves Recall@1 to 82.6% on the newly released SearchTome benchmark, outperforming baselines like BM25, DPR, and Mistral. The authors also release SearchTome, a diverse dataset of 18 books across six domains, to encourage further research in ToC‑based retrieval.
The Dice Roll Method is a standardized protocol for auditing large language model brand recommendations through repeated queries. It decomposes total response variance into sampling, prompt‑phrasing, run‑to‑run, and model‑version components, and uses a negative‑binomial mixed model, Cliff’s delta, and bootstrap techniques to guide iteration counts. The study identifies three iteration tiers—exploratory (n=5), confirmatory (n=10), and rigorous (n=15)—and recommends a compact battery of four complementary metrics for robust evaluation.
SV-WAM is a surround‑view world‑action model that keeps all six camera views while enabling efficient inference by discarding the video branch at deployment. It uses future‑video prediction as dense training supervision and an action‑centered causal mask to prevent action tokens from attending to future‑video tokens during joint denoising. A differentiable drivable‑area compliance regularizer penalizes vehicle‑footprint corners near or crossing drivable boundaries, improving safety and boundary awareness. Experiments on NAVSIMv2 and nuScenes show state‑of‑the‑art planning performance with low latency and strong zero‑shot transfer.
The paper proposes a test‑time method to enhance the faithfulness of large language model (LLM) explanations by removing concepts that are not credited in the model’s explanation before re‑querying the model. This approach directly addresses the incompleteness dimension of unfaithful explanations, unlike prior methods that mainly target unsoundness. Experiments across two datasets, multiple model families, and two faithfulness metrics show that the method improves explanation faithfulness over standard prompting and faithfulness‑encouraging prompts, and it is model‑agnostic and requires no parameter changes.
FLY-EVAL++ is an evidence-driven evaluation protocol designed for safety-constrained flight prediction with large language models. It combines deterministic verification of protocol compliance, physical feasibility, and safety constraints, then aggregates results into interpretable multi-dimensional scores. Applied to Flight Trajectory and Attitude Prediction, the protocol reveals that safety compliance is the most discriminative metric, with models of similar predictive accuracy differing by over 28 points in safety score and exhibiting recurrent safety violations and instability in multi-step rollouts.
The paper presents a differentiable hybrid modelling framework that combines a JAX finite volume population balance solver with neural network components to learn constitutive laws and initial conditions directly from experimental data. This approach addresses biases from hand‑picked models and the limitations of black‑box surrogates, enabling more accurate transport predictions. The framework’s differentiability also facilitates optimisation of experimental settings for desired process outcomes.
The paper introduces Test Cases Scaling (TCS), a two‑stage reinforcement learning framework that automatically generates sound and discriminative test cases for code‑generation large language models. Stage 1 trains a test generator to produce tests aligned with a reference solution, while Stage 2 refines the generator to produce counterexample tests that target the model’s current failure modes. Experiments on TACO and LiveCodeBench show that TCS improves pass@1 and inference‑time answer selection, and the learned generator also aids in selecting better outputs from other LLMs.
The paper investigates how to better detect factual errors, or hallucinations, in long-form medical chatbot responses. It introduces a multi‑perspective annotation workflow that combines first‑pass labeling, a large language model acting as a judge (LaJ) to surface candidate errors, and two adjudication steps—expert medical review and evidence‑based fact‑checking. The study finds that single‑pass benchmarks miss many errors, that LaJ alone is insufficient, and that adjudicators disagree, indicating that multi‑pass adjudication improves coverage but still depends on human judgment and evidence.
The paper introduces CAPA, a Collaborative Agent Predictive Architecture designed to improve large language model (LLM) participation in online meetings. CAPA tracks meeting state with a Perceiver, predicts conversation flow, decides when and what to speak, and generates contributions in the participant’s style, all while being calibrated by judges. In experiments on 137 AMI meetings, CAPA cuts the LLM’s silence rate from 51.4% to 2.5%, doubles credited recovery, and maintains low hallucination.