A new arXiv study (2607.27564v1) finds that a simple order-vote policy outperforms more complex agentic strategies for multi-image medical visual question answering, achieving 57.89% ± 0.65% final-test accuracy on MedFra…
A new arXiv preprint (2607.27566v1) finds that direct answer-only supervised fine-tuning (SFT) on MedGemma-1.5-4B is the most robust adaptation method for multi-frame medical visual question answering on the MedFrameQA b…
Researchers developed an LLM pipeline that converts audio interviews into transcripts, maps them into the ten MADRS symptom items, estimates severity, and identifies problematic clinical ratings, achieving a correlation …
Researchers propose Generator-Gate-Corrector (GGC), a framework for reliable LLM-based Text-to-SPARQL generation, which improves query-level accuracy from 90.23% to 98.33% on MCQA while reducing inference overhead by 45%…
Researchers introduced FinSMART, the first market-aligned reinforcement learning framework for financial sentiment analysis, which directly optimizes sentiment signals using realized market outcomes and improves cumulati…
A new open-source benchmark framework, ParliamentBench, based on the social deduction game Secret Hitler, evaluates whether large language models (LLMs) can deceive, persuade, and reason under information asymmetry. Test…
A pre-registered audit of 1,179 tutor turns found that general-purpose helpfulness rubrics cannot reliably distinguish direct answer-giving from pedagogical guidance, while a pedagogy-targeted rubric perfectly rank-separ…
Researchers introduced Reflective Retrieval Memory (RRM), a framework that improves long-horizon multimodal reasoning by adding reflective experience memory to an entity-centric multimodal memory graph, distilling reusab…
Researchers propose a novel framework using inverted causal self-attention within a transformer architecture for causal discovery in multivariate time series, outperforming existing methods on linear and nonlinear datase…
Researchers introduced CoMem (Comprehension Memory), a method that leverages the depth division of labor in large language models to enable unbounded-context memory. In evaluations, CoMem achieved 97.05 on RULER and 38.2…
Researchers at the University of Arizona tested a convolutional neural network for low-order wavefront sensing on the Tiny Observatory for Telescope Optimization (TOTO) optical testbed, as detailed in a paper submitted t…
A new study from arXiv (2607.28347v1) found that six large language models, including Qwen3-32B and GPT-5-Mini, fail to simulate human belief updates in controlled environments, with only some matching post-stance distri…
A new study from researchers at arXiv introduces the SaliTrap Benchmark and finds that 12 state-of-the-art large language models, including mainstream models, exhibit 'Salience Bias' in commonsense reasoning, where expli…
Researchers at arXiv introduced a novel evaluation framework that uses pairwise comparisons by multiple large language models (LLMs) and an Elo rating system to approximate expert-level assessments of LLM-generated conte…
Researchers introduced Fairness Pruning, a method that locates demographic bias in GLU-MLP layers of large language models by identifying neurons that react differentially to demographic attributes. Testing on models up …
Researchers from an unnamed institution proposed a structured information extraction framework using LLaMA-3.1-70B to extract six semantic dimensions from financial news, finding that combining these with FinBERT sentime…
ShowTrust, a testimonials collection tool and widget embed, launched with a free tier and a $4.99/mo Pro plan, allowing users to collect unlimited testimonials and embed a lightweight (~5kb) widget via a single snippet o…
Gail Weiner, a former C-suite executive at a tech company, reflects on how status-seeking behavior once tied to luxury goods like her Mercedes is now manifesting in the AI industry, where access to premium models like Fa…
An engineer argues that AI harnesses are essentially middleware and that the security risks they pose are not novel, but rather a rehash of long-known integration security issues. The post emphasizes that the real proble…
Eloqra, a new live chat tool, integrates with Telegram to let website owners answer visitor messages directly from their Telegram app, eliminating the need for a separate dashboard. The service, currently in early access…