This paper presents a unified evaluation of methods aimed at improving cross-lingual consistency in multilingual language models for question answering. The authors compare inference-time interventions and post-training approaches across three model families and three closed-form benchmarks. Post-training methods, especially direct distribution alignment, prove more reliable, while cross-domain transfer remains limited. Tests on culturally diverse QA show no systematic closed-form degradation, but open-ended generation occasionally loses accuracy for non-English responses.

A new paper investigates why automatic speech recognition (ASR) systems produce fluent text unrelated to the input speech. By analyzing two independently trained Conformer-Large models (CTC and RNN-T) under environmental degradation and speaker-background shift, the authors find that the final encoder stage is a critical boundary: bypassing it causes divergence on nearly every utterance, while bypassing middle blocks has little effect. At this stage, representations become more compact, text becomes readable by the decoder, and grapheme information becomes explicit. However, the intervention yields garbled or repetitive output rather than fluent fabrication, so the study identifies a mechanistic precondition for hallucination—failure to produce adequately grounded output—not the complete…

The paper introduces a novel evaluation framework based on the Hüllermeier-Rifqi Index, a generalized variant of the Rand Index, to assess how faithfully phonetic encoding algorithms preserve the similarity structure of word-level IPA transcriptions. Pairwise similarities are computed with normalized edit distance, and a discordance score is derived from the absolute difference between ground-truth and encoded similarities, then calibrated against a random-string baseline. The authors benchmark a range of phonetic encoders on multilingual transcription datasets, supplementing the evaluation with recall analysis based on collision rates, and demonstrate the framework's value for measuring orthographic transparency.

This paper introduces a benchmark for evaluating whether LLMs can recover situated pragmatic meanings from indirect and playful Chinese online comments. Built from over 200,000 public Chinese social media interaction records, the benchmark contains 4,735 human-validated diagnostic items. Across eight LLMs, the mean leave-writer-out accuracy is 68.70% and the best model reaches 81.42%, while human accuracy is 90.8%. Case analyses show models often detect broad irony or playfulness but fail to identify the precise mechanism or interactional move.

Early-onset colorectal cancer is increasing among younger adults, yet red-flag symptoms in this age group lack evidence-based guidelines for follow-up testing, and structured encounter data miss key details such as symptom duration, context, and family history. Researchers developed VERGE, an agentic workflow that combines retrieval-augmented generation with a bounded verification-refinement cycle to extract six red-flag symptoms and family-history risk status from free-text clinical notes. On 4,033 clinician-labeled note-finding pairs, VERGE improved precision from 0.764 to 0.849 and MCC from 0.681 to 0.730 while requiring human review for only 1.5% of claims.

A new study introduces a prediction task: can we tell early whether volunteer mental-health crisis counselors will eventually improve at steering conversations toward positive outcomes? The authors find that tracking how counselors adapt after struggling with specific kinds of conversational moments predicts long-term skill growth better than models trained directly on full transcripts.

MedProb is a lightweight probing framework that answers multiple-choice medical visual question answering (Med-VQA) directly from frozen vision-language model representations, without generating free text. Across PATH-VQA, SLAKE, and VQA-RAD, it recovers substantially more answer-relevant signal than prompting and outperforms medical VLMs and agentic systems. Probing also narrows the apparent gap between small and large models, while free-text generation shows an answer-position bias of up to 10 percentage points.

A new distributional theory shows that LLMs do not treat external evidence as neutral truth; instead, the model's prior beliefs shape whether and how it adopts candidate answers. Across 10 million trials and 12 models, plausible candidates are more persuasive, models absorb their own characteristic errors more readily, and identical evidence can help weaker models while hurting stronger ones—even when the model internally knows the evidence is invalid.

A new preprint compares mean dependency-distance estimates across 38 same-language treebank pairs in Universal Dependencies v2.18. Cross-treebank agreement is only moderate: changing treebanks reverses about 40% of pairwise language orderings and explains roughly 29% of between-group variance, persisting across all twelve preprocessing specifications. Still, every treebank confirms dependency-length minimization, suggesting MDD reflects a corpus-conditioned combination of grammatical, register, and annotation factors rather than a stable language-level parameter.

Researchers introduce QPRU, a lightweight quantum recurrent architecture that keeps forecasting performance on par with state-of-the-art classical and quantum baselines while using far fewer trainable parameters, offering a more scalable and parameter-efficient alternative to models such as LSTM, GRU, QLSTM, and QGRU.

This paper proposes a modular deep recurrent neural network (RNN) design that eases deployment of varied RNN architectures and automates derivative computation for gradient-based learning. The modular approach yields new configurations, including feedforward inter-layer connections, which substantially improve learning of high-order dynamics and nonlinearities while mitigating gradient vanishing/exploding issues across layers. Altitude dynamics modeling for quadrotors demonstrates that existing methods fail to generalize quickly or at all.

BER-PEF is a framework that reframes Bayes error rate estimation as human mobility predictability estimation, providing a unified protocol to compare predictability estimators when ground-truth predictability is unobservable. It maps symbolic sequences, numeric trajectories, contextual features, and learned representations into a shared feature-label space, then measures deviations relative to a reference interval along controlled perturbation curves. Experiments on Foursquare NYC/TKY, GeoLife, and T-Drive show several BER-based estimators achieve lower reference discrepancy than existing methods, while aggregation across perturbation levels improves estimator selection.

This study examines whether large language models (LLMs) can predict weather-related forced outage risk in a zero-shot framework using six years of outage records and high-resolution weather data from a central Texas utility service area. Framing the task as binary severity classification across 3h, 6h, and 12h horizons, the authors benchmark four zero-shot LLMs against two supervised classifiers. Supervised models generally score higher on macro-F1 and precision, while newer LLM generations achieve competitive performance. LLMs further offer complementary strengths in actionable reasoning and geographic scalability, suggesting hybrid approaches may be best.

Wi-Fi-based human activity recognition (HAR) relies on deep models that are memory-hungry to train and run. A new Q-MET framework uses hybrid quantum-classical neural networks plus structured pruning to cut trainable parameters by 90–95% while preserving or even improving classification accuracy.

This paper presents a data fusion correction framework that adapts a CFD-trained deep learning surrogate using pressure-sensitive paint wind-tunnel measurements without retraining the underlying model. Applied to the NASA CRM wing-body configuration, it learns systematic CFD-to-experiment discrepancies and substantially improves agreement with PSP data at Mach 0.85, including the suction peak, shock location, and pressure recovery. The corrected surrogate stays within 2.3–2.7% of the measured Cp range on held-out angles of attack and outperforms direct interpolation between measured conditions.

ProToMEx is a model-agnostic explainability framework that uses probabilistic topic models to learn latent 'topics' as high-level reasons for classifier decisions. It provides both global and local explanations, matches SHAP and LIME in fidelity, and generates local explanations about 30–40x faster on standardized tabular and synthetic datasets.

This paper identifies a new failure mode in latent world models called physical representation laziness, which can persist even when representation collapse is prevented. The authors introduce a lightweight Fourier auxiliary head that imposes physically informed structure on the latent space at training time with no inference-time cost, improving planning success rates—especially in dynamic and low-data environments.

HarvestBench is the first benchmark to place an explicit cost on avoiding a side effect and to identify that side effect as a living creature. Introduced in an arXiv paper, the farm simulation has LLM sub-agents run two tractors through a cooperative corn harvest while animals are present. Across nine models and 7,201 priced decisions, 3,951 involved animals; kill rates ranged from 0.4% to 98.8%, independent of model capability. Four of six models were significantly price-sensitive, and all models ran over wild animals more often than farmed ones. Morality briefings were decisive: with the briefing, five of six reasoning models kept kill rates below 6%; without it, all six exceeded 84%. No LLM grader is used; scoring derives from the game log.

This paper proposes the Corporate Language Model, an architecture that converts structured, unstructured, multimodal, and tacit corporate knowledge into an ontology-grounded foundation for reasoning and governed execution. It introduces four architectural pillars, including a Neurosymbolic Mesh, Skill Graph, Living Digital Twins, and Deep Security Layer, plus Spec-as-Code, with real-world evidence from a JCI-accredited tertiary hospital in Brazil under LGPD.

A new arXiv paper warns that deploying more capable LLMs in markets and other high-stakes systems can increase systemic risk due to correlated behavior. Simulations show frontier models act alike, and when they share false information, this homogeneity amplifies risk instead of diversifying it.

This paper introduces a model-agnostic test-time method to improve the faithfulness of LLM explanations. It targets incompleteness by removing from the input concepts not credited in the model's explanation and re-querying the model. The authors show this improves explanation faithfulness across two datasets, multiple model families, and two independent metrics compared to standard prompting and prompting that encourages faithfulness.

The paper introduces Iris-mini and Iris-pro, open-source search agents trained at 35B-A3B and 397B-A17B scales. Tasks are reverse-constructed from web hyperlink structures to create multi-hop questions, then filtered for SFT and optimized via RL against live search. With inference-time context management enabled, the models achieve strong results across BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE.

This paper investigates machine learning classifiers for power system contingency security, categorizing scenarios as safe, moderate, or severe. Using Newton-Raphson load flow and the Overall Performance Index, the authors applied SMOTE and PCA preprocessing, then evaluated KNN, Random Forest, and SVM on N-k contingencies for IEEE-14 and IEEE-30 bus systems. Random Forest achieved the best F1 scores (0.97 IEEE-30, 0.86 IEEE-14), with PCA contributing more to performance than SMOTE. The study suggests ML offers scalable real-time security assessment.

This paper introduces Harbor Adapters, a unified evaluation infrastructure for agentic benchmarks that ports more than 80 benchmarks to evaluate arbitrary agents, validated via rigorous code review and parity experiments. The authors conduct large-scale evaluations of 8 models across 54 benchmarks using Terminus-2 and one of 3 native harnesses. They also present Harbor-Index, a curated meta-dataset of 82 difficult, diverse, high-quality tasks spanning 29 benchmarks. No tested model-harness combination exceeds a 30% pass rate, with the strongest (GPT-5.5 with Codex) reaching 28.0%; all artifacts are open-sourced.

A new technical report introduces EXAONE Finance, a financial time-series foundation model designed for forecasting. It uses an attention-free architecture with causal 1D convolution and group-aware pooling MLP, plus masked context augmentation for missing data, achieving top results on the FinVerse benchmark across point forecasts, asset ranking, and portfolio profitability.

Clipnote saves your AI conversations so they remain available after you close the browser tab.

Simon Willison highlights OpenAI’s new post on research acceleration, which he says marks “RSI day” at the company—RSI standing for Recursive Self-Improvement. The post details how OpenAI’s own research team has adopted coding agents, with median daily AI spend per researcher rising from near zero in February 2026 to about $600 in late August. Willison suspects the late-July surge coincides with employee access to the model later released as GPT-6 Astra.

Airuncode is a developer tool for running multiple coding agents locally on your machine. It is currently listed on Product Hunt, with discussion and links available.

Tucky is a note-taking tool that docks notes to the screen edge and includes an AI agent, aiming to keep quick capture and AI assistance close at hand.

Discussion | Link