The paper introduces RAG-Pref, a novel, training-free Retrieval Augmented Generation (RAG) method for preference alignment that significantly improves LLM refusal guardrails against agentic attacks wi…

This paper demonstrates that targeted adaptation of the small language model Phi Silica, using dataset curation and fine-tuning, significantly improves its performance in short-form text rewriting, na…

The paper proposes EKSFT, a selective fine-tuning method that masks high-entropy or high-KL divergence tokens during Supervised Fine-Tuning (SFT) to prevent distribution shift and improve subsequent R…

The paper proposes DOA, a training-free attention policy that leverages self-attention in decoder-only SpeechLLMs to achieve high-quality, low-latency simultaneous long-form translation without requir…

The paper introduces Script-Normalized WER (SN-WER), a novel evaluation metric that transliterates ASR transcripts into a canonical script to accurately measure speech recognition performance across d…

The paper analyzes token reduction for efficient unified VLM training, finding that while task-specific acceleration saves computation, it destroys the mutual performance gains achieved through joint…

This paper proposes a domain-specialized large language model, PoetryQwen, for precise translation and emotional understanding of classical poetry.

The paper introduces semantic motion anchors, a method that bridges the gap between spoken text and gesture meaning by providing structured, semantically grounded supervision, significantly improving…

PolySpeech-100 introduces a massive, multi-lingual benchmark covering 110 linguistic variants to rigorously test Speech-LLMs, demonstrating that open-source models struggle with low-resource languages…

The paper proposes an advanced auditing framework for classical-to-modern LLM translations, demonstrating that embedding drift signals potential error severity rather than error itself, and identifyin…

The paper introduces three new Korean speech benchmarks (KVoiceBench, KOpenAudioBench, and KMMAU) to evaluate SpeechLMs, demonstrating that English-centric evaluation fails to capture performance gaps…

The study investigates the generalization of auto-generated natural-language labels for language model features, finding that while the underlying features show cross-lingual semantic consistency, the…

The paper benchmarks local, offline LLMs for confidential translation workflows, demonstrating that while they are viable for privacy-sensitive use, they generally lag behind top commercial NMT system…

The paper proposes SubFit, a novel compression technique that achieves superior LLM compression by replacing non-contiguous, submodule-level components (Attention and FeedForward) with lightweight res…

The paper demonstrates that the AI-like style introduced by post-training alignment can be measured, localized, and causally removed using a novel ablation technique called PASTA.

The paper proposes using fine-grained quality signals, such as pairwise self-judgments and token-level entropy, instead of simple binary correctness to improve LLM performance on saturated datasets, s…

The paper proposes Sensitivity-Uncertainty Alignment (SUA), a framework that measures the misalignment between a model's prediction instability and its stated uncertainty to improve model reliability.

The paper identifies a failure mode called spatial lexical bias in MLLMs, where adding a spatial word to options biases the model's choice, and demonstrates that this failure originates primarily from…