Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2609.38036v2 Announce Type: replace-cross Abstract: Understanding gender biases in large language models (LLMs) is increasingly important as these systems become embedded in decision-support to...
arXiv:2607.19889v3 Announce Type: replace Abstract: Recognizing instrument-tissue interactions is essential for context-aware surgical AI. Vision-language models offer a natural way to inject semanti...
The paper introduces BARRAC, a method that adapts an English aspect‑based sentiment analysis framework for Arabic dialect classification tasks. It replaces English consumer‑review attribute pools with Arabic linguistic markers for sentiment, sarcasm, and dialect identification, and swaps noisy self‑training for a two‑stage training process. Evaluated on five Arabic dialect datasets, BARRAC achieves a mean macro‑F1 of 63.93%, surpassing the best few‑label state‑of‑the‑art by 3% and outperforming GPT‑4o on four of the five tasks, while error analysis highlights remaining challenges.
LexReward is a taxonomy-driven reward framework designed for legal language models, evaluating responses across three dimensions: Style (lexical and syntactic quality), Element (legal subjects, facts, statutes, and decisions), and Chain (order, completeness, correctness, and non-redundancy of reasoning). The framework introduces rubrics that define evaluation criteria and quality levels for each dimension, generating pairwise preference data used for Direct Preference Optimization (DPO) and reward-model training. Experiments demonstrate that rubric-based rewards effectively differentiate legal responses of varying quality, and that DPO training improves performance across all dimensions; the resulting reward models (LexRM) enable reinforcement learning to enhance policy performance in each specific dimension without needing reference answers.
RAST is a resolution‑aware transfer framework that addresses the mismatch between high‑resolution (HR) audio available during training and low‑resolution (LR) audio used at inference for human activity recognition. By compressing HR teacher representations while preserving token‑level information and neighborhood structure, RAST performs localized HR‑LR alignment. Experiments on the SAMoSA and AudioIMU datasets show that RAST consistently outperforms LR‑only training and direct teacher‑transfer baselines, improving LR‑only recognition by up to approximately 7.8% while requiring only LR audio at inference.
The paper proposes probe-guided fine-tuning, a method that uses probes detecting undesired properties in model activations as a direct training signal. Experiments show that continuously updated probes reduce harmfulness and improve honesty while preserving utility, outperforming DPO and inference-time steering in safety-utility trade-offs and robustness to jailbreak and abliteration attacks. Importantly, the concepts remain linearly encoded after fine-tuning, maintaining monitorability.
arXiv:2609.34330v2 Announce Type: replace Abstract: Multimodal large language models (MLLMs) have demonstrated impressive performance in multimodal understanding, but processing large numbers of visu...
arXiv:2609.38630v1 Announce Type: new Abstract: Privacy redaction must remove personal information while preserving relationships expressed in text. We develop a multilingual named-entity tagger with...
HalluScoring 2026 is a shared task that evaluates hallucination detection and factual verification in Arabic question answering, focusing on generalization to unseen questions and LLMs. It comprises two main tasks with four subtasks: binary hallucination detection (Subtasks 1.1 and 1.2) and answer verification against six candidates in Islamic and general knowledge domains (Subtasks 2.1 and 2.2). Thirteen teams participated, with the top system achieving AUC‑ROC scores of 0.772 and 0.767 for detection, and 0.882 and 0.857 for verification.
The paper investigates regression when outcomes are observed only after passing through selection filters that depend on both covariates and outcomes, a common issue in fields such as clinical trials, labor markets, and auctions. It provides a complete characterization of the minimal assumptions on the functional forms of selection processes that allow regression to remain possible, and shows that the regression function can sometimes be identified even when the selection filter itself cannot. Under stronger identification conditions, the authors also deliver finite‑sample estimation guarantees, explicit convergence rates, and oracle‑efficient algorithms, offering the first general‑purpose estimation method for this broad class of selection problems.
MetaSteer is a new method for steering large language models that learns nonlinear, context-dependent interventions applied to attention projection matrices. Unlike traditional linear, context-independent techniques, MetaSteer adapts its effects based on the input, requiring no linear concept-geometry assumption. Trained once on a pooled preference corpus, it transfers zero‑shot to unseen concepts and out‑of‑distribution contexts, matching or surpassing strong task‑specific baselines on multiple benchmarks and model families.
Cogentic is a multi‑agent system designed to automate proof discovery for open research problems. It uses an iterative prove‑verify loop where an orchestrator assigns independent provers to different proof directions, verifies their outputs with specialized components, and records confirmed intermediate results in a persistent ledger for future rounds. Built on Gemini, Cogentic has produced novel results on five open problems in online learning, auction theory, and mechanism design, each verified by domain experts and detailed in companion papers.
arXiv:2307.15494v3 Announce Type: replace-cross Abstract: Hindsight Experience Replay (HER) enhances sample efficiency in goal-conditioned reinforcement learning (RL) by relabelling failed trajectori...
The paper reports that fine‑tuning large language models on aligned data can unintentionally cause misaligned responses in other contexts—a phenomenon termed *context confusion*. The authors demonstrate this effect in gender equality, privacy, and physical safety domains, showing that it differs from emergent misalignment and is not mitigated by general alignment data but can be reduced with domain‑specific alignment or in‑context examples. They provide a mechanistic explanation based on representational shifts during fine‑tuning that lead to behavioral feature transfer across contexts.
The paper introduces Demographic Pluralism, an inference-time framework that estimates population-level opinion distributions without requiring training data or task-specific fine-tuning. It generates multiple perspectives within demographically grounded groups and, across four model backbones on GlobalOpinionQA and VITAL, reduces Jensen-Shannon distance by 8.4%–26.4% compared to Modular Pluralism. The study finds that equal weighting of group perspectives yields the best overall performance, while weighted aggregation performs worse due to increased group-level error.
MADBench is a benchmark that evaluates the security of Multi-Agent Debate (MAD) systems, which allow large language models to exchange and critique answers. The study categorizes attacks into a layered taxonomy aligned with the MAD workflow and tests six attack families across 356 source tasks and 3,958 test cases. Results indicate that while MAD can reduce answer accuracy attacks compared to single-agent baselines, it may amplify unauthorized reads or writes, and even with collusion among agents, the final answer changes from correct to wrong only 28.30% of the time.
The paper proposes measuring a language model’s understanding via no‑arbitrage, defining it as the inability of a bounded trader to profit from Dutch books against the model’s probabilities on logically related claims. It shows that full logical coherence is computationally infeasible, that standard next‑token training yields incoherent predictions across formats, and that uncertainty grows predictably along reasoning chains, creating arbitrage opportunities. The authors introduce Arbitr, a training framework that penalizes logical inconsistencies while maintaining accuracy, reducing exploitability by orders of magnitude and revealing a scaling illusion where large models appear coherent yet exhibit extreme unjustified confidence.
The paper introduces RAE-CoD, a diffusion-based compression method that operates in a representation autoencoder space to preserve recognizable content even at extremely low bitrates. It addresses the problem of semantic collapse observed in existing codecs when the bitrate approaches zero, showing that reconstruction losses conflict with semantic objectives and that VAE diffusion models lose efficiency in preserving semantics. Experiments on MSCOCO-30K demonstrate that RAE-CoD outperforms competitors, reducing VFM feature MSE and Fréchet Distance ratios by at least 25.7% and 69.1% at 0.001–0.008 bpp while maintaining stable recognizability and quality.
PCB-MC is a new dataset for detecting missing components on printed circuit boards, featuring 197 distinct board types with footprint-level annotations derived from the RF100 dataset. The dataset includes multiple augmented samples per board type and introduces board type-aware cross‑validation splits to prevent layout leakage between training and test sets. Benchmark results show that supervised models suffer high false‑negative rates on unseen board designs, while unsupervised anomaly detection methods fail due to misalignment with board-specific references, highlighting the remaining challenges in missing component detection across diverse PCB layouts.
Soft Spatial Reasoning introduces a post‑training framework for Large Vision‑Language Models that replaces hard, token‑by‑token chain‑of‑thought reasoning with a soft, continuous state formed by mixing token embeddings at each intermediate step. The method employs AdaptSoft, a controller that adjusts the degree of softness based on hidden states and predictive uncertainty, guided by a gradient‑alignment learning objective that requires no intermediate supervision. Experiments on diverse spatial benchmarks show that this approach outperforms both hard and fixed‑soft chain‑of‑thought baselines and several existing LVLMs.