The Archive

PerceptRL: reinforcement learning approach to improve visual perception in MLLMs during post-training via multi-aspect reward design.

NPC-RubricBench evaluates LLMs across six longitudinal cancer care phases using 941 patient-phase units from 405 nasopharyngeal carcinoma cases.

Connector-based Speech LLMs expand to new languages via targeted fine-tuning and TIES merging while preserving source-language ASR performance.

Study compares selection bias across Jev-style System One models, Jev reproductions, and generative LLMs using multi-option shuffling tests to evaluate whether decision structure reduces bias.

dSpark applies schedule distillation to teach diffusion LLMs when to commit tokens during parallel generation, prioritizing high-context predictions over high-probability ones for faster decoding.

ComFair benchmark audits intersectional fairness in LLMs across 479k dual/multi-attribute instances, exposing harms missed by single-attribute evaluation.

RL method trains LLMs to abstain with user-specified confidence targets, enabling inference-time adaptation of accuracy-abstention tradeoffs.

CoCHI improves LLM calibration by identifying and intervening on internal heads responsible for overconfident errors via causal analysis.

Reversible pseudonymisation benchmark and engine for Russian/English documents sent to external LLMs, evaluating five tools and regex baseline.

Five LLMs execute harmful tool calls 27–99% when given misinformation rated untrustworthy; credibility judgment unreliable.

Taxonomy and deterministic classifier for schema-conformance failures in Pydantic-based structured output validation across nine LLMs, moving beyond aggregate validity rates.

Counterfactual Test-Time Augmentation removes spurious correlations in LLMs without training access by generating counterfactual edits via auxiliary models.

Dataset of 11k controversial topics paired with value-based personas identifies safety gaps in LLM alignment when persona-conditioned requests emerge from diverse value systems.

Industrial case study: SE expertise guides training of 92B coding LLM serving 100k+ developers; RefactorForge pipeline targets repository-level refactoring capabilities.

Proposes instruction-tuned LLMs as implementing a distinct genre combining conversational and formal discursive patterns.

Foyer self-hosted inbox agent with novel measurement protocol (stratified sampling, human-labeled ground truth) compares typed decision models, LLMs, and baselines on 739 multilingual emails.

Empirical evaluation across 11 open-source LLMs shows performance-optimized code via self-refinement may not reduce energy costs or offset training overhead.

LGMT: open-source tool detecting logical inconsistencies in LLM reasoning via metamorphic testing grounded in first-order logic.

PHYRD benchmark evaluates LLMs and agents on 106 theoretical physics tasks (reproduction and discovery), showing 49-68% accuracy on reproduce and 19-74% on discover problems.

Discord's latest update released today has a big improvement for Markdown previews: they should now render correctly instead of with "a bunch of symbols," according to the company. LLMs and AI agents often rely on Markdown files for instructions and context, so these improved previews could be helpful for people sharing information about AI projects on the platform. "When you upload a Markdown (.md) file, you'll now see the file rendered all nice and neat and pretty instead of with a bunch of symbols," Discord says. "This includes tables, formatting, all that good stuff. Totally great for whe...

Evaluates four open LLMs (8-27B) on Spanish MIR medical exam for uncertainty calibration, reasoning soundness, and multimodal grounding beyond accuracy.

Empirical study: users defer to confident but wrong LLMs over correct information; demonstrates trust calibration collapse independent of model sycophancy.

TTR-Compressor uses test-time regression for sequence compression in LLMs; learns end-to-end via backprop, computes coefficients in closed form at inference.

Diffusion language models improve adversarial prompt attacks on black-box LLMs by generating natural, harder-to-detect malicious inputs via learned text decoding.

GPU applications are processing ever-growing datasets with increasingly powerful accelerators. According to the 2025 Stanford AI Index, compute for notable AI... GPU applications are processing ever-growing datasets with increasingly powerful accelerators. According to the 2025 Stanford AI Index, compute for notable AI models has been doubling approximately every five months, while the size of datasets used to train LLMs has been doubling approximately every eight months. As computation and data scale together, moving data efficiently from storage becomes… Source

Empirical benchmark of recent LLMs vs. fine-tuned encoder-decoder models on knowledge-graph-to-text generation, narrowing performance gap.

Mechanistic interpretability reveals causal subspaces controlling metaphor representation in LLMs across architectures.

DIAL enables fine-grained numerical control over instruction-adherence in LLMs via distillation, scaling across diverse instructions efficiently.

LLMs combined with constraint reasoning to generate contrastive explanations for MIP-based optimization decisions, addressing interpretability in DSS.

Study reveals Theory of Mind benchmarks conflate explicit belief recognition with applied behavior prediction; Llama-3.1-8B shows 51.5% aggregate but uneven competencies.