Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.
Teochew has a substantial speaker community and exhibits distinctive lexical, syntactic, and pragmatic features, yet textual resources for evaluating large language models remain limited. We present T...
Large language model (LLM) agents augmented by tools can automate complex, multi-step tasks, such as web navigation, code generation, and workflow orchestration, by acting on external systems through...
AI-driven penetration testing has been demonstrated with premium frontier models such as GPT-4, but the per-engagement token cost makes continuous, automated testing unaffordable for the smaller organ...
LLM serving is typically offered as a shared, multi-tenant service, where high-demand workloads from one client can cause latency SLO violations for others. Existing solutions for performance isolatio...
The paper explores how Large Language Models (LLMs) such as GPT-4o-mini, Gemini-1.5-pro, and LLaMA can be used to automatically simplify complex medical texts into plain language, a process known as Plain Language Adaptation (PLA). It evaluates different prompting strategies, fine‑tunes models with QLoRA, and incorporates Mixture‑of‑Agents (MoA) techniques to improve adaptability and robustness. The results show that LLM‑driven PLA effectively makes healthcare information more comprehensible while preserving essential content.
The paper introduces MIMIC, a framework that uses executable code to generate rigorous reasoning data for large language models (LLMs). By converting algorithms into verifiable reasoning trajectories through narrative fusion, code-guided test synthesis, and dynamic code instrumentation, MIMIC creates a Code-Instrumented Reward (CIR) that supplies dense, high‑fidelity supervision for reinforcement learning. Models trained with MIMIC’s synthetic dataset show significant, consistent improvements in general reasoning, complex mathematics, and fine‑grained deterministic tasks.
TEDi is a Temporal memory-Enhanced and Denoising Transformer designed for surgical instrument segmentation. It introduces a query-level memory bank with a memory search enhancement encoder to incorporate discriminative representations from past frames, and a temporal consistency denoising module that builds a cross‑frame semantic anchor to stabilize predictions. Experiments on EndoVis 2017 and EndoVis 2018 show that TEDi outperforms existing state‑of‑the‑art methods, indicating its effectiveness for computer‑assisted surgery.
EMODY Flow is a lightweight flow‑matching framework that generates synchronized full‑body motion and facial expressions conditioned on speech and emotion. It attaches to a frozen Qwen‑3 Omni model, reusing its audio codecs to drive two parallel DiT generators for SMPL‑X body pose and FLAME facial expressions. An auxiliary emotion classifier at training time restores emotion sensitivity, enabling EMODY Flow to achieve state‑of‑the‑art gesture quality on BEAT2 and zero‑shot facial animation on TFHP, with significant improvements in FGD, Beat Correlation, and Diversity metrics.
HintMiner is an automatic tool that mines question hints from web Q&A posts using a language‑model‑based MiningNet. It retrieves many Q&A posts, extracts hints via a transformer‑based encoder‑decoder with copying mechanisms, and is trained with a self‑supervised objective on large online data. Evaluated on 60,000 Stack Overflow questions, HintMiner achieves an average BLEU score of 36.17% and ROUGE‑2 of 36.29%.
The paper investigates counterfactual self‑explanations in large language models, where a model edits an input minimally to change its own prediction. Experiments on sentiment analysis and natural language inference with ten instruction‑tuned models from the LLaMA‑3 and Qwen‑2.5 families show that larger models produce more faithful, minimal, and human‑aligned counterfactuals. While rationale‑guided prompts improve minimality and alignment, they do not consistently enhance faithfulness, indicating that explanation quality depends heavily on model capacity and requires empirical validation.
QART is a quantum‑classical hybrid architecture that augments a language model with quantum encoding, CIM‑based QUBO optimization, and quantum decoding to improve long‑horizon reasoning. The authors claim that, under certain assumptions, QART can maintain a non‑zero probability of recovering an optimal reasoning path while traditional autoregressive LLMs see their acceptance probability drop to zero as cumulative risk grows. Experiments on six benchmarks with three backbone models show that QART outperforms the baselines in 14 of 15 pairings, with relative gains up to 84.0% on SciCode.
The paper introduces Style‑Debiased DPO (SD‑DPO), a method that refines large language models’ ability to retrieve stored knowledge by using preference optimization that corrects for style differences while preserving factual accuracy. SD‑DPO evaluates on the EntiGraph storing‑side framework and outperforms baseline CPT on the QuALITY reading‑comprehension benchmark, achieving higher accuracy with far fewer training tokens. In a knowledge‑editing setting (AToKE), SD‑DPO attains an overall accuracy of 0.982, correctly answering queries with either new or old facts based on the requested time period.
EviScope is a new paired counterfactual benchmark that evaluates grounded language models by fixing the question while manipulating evidence—adding, removing, distracting, or contradicting it. The v1.1 dataset includes 40 four‑condition quartets with repaired counterfactual claims and span‑level support labels for automated assessment. Experiments on Qwen2.5‑7B, Llama 3.1 8B, and Gemini 3.5 Flash show that paired metrics reveal grounding behaviors hidden by simple answer accuracy, such as unsupported answers, conflict blindness, and incorrect non‑answer actions.
The article examines how AI policies implemented by software organisations affect developers, using insights from 19 interviews. It finds that while such policies aim to reduce risks like data leaks and unauthorized use, they can also impede developers if not properly engaged. The authors propose developer‑centric strategies to help managers and decision makers implement effective AI policies.
Lit3R is a system developed by tus-nlp for the LitTraceQA shared task, which focuses on evidence-grounded question answering over scientific literature. The system combines off-the-shelf retrieval, reranking, and large language model components without task-specific training, using an iterative retrieval process that merges BM25-based sparse and dense retrieval, cross-encoder reranking, and LLM verification, along with paper-to-paper expansion. In the official test set, Lit3R achieved a 4th place ranking on the leaderboard.
The paper introduces a panoramic aerodynamic performance prediction framework for turbomachinery cascades that first predicts basic Navier-Stokes parameters (temperature, pressure, density) and then derives key turbine stage performance metrics. A transformer‑enhanced neural operator (TNO) is trained on Rotor 37 blades to accurately predict transonic compressor blade performance, outperforming state‑of‑the‑art deep learning operators like FNO and DeepONet. The TNO also supports downstream tasks such as sensitivity analysis and optimization, achieving CFD‑like results while cutting computational cost by four orders of magnitude.
The paper evaluates a learned request‑routing policy for disaggregated large‑language‑model serving, where compute‑heavy prefill and memory‑heavy decode stages run on separate GPU pools. Using a discrete‑event simulator and real NVIDIA A40 GPUs, the calibrated router—leveraging prompt length, predicted output length, KV‑cache pressure, and SLO class—outperforms round‑robin, least‑loaded, and length‑based heuristics, achieving the highest mean goodput (0.864) and lowest variance across three mixed, bursty arrival traces. Hardware calibration proves critical, providing a 4.5‑point goodput boost and roughly 40 % of the tail‑latency advantage, and the learned router can match round‑robin performance with one fewer GPU in certain scenarios.
The paper investigates how reinforcement learning can cause large language model agents to adopt shortcut policies for tool use, relying on superficial prompt cues rather than actual task needs. By creating synthetic environments that mix factual QA and math reasoning, the authors show that agents often invoke tools when cues are present, even when those tools are unnecessary, with spurious invocation rates rising up to 39%. They find that shortcut learning occurs mainly when agents have already mastered the target tool and that semantic alignment between cues and tools amplifies the effect. To counter this, they propose a dense, decision-level reward where an LLM judge assesses tool necessity, which reduces cue-driven tool use while maintaining performance.
The paper introduces MADA, a Large Language Model–powered multi‑agent framework that coordinates specialized agents—Job Management, Geometry, and Inverse Design—to automate complex design workflows on high‑performance computing systems. In the context of Richtmyer–Meshkov Instability suppression for Inertial Confinement Fusion, MADA iteratively refines designs by launching ensemble simulations, generating meshes, and proposing new designs based on simulation outcomes, achieving improved suppression with minimal manual effort. The framework demonstrates how coordinated reasoning, simulation, and specialized tools can be scaled for rapid, automated design exploration.
Cascade is a hierarchical recoverability control framework designed to improve large language model unlearning. It combines path-level routing, representation-level compression, and decoding-level intervention to minimize the internal identifiability of target knowledge. Experiments on TOFU, MUSE-News, and WMDP demonstrate that Cascade reduces recoverability while preserving model utility.