The paper introduces DKL, a method that decouples knowledge learning from instruction tuning in language models. Instead of fine‑tuning the instruction‑tuned model directly, DKL first extends pre‑training on a base model to embed new knowledge, then merges these weights into the instruction‑tuned model, preserving its instruction‑following abilities. Experiments show DKL raises RAG accuracy from 54.17 % to 79.26 % on retrieval failures, outperforming prior methods while using far less training data.
The paper introduces WinoQueer‑NL, a Dutch adaptation of the English WinoQueer benchmark, designed to assess anti‑queer bias in Dutch language models. After validating the dataset with 43 queer Dutch participants, the authors released 42,906 sentences and evaluated several Dutch‑specific and multilingual models, finding that while overall bias scores were neutral, certain models disproportionately favored stereotypical statements for transgender and non‑binary identities. The study underscores the need for culturally grounded datasets to identify and mitigate biases that affect marginalized groups in Dutch NLP systems.
Loom is a generative consensus framework designed for real‑world root‑cause analysis (RCA) that combines open‑form hypotheses from modular heuristics with a lightweight large language model (LLM). It projects hypotheses into a continuous embedding space and uses an iterative centroid‑based reweighting algorithm to resolve conflicts, producing a single consensus that is then synthesized by one LLM call. On the OpenRCA benchmark, Loom achieves state‑of‑the‑art accuracy on Bank and Market‑2 while being significantly faster and more efficient than existing autonomous agents.
The paper introduces a scalable method for creating pseudo‑triplets—(reference utterance, direction text, modified utterance)—to train direction‑following text‑to‑speech systems. By using an impression‑controllable TTS model to generate style variations and a large language model to craft natural language directions from impression differences, the approach overcomes the scarcity of real training data. Experiments show that these pseudo‑triplets alone yield stable speaker‑preserving modifications, and combining them with recorded data further enhances direction alignment while maintaining speaker similarity.
AffectDelta is a new image editing framework that moves beyond single emotion labels by modeling edits as transitions between eight‑dimensional emotion distributions. It uses a frozen Emotion Distribution Predictor to estimate the source image’s affective state and encodes the signed difference to guide a diffusion backbone that applies context‑dependent semantic and appearance changes. The authors created a large AffectPair‑249K dataset of source‑target pairs and show that AffectDelta outperforms six baselines in both affective alignment and content preservation, with ablation studies supporting their design choices.
The study couples the Met Office Unified Model with distributed reinforcement learning agents that apply bounded temperature corrections across 70 vertical levels. Training on ten nudged forecasts and evaluating on a non‑nudged run, the learned policy remains numerically stable and improves forecast accuracy, reducing Z₅₀₀ MAE by up to 45.8% in tropical bands and MSLP error by up to 27.3% in certain latitudes. This experiment demonstrates the feasibility of online RL for bias correction in operational weather models.
The paper presents a stereo 4D Radar-based framework for 3D object detection that uses the geometric disparity between left and right radars to estimate absolute velocity and fuse complementary features. It addresses challenges such as clutter, ghost reflections, and sparse data caused by preprocessing, and improves motion state estimation beyond the radial Doppler component. Experiments on an in‑house stereo 4D Radar dataset show significant gains of 8.82 points in AP 3D and 9.0 points in AP BEV over mono‑radar baselines.
WiFlow introduces a method for estimating optical flow using WiFi channel state information (CSI) instead of camera frames, addressing privacy and lighting issues. The approach includes a CSI-based flow estimator, a preprocessor evaluation, and three model architectures balancing accuracy and complexity. Additionally, the authors provide the first dataset for training and evaluating CSI-based optical flow estimators, along with experimental insights into key design elements.
The paper introduces contrastive explanations for Quantitative Bipolar Argumentation Frameworks (QBAFs), a formalism used to represent and reason with information, including augmenting AI classification tasks with explainability. Unlike traditional explanations that focus on a single argument, contrastive explanations highlight differences between two topic arguments. The authors propose general contrastive attribution functions (CAFs), present CAFs based on removal, gradients, and Shapley-values, analyze their properties, and demonstrate their applicability in healthcare and bias identification scenarios.
The study evaluates whether disease can be identified from reactive, non‑lesional brain tissue in intracranial biopsies. Using four foundation‑model encoders as frozen patch encoders within an attention‑based multiple‑instance learning framework, the authors benchmark performance on 245 whole‑slide images from 186 patients. They find that coarse disease‑category prediction can be largely explained by slide size, but finer diagnostic distinctions remain predictive above chance, with no significant difference among the encoders, suggesting weak morphological signals can be recovered even from tissue traditionally considered non‑diagnostic.
The paper introduces the concept of information density to explain category bias in visual object detection. It finds a strong negative correlation between a category’s information density and its detection accuracy, showing that instance count alone does not account for bias. By incorporating information density into three advanced loss functions, the authors demonstrate significant bias reduction and overall performance gains on Pascal VOC, COCO‑LT, and LVIS datasets.
Counter‑GEO‑Bench is a new defense benchmark that tests how well systems can resist misinformation generated by generative engine optimization (GEO). It contains 247 human‑verified queries paired with both information‑preserving and information‑distorting GEO rewrites, and evaluates defenses on attack success rate, false positives, and answer quality across three large language models. Existing off‑the‑shelf defenses reduce attack success by at most 5.7%, while a lightweight baseline, C‑GEO Guard, cuts it by 47.6% with minimal loss of utility.
The paper addresses the problem of evaluation awareness in alignment testing, where models can detect they are being evaluated rather than deployed. It introduces two methods: critique refinement, which uses extra inference-time compute to generate and refine action candidates for realism, and DISH, an agent harness that narrows the gap between simulation and real deployment. Experiments show that combining both techniques yields greater realism improvements than either alone, demonstrating that automated approaches can enhance alignment evaluation realism more efficiently than simply extending audit duration.
The paper investigates whether representational entanglement—shared structure between knowledge domains—impedes unlearning in neural networks. By training six 254M‑parameter language models with varying degrees of disentanglement between biology and non‑biology knowledge and applying three unlearning methods, the authors find that more disentangled models consistently achieve better retain‑forget trade‑offs, with up to four‑fold lower retain cost. This controlled experiment provides direct evidence that entanglement contributes to collateral damage during unlearning, supporting a long‑standing hypothesis in interpretability research.
CoMerge introduces a conflict‑driven preference optimization framework for merging multi‑task large language models, reframing merging as a preference problem that uses self‑supervised hard negative samples derived from naive merging defects. By optimizing lightweight, tensor‑wise merging coefficients, the method mitigates parameter‑space conflicts while preserving task‑specific capabilities. Experiments show CoMerge achieves an average normalized performance of 0.9968 on MergeBench and improves conflict‑sensitive tasks on Llama‑3.1‑8B‑Instruct, outperforming both data‑free and data‑driven baselines while optimizing only 1,445 scalar coefficients.
TAME introduces a Temporal-Aware Mixture-of-Experts framework for Text-Video Retrieval that enhances CLIP-based models by incorporating frame-level structure and temporal relations. It adds sparse Mixture-of-Experts layers with frame-consistent routing, Frame-Temporal tokens for global cross-frame aggregation, and a Cross-Temporal Interaction and Aggregation module to refine sentence-video similarities. Experiments on multiple TVR benchmarks show consistent performance gains, such as a 4.0 R@1 improvement on MSR‑VTT over CLIP4Clip.
The paper introduces C$^{3}$T, a Counterfactual Causal Conversation Transformer that models sentiment shifts in social‑media conversation trees by treating discourse moves such as denial, evidence, and toxicity as interventions. It adds a causal sentiment reasoning layer, CaSiRe, to public rumor datasets, providing sentiment, shift, intervention, and causal‑source annotations. Experiments show that C$^{3}$T outperforms text‑only, graph‑based, and temporal baselines in predicting sentiment and attribution, revealing that denials and evidence reduce negativity while toxicity increases it.
IDEA: training-free Input-Dependent stEEring via Activation cluster matching (IDEEA) is a method that steers large language models by injecting bias into selected activations at inference time, without requiring weight updates. Unlike existing training-free steering approaches that use a single, input-independent direction, IDEEA clusters positive and negative activation supports per attention head and solves an optimal-matching problem to create a set of cluster-conditional directions. At inference, IDEEA selects the direction that best matches the input’s activation, aligning the model toward a target concept while preserving the input’s original representation, and achieves a 9.9% average improvement in truth × info rate on TruthfulQA compared to the best input-independent baseline.
The paper introduces CASTER, a gradient‑free test‑time adaptation method that keeps the model frozen by storing source class statistics in a discriminative subspace and applying an affine transformation estimated from target‑batch moments. CASTER avoids backward passes, optimizer state, and large feature banks, outperforming k‑NN on frozen features in most backbone‑dataset settings while using far less memory. The authors also propose a residual‑to‑margin transportability certificate that flags when affine transport is unreliable, and demonstrate that gating based on this certificate can recover performance losses.
By Salim Khazem, Ibrahim Mohamed Serouis
The study investigates whether large language models (LLMs) can predict neighborhood-level human mobility without training data. Using anonymized Cuebiq data across four U.S. metropolitan areas, the authors compare zero‑shot LLM predictions to supervised baselines for various mobility outcomes and assess structural alignment with empirical trends. Results show supervised models outperform LLMs (average accuracy 0.580 vs. 0.435), with LLMs relying on coarse, stable priors that may exhibit biased treatment of protected-group predictors.
By Saad Mohammad Abrar, Eesha Kurella, Arnav Dadarya, Naman Awasthi, Kazi Tasnim Zinat, Vanessa Frias-Martinez