Large language models (LLMs) face a critical alignment challenge: balancing safety with helpfulness. Excessive safety can lead to over-refusal, where models reject harmful-looking yet benign queries, severely limiting utility.Existing training-free interventions offer an efficient way to mitigate over-refusal without re-training, but suffer from high inference overhead and architecture dependency. Our work explores a complementary direction: rather than applying post-hoc corrections to model outputs, our goal is to intrinsically reshape the distributions of harmful and benign samples within the model’s decision space. In this paper, we argue that a lightweight training-based approach can more effectively distinguish between harmful and benign samples. We propose Single Token Alignment (STA), which optimizes only a single-token prefix (e.g., 4,096 parameters) while keeping the base model frozen. To address the inherent challenge of achieving robust refinement through such a minimal parameter interface, STA employs a mixed weighting mechanism integrated with its optimization objective. This mechanism incorporates hard weighting via stringent data filtering to provide clear, unbiased learning signals, and soft weighting through a focal mechanism to prioritize challenging cases.Extensive experiments across 9 models and 10 datasets demonstrate that STA achieves a superior safety-helpfulness balance for LLMs, MLLMs, and reasoning models, offering a highly efficient and generalizable solution for refining safety alignment.

Although Large Audio-Language Models (LALMs) deliver state-of-the-art (SOTA) performance, they frequently suffer from hallucinations, e.g., generating text not grounded in the audio input. We analyze these grounding failures and identify a distinct taxonomy: Event Omission, False Event Identity, Temporal Relation Error, and Quantitative Temporal Error. To address this, we introduce the AHA (Audio Hallucination Alignment) framework. By leveraging counterfactual hard negative mining, our pipeline constructs a high-quality preference dataset that forces models to distinguish strict acoustic evidence from linguistically plausible fabrications. Additionally, we establish AHA-Eval, a diagnostic benchmark designed to rigorously test these fine-grained reasoning capabilities. We apply this data to align Qwen2.5-Omni. The resulting model, Qwen-Audio-AHA, achieves a 13.7% improvement on AHA-Eval. Crucially, this benefit generalizes beyond our diagnostic set. Our model shows substantial gains on public benchmarks, including 1.3% on MMAU-Test and 1.6% on MMAR, outperforming latest SOTA methods.

Retrieval-augmented generation reduces hallucination by grounding model outputs in external evidence, yet hallucinations can still occur even when the retrieved context is accurate and sufficient. From the perspective of information routing in the residual stream, this reflects an imbalance where internal parametric knowledge overwhelms external context during generation. We present an attention-centric analysis of RAG hallucination under valid evidence, showing that hallucinated and factual tokens diverge in mid-to-late Transformer layers as context-selective attention routing weakens, allowing parametric influence to dominate the residual stream. Motivated by prior studies showing that some attention heads—often referred to as copying heads—exhibit stronger information transport capacity, we aim to extend similar evidence-carrying behavior to a broader set of attention heads. To this end, we introduce CoDA, a lightweight inference-time attention intervention that amplifies evidence-aligned value states, enabling more attention heads to transport reliable external evidence in a copy-encouraged manner. Experiments demonstrate that CoDA improves contextual faithfulness, reduces hallucination, and remains robust under long and noisy contexts with modest and stable inference overhead.

Distilling the capabilities from a large reasoning model (LRM) to a smaller student model often involves training on substantial amounts of reasoning data. However, knowledge distillation (KD) over lengthy sequences with prompt (P), chain-of-thought (CoT), and answer (A) sections makes the process computationally expensive. In this work, we investigate how the allocation of supervision across different sections (P, CoT, A) affects student performance. Our analysis shows that selective KD over only the CoT tokens can be effective when the prompt and answer information is encompassed by it. Building on this insight, we establish a truncation protocol to quantify computation-quality tradeoffs as a function of sequence length. We observe that beyond a specific length, longer training sequences provide marginal returns for downstream performance but require substantially higher memory and FLOPs. To this end, training on only the first 50% of tokens of every training sequence can retain, on average, ≈91% of full-sequence performance on math benchmarks while reducing training time, memory usage, and FLOPs by about 50% each. Codes are available at https://github.com/weiruichen01/distilling-the-essence.

Large language models (LLMs) have demonstrated remarkable performance across a wide range of industrial applications, from search and recommendation systems to generative tasks. Although scaling laws indicate that larger models generally yield better generalization and performance, their substantial computational requirements often render them impractical for many real-world scenarios at scale. In this paper, we present a comprehensive set of insights for training and deploying small language models (SLMs) that deliver high performance for a variety of industry use cases. We focus on two key techniques: (1) knowledge distillation and (2) model compression via structured pruning and quantization. These approaches enable SLMs to retain much of the quality of their larger counterparts while significantly reducing training/serving costs and latency. We detail the impact of these techniques on a variety of use cases in a large professional social network platform and share deployment lessons, including hardware optimization strategies that improve speed and throughput for both predictive and reasoning-based applications in Recommendation Systems.

This paper describes the system we submitted to the IWSLT 2023 multilingual speech translation track, with input being English speech and output being text in 10 target languages. Our system consists of CNN and Transformer, convolutional neural networks downsample speech features and extract local information, while transformer extract global features and output the final results. In our system, we use speech recognition tasks to pre-train encoder parameters, and then use speech translation corpus to train the multilingual speech translation model. We have also adopted other methods to optimize the model, such as data augmentation, model ensemble, etc. Our system can obtain satisfactory results on test sets of 10 languages in the MUST-C corpus.

In this paper we introduce our Chinese-English simultaneous translation system participating in AutoSimulTrans2021. In simultaneous translation, translation quality and delay are both important. In order to reduce the translation delay, we cut the streaming-input source sentence into segments and translate the segments before the full sentence is received. In order to obtain high-quality translations, we pre-train a translation model with adequate corpus and fine-tune the model with domain adaptation and sentence length adaptation. The experimental results on the evaluation data show that our system performs better than the baseline system.