Hakaze Cho

Hakaze Cho, Yoshihiro Sakai, Mariko Kato et al.

In-Context Learning (ICL) typically utilizes classification criteria from output probabilities of manually selected label tokens. However, we argue that such token-based classification criteria lead to suboptimal decision boundaries, despite delicate calibrations through translation and constrained rotation applied. To address this problem, we propose Hidden Calibration, which renounces token probabilities and uses the nearest centroid classifier on the LM's last hidden states. In detail, we assign the label of the nearest centroid previously estimated from a calibration set to the test sample as the predicted label. Our experiments on 6 models and 10 classification datasets indicate that Hidden Calibration consistently outperforms current token-based baselines by about 20%~50%, achieving a strong state-of-the-art in ICL. Our further analysis demonstrates that Hidden Calibration finds better classification criteria with less inter-class overlap, and LMs provide linearly separable intra-class clusters with the help of demonstrations, which supports Hidden Calibration and gives new insights into the principle of ICL. Our official code implementation can be found at https://github.com/hc495/Hidden_Calibration.

Haolin Yang, Hakaze Cho, Yiqiao Zhong et al.

The unusual properties of in-context learning (ICL) have prompted investigations into the internal mechanisms of large language models. Prior work typically focuses on either special attention heads or task vectors at specific layers, but lacks a unified framework linking these components to the evolution of hidden states across layers that ultimately produce the model's output. In this paper, we propose such a framework for ICL in classification tasks by analyzing two geometric factors that govern performance: the separability and alignment of query hidden states. A fine-grained analysis of layer-wise dynamics reveals a striking two-stage mechanism: separability emerges in early layers, while alignment develops in later layers. Ablation studies further show that Previous Token Heads drive separability, while Induction Heads and task vectors enhance alignment. Our findings thus bridge the gap between attention heads and task vectors, offering a unified account of ICL's underlying mechanisms.

Hakaze Cho, Naoya Inoue

Classification tasks are widely investigated in the In-Context Learning (ICL) paradigm. However, current efforts are evaluated on disjoint benchmarks and settings, while their performances are significantly influenced by some trivial variables, such as prompt templates, data sampling, instructions, etc., which leads to significant inconsistencies in the results reported across various literature, preventing fair comparison or meta-analysis across different papers. Therefore, this paper proposes a standardized and easy-to-use evaluation toolkit (StaICC) for in-context classification. Including, for the normal classification task, we provide StaICC-Normal, selecting 10 widely used datasets, and generating prompts with a fixed form, to mitigate the variance among the experiment implementations. To enrich the usage of our benchmark, we also provide a sub-benchmark StaICC-Diag for diagnosing ICL from several aspects, aiming for a more robust inference processing.

Haolin Yang, Hakaze Cho, Kaize Ding et al.

Large Language Models (LLMs) can perform new tasks from in-context demonstrations, a phenomenon known as in-context learning (ICL). Recent work suggests that these demonstrations are compressed into task vectors (TVs), compact task representations that LLMs exploit for predictions. However, prior studies typically extract TVs from model outputs or hidden states using cumbersome and opaque methods, and they rarely elucidate the mechanisms by which TVs influence computation. In this work, we address both limitations. First, we propose directly training Learned Task Vectors (LTVs), which surpass extracted TVs in accuracy and exhibit superior flexibility-acting effectively at arbitrary layers, positions, and even with ICL prompts. Second, through systematic analysis, we investigate the mechanistic role of TVs, showing that at the low level they steer predictions primarily through attention-head OV circuits, with a small subset of "key heads" most decisive. At a higher level, we find that despite Transformer nonlinearities, TV propagation is largely linear: early TVs are rotated toward task-relevant subspaces to improve logits of relevant labels, while later TVs are predominantly scaled in magnitude. Taken together, LTVs not only provide a practical approach for obtaining effective TVs but also offer a principled lens into the mechanistic foundations of ICL.

Haolin Yang, Hakaze Cho, Naoya Inoue

We investigate the mechanistic underpinnings of in-context learning (ICL) in large language models by reconciling two dominant perspectives: the component-level analysis of attention heads and the holistic decomposition of ICL into Task Recognition (TR) and Task Learning (TL). We propose a novel framework based on Task Subspace Logit Attribution (TSLA) to identify attention heads specialized in TR and TL, and demonstrate their distinct yet complementary roles. Through correlation analysis, ablation studies, and input perturbations, we show that the identified TR and TL heads independently and effectively capture the TR and TL components of ICL. Using steering experiments with geometric analysis of hidden states, we reveal that TR heads promote task recognition by aligning hidden states with the task subspace, while TL heads rotate hidden states within the subspace toward the correct label to facilitate prediction. We further show how previous findings on ICL mechanisms, including induction heads and task vectors, can be reconciled with our attention-head-level analysis of the TR-TL decomposition. Our framework thus provides a unified and interpretable account of how large language models execute ICL across diverse tasks and settings.

Hakaze Cho, Haolin Yang, Gouki Minegishi et al.

In-context Learning (ICL) is an emerging few-shot learning paradigm based on modern Language Models (LMs), yet its inner mechanism remains unclear. In this paper, we investigate the mechanism through a novel perspective of information removal. Specifically, we demonstrate that in the zero-shot scenario, LMs encode queries into non-selective representations in hidden states containing information for all possible tasks, leading to arbitrary outputs without focusing on the intended task, resulting in near-zero accuracy. Meanwhile, we find that selectively removing specific information from hidden states by a low-rank filter effectively steers LMs toward the intended task. Building on these findings, by measuring the hidden states on carefully designed metrics, we observe that few-shot ICL effectively simulates such task-oriented information removal processes, selectively removing the redundant information from entangled non-selective representations, and improving the output based on the demonstrations, which constitutes a key mechanism underlying ICL. Moreover, we identify essential attention heads inducing the removal operation, termed Denoising Heads, which enables the ablation experiments blocking the information removal operation from the inference, where the ICL accuracy significantly degrades, especially when the correct label is absent from the few-shot demonstrations, confirming both the critical role of the information removal mechanism and denoising heads.

Hakaze Cho, Haolin Yang, Brian M. Kurkoski et al.

Existing works are dedicated to untangling atomized numerical components (features) from the hidden states of Large Language Models (LLMs) for interpreting their mechanism. However, they typically rely on autoencoders constrained by some implicit training-time regularization on single training instances (i.e., $L_1$ normalization, top-k function, etc.), without an explicit guarantee of global sparsity among instances, causing a large amount of dense (simultaneously inactive) features, harming the feature sparsity and atomization. In this paper, we propose a novel autoencoder variant that enforces minimal entropy on minibatches of hidden activations, thereby promoting feature independence and sparsity across instances. For efficient entropy calculation, we discretize the hidden activations to 1-bit via a step function and apply gradient estimation to enable backpropagation, so that we term it as Binary Autoencoder (BAE) and empirically demonstrate two major applications: (1) Feature set entropy calculation. Entropy can be reliably estimated on binary hidden activations, which we empirically evaluate and leverage to characterize the inference dynamics of LLMs and In-context Learning. (2) Feature untangling. Similar to typical methods, BAE can extract atomized features from LLM's hidden states. To robustly evaluate such feature extraction capability, we refine traditional feature-interpretation methods to avoid unreliable handling of numerical tokens, and show that BAE avoids dense features while producing the largest number of interpretable ones among baselines, which confirms the effectiveness of BAE serving as a feature extractor.

Hakaze Cho, Peng Luo, Mariko Kato et al.

In-context Learning (ICL) utilizes structured demonstration-query inputs to induce few-shot learning on Language Models (LMs), which are not originally pre-trained on ICL-style data. To bridge the gap between ICL and pre-training, some approaches fine-tune LMs on large ICL-style datasets by an end-to-end paradigm with massive computational costs. To reduce such costs, in this paper, we propose Attention Behavior Fine-Tuning (ABFT), utilizing the previous findings on the inner mechanism of ICL, building training objectives on the attention scores instead of the final outputs, to force the attention scores to focus on the correct label tokens presented in the context and mitigate attention scores from the wrong label tokens. Our experiments on 9 modern LMs and 8 datasets empirically find that ABFT outperforms in performance, robustness, unbiasedness, and efficiency, with only around 0.01% data cost compared to the previous methods. Moreover, our subsequent analysis finds that the end-to-end training objective contains the ABFT objective, suggesting the implicit bias of ICL-style data to the emergence of induction heads. Our work demonstrates the possibility of controlling specific module sequences within LMs to improve their behavior, opening up the future application of mechanistic interpretability.

Mariko Kato, Hakaze Cho, Yoshihiro Sakai et al.

The performance of In-Context Learning (ICL) is highly sensitive to the selected demonstrations. Existing approaches to demonstration selection optimize different objectives, yielding inconsistent results. To address this, we propose a unified metric--affinity and diversity--that leverages ICL model's internal representations. Our experiments show that both affinity and diversity strongly correlate with test accuracies, indicating their effectiveness for demonstration selection. Moreover, we show that our proposed metrics align well with various previous works to unify the inconsistency.

Hakaze Cho, Yoshihiro Sakai, Kenshiro Tanaka et al.

In this paper, we investigate the output token probability information in the output embedding of language models. We find an approximate common log-linear encoding of output token probabilities within the output embedding vectors and empirically demonstrate that it is accurate and sparse. As a causality examination, we steer the encoding in output embedding to modify the output probability distribution accurately. Moreover, the sparsity we find in output probability encoding suggests that a large number of dimensions in the output embedding do not contribute to causal language modeling. Therefore, we attempt to delete the output-unrelated dimensions and find more than 30% of the dimensions can be deleted without significant movement in output distribution and sequence generation. Additionally, in the pre-training dynamics of language models, we find that the output embeddings capture the corpus token frequency information in early steps, even before an obvious convergence of parameters starts.