Jiachen Zhao

Jiachen Zhao, Wenlong Zhao, Andrew Drozdov et al.

We study semi-supervised sequence generation tasks, where the few labeled examples are too scarce to finetune a model, and meanwhile, few-shot prompted large language models (LLMs) exhibit room for improvement. In this paper, we present the discovery that a student model distilled from a few-shot prompted LLM can commonly generalize better than its teacher to unseen examples on such tasks. We find that the student is able to learn a general pattern from the high-quality pseudolabels produced by the teacher during knowledge distillation (KD), and favorably not a general pattern from the low-quality pseudolables. Leveraging this discovery, we propose a new method, Multistage Collaborative Knowledge Distillation from an LLM (MCKD), for these tasks. MCKD first few-shot prompts an LLM to produce pseudolabels for unlabeled data. Then at each stage of an iterative KD process, a new pair of students is trained on disjoint partitions of the pseudolabeled data, and produces new and improved pseudolabels for their unseen partitions. We conduct extensive experiments on four syntactic and semantic parsing datasets and show the effectiveness of MCKD for low-resource semi-supervised sequence generation. On CRAFT biomedical parsing, for example, 3-stage MCKD with 50 labeled examples outperforms an LLM teacher and vanilla KD by 7.5% and 3.7% parsing F1, respectively, and matches the performance of supervised finetuning with 500 labeled examples.

Jiachen Zhao, Zonghai Yao, Zhichao Yang et al.

In this work, we study in-context teaching (ICT), where a teacher provides in-context example rationales to teach a student to reason over unseen cases. Human teachers are usually required to craft in-context demonstrations, which are costly and have high variance. We ask whether a large language model (LLM) can serve as a more effective in-context teacher for itself or other LLMs, compared to humans. Inspired by the Encoding Specificity Hypothesis from human episodic memory, we hypothesize that in-context exemplars crafted by the teacher should match the training data of the student. This hypothesis motivates us to propose Self-Explain where an LLM's self-elicited explanations are used as in-context demonstrations for prompting it as they are generalized from the model's training examples. Self-Explain is shown to significantly outperform using human-crafted exemplars and other baselines. Furthermore, we reveal that for ICT, rationales from different teacher LLMs or human experts that more resemble the student LLM's self-explanations are better in-context demonstrations. This supports our encoding specificity hypothesis. We then propose Teach-Back that aligns a teacher LLM with the student to enhance the ICT performance. For example, Teach-Back enables a 7B model to teach the much larger GPT-3.5 in context, surpassing human teachers by around 5% in test accuracy on medical question answering.

Kia-Jüng Yang, Dominik Meier, Jiachen Zhao et al.

Large reasoning models (LRMs) generate chain-of-thought (CoT) traces before producing final outputs, introducing a dynamic internal state that may complicate control mechanisms such as refusal. Unlike instruction-tuned LLMs, where refusal is mediated by a single directional subspace, refusal in large reasoning models (LRMs) additionally depends on the CoT. In DeepSeek-R1-Distill-LLaMA-8B, activation steering reverses refusal in only 39% of cases when the CoT is kept fixed, but removing the CoT entirely increases this to 70%, indicating that the CoT actively reinforces refusal. In a two-stage intervention where the model regenerates its CoT under activation steering, refusal is reversed in 94% of cases, while the resulting CoT alone retains 48% of this effect even after steering is removed. This suggests that the CoT can carry and reconstruct the compliance signal independently. These findings indicate that refusal in LRMs is jointly encoded in residual stream activations and CoT. This joint activation makes LRM more robust against activation-level interventions alone, but exposes CoT to a possible alternative surface attack.

Jiachen Zhao

Recently, large language models (LLMs) have made remarkable progress in natural language processing. The most representative ability of LLMs is in-context learning (ICL), which enables LLMs to learn patterns from in-context exemplars without training. The performance of ICL greatly depends on the exemplars used. However, how to choose exemplars remains unclear due to the lack of understanding of how in-context learning works. In this paper, we present a novel perspective on ICL by conceptualizing it as contextual retrieval from a model of associative memory. We establish a theoretical framework of ICL based on Hopfield Networks. Based on our framework, we look into how in-context exemplars influence the performance of ICL and propose more efficient active exemplar selection. Our study sheds new light on the mechanism of ICL by connecting it to memory retrieval, with potential implications for advancing the understanding of LLMs.

Jiachen Zhao, Haiqin Yang

Event detection (ED), aiming to detect events from texts and categorize them, is vital to understanding actual happenings in real life. However, mainstream event detection models require high-quality expert human annotations of triggers, which are often costly and thus deter the application of ED to new domains. Therefore, in this paper, we focus on low-resource ED without triggers and aim to tackle the following formidable challenges: multi-label classification, insufficient clues, and imbalanced events distribution. We propose a novel trigger-free ED method via Derangement mechanism on a machine Reading Comprehension (DRC) framework. More specifically, we treat the input text as Context and concatenate it with all event type tokens that are deemed as Answers with an omitted default question. So we can leverage the self-attention in pre-trained language models to absorb semantic relations between input text and the event types. Moreover, we design a simple yet effective event derangement module (EDM) to prevent major events from being excessively learned so as to yield a more balanced training process. The experiment results show that our proposed trigger-free ED model is remarkably competitive to mainstream trigger-based models, showing its strong performance on low-source event detection.

Jiachen Zhao, Zhun Deng, David Madras et al.

As the number of large language models (LLMs) released to the public grows, there is a pressing need to understand the safety implications associated with these models learning from third-party custom finetuning data. We explore the behavior of LLMs finetuned on noisy custom data containing unsafe content, represented by datasets that contain biases, toxicity, and harmfulness, finding that while aligned LLMs can readily learn this unsafe content, they also tend to forget it more significantly than other examples when subsequently finetuned on safer content. Drawing inspiration from the discrepancies in forgetting, we introduce the "ForgetFilter" algorithm, which filters unsafe data based on how strong the model's forgetting signal is for that data. We demonstrate that the ForgetFilter algorithm ensures safety in customized finetuning without compromising downstream task performance, unlike sequential safety finetuning. ForgetFilter outperforms alternative strategies like replay and moral self-correction in curbing LLMs' ability to assimilate unsafe content during custom finetuning, e.g. 75% lower than not applying any safety measures and 62% lower than using self-correction in toxicity score.

Jiachen Zhao

Knowledge distillation (KD) has been widely employed to transfer knowledge from a large language model (LLM) to a specialized model in low-data regimes through pseudo label learning. However, pseudo labels generated by teacher models are usually noisy and may influence KD performance. This study delves into KD with noisy teachers and uncovers that the student model can already generate more accurate predictions than the teacher labels used to train it during KD, indicating its inherent ability to denoise noisy teacher labels. Motivated by this finding, we propose Peer-Advised KD to improve vanilla KD from noisy teachers. Experiments show that Peer-Advised KD can outperform LLM by approximately 5% with 50 human-labeled data, and even competitive to standard supervised finetuning with 750 human-labeled data.