Emergent misalignment, where fine-tuning on narrow, non-harmful tasks induces harmful behaviors, poses a key challenge for AI safety in LLMs.Despite growing empirical evidence, its underlying mechanism remains unclear.To uncover the reason behind this phenomenon, we propose a geometric account based on the geometry of feature superposition.Because features are encoded in overlapping representations, fine-tuning that amplifies a target feature also unintentionally strengthens nearby harmful features in accordance with their similarity.We give a simple gradient-level derivation of this effect and empirically test it in multiple LLMs (Gemma-2 2B/9B/27B, LLaMA-3.1 8B, gpt-oss 20B).Using sparse autoencoders (SAEs), we identify features tied to misalignment-inducing data and to harmful behaviors, and show that they are geometrically closer to each other than features derived from non-inducing data.This trend generalizes across domains (e.g., health, career, legal advice).Finally, we show that a geometry-aware approach-filtering training samples closest to toxic features-reduces misalignment by 34.5%, substantially outperforming random removal and achieving comparable or slightly lower misalignment than LLM-as-ajudge-based filtering.Our study links emergent misalignment to feature superposition, providing a basis for understanding and mitigating this phenomenon.This paper analyzes datasets and model outputs that include offensive language.

Large Language Models (LLMs) are increasingly used for Automated Essay Scoring (AES), yet the scoring rubrics they rely on are typically designed for human raters and may not be optimal for LLMs.Inspired by the calibration process that human raters undergo before formal scoring, we propose Reflect-and-Revise, an iterative framework that refines scoring rubrics by prompting models to reflect on their own chain-of-thought rationales and score discrepancies with human labels.At each iteration, the model identifies scoring-error patterns from sampled mismatches and revises the rubric accordingly.Experiments on three essay scoring benchmarks (ASAP, ASAP 2.0, and TOEFL11) with three LLMs (GPT-5 mini, Gemini 3 Flash, and Qwen3-Next-80B-A3B-Instruct) demonstrate that our method yields improvements in Quadratic Weighted Kappa (QWK), achieving gains of up to +0.403 over human-authored rubrics.Starting from a minimal seed rubric that specifies only the score scale, our method matches or exceeds expert rubric performance in most dataset-model combinations, indicating that iterative refinement can reduce the manual effort of rubric authoring.Analysis of the refined rubrics reveals that the refinement process introduces explicit procedural structures, such as conditional gating rules and quantitative thresholds, that are absent from humanauthored rubrics, highlighting a gap between rubrics designed for human raters and those effective for LLMs. 1

Large language models (LLMs) have achieved significant performance improvements in natural language processing (NLP) domain. However, these models often require large computational resources for training and inference. Recently, Mamba, a language model architecture based on State-Space Models (SSMs), has achieved comparable performance to Transformer models while significantly reducing costs by compressing context windows during inference. We focused on the potential of the lightweight Mamba architecture by applying BitNet quantization method to the model architecture. In addition, while prior BitNet methods generally quantized only linear layers in the main body, we extensively quantized the embedding and projection layers considering their significant proportion of model parameters. In our experiments, we applied ternary quantization to the Mamba-2 (170M) architecture and pre-trained the model with 150 B tokens from scratch. Our method achieves approximately 90.0% reduction in the bits used by all parameters, achieving a significant improvement compared with a 48.4% reduction by the conventional BitNet quantization method. In addition, our method experienced minimal performance degradation in both the pre-training perplexity and downstream tasks. These findings demonstrate the potential of incorporating lightweight language models into edge devices, which will become more demanding in the future.

Recent Foundation Model-enabled robotics (FMRs) display greatly improved general-purpose skills, enabling more adaptable automation than conventional robotics. Their ability to handle diverse tasks thus creates new opportunities to replace human labor. However, unlike general foundation models, FMRs interact with the physical world, where their actions directly affect the safety of humans and surrounding objects, requiring careful deployment and control. Based on this proposition, our survey comprehensively summarizes robot control approaches to mitigate physical risks by covering all the lifespan of FMRs ranging from pre-deployment to post-accident stage. Specifically, we broadly divide the timeline into the following three phases: (1) pre-deployment phase, (2) pre-incident phase, and (3) post-incident phase. Throughout this survey, we find that there is much room to study (i) pre-incident risk mitigation strategies, (ii) research that assumes physical interaction with humans, and (iii) essential issues of foundation models themselves. We hope that this survey will be a milestone in providing a high-resolution analysis of the physical risks of FMRs and their control, contributing to the realization of a good human-robot relationship.

Recent large-scale reasoning models have achieved state-of-the-art performance on challenging mathematical benchmarks, yet the internal mechanisms underlying their success remain poorly understood. In this work, we introduce the notion of a reasoning graph, extracted by clustering hidden‐state representations at each reasoning step, and systematically analyze three key graph-theoretic properties: cyclicity, diameter, and small-world index, across multiple tasks (GSM8K, MATH500, AIME~2024). Our findings reveal that distilled reasoning models (e.g., DeepSeek-R1-Distill-Qwen-32B) exhibit significantly more recurrent cycles (about 5 per sample), substantially larger graph diameters, and pronounced small-world characteristics (about 6x) compared to their base counterparts. Notably, these structural advantages grow with task difficulty and model capacity, with cycle detection peaking at the 14B scale and exploration diameter maximized in the 32B variant, correlating positively with accuracy. Furthermore, we show that supervised fine-tuning on an improved dataset systematically expands reasoning graph diameters in tandem with performance gains, offering concrete guidelines for dataset design aimed at boosting reasoning capabilities. By bridging theoretical insights into reasoning graph structures with practical recommendations for data construction, our work advances both the interpretability and the efficacy of large reasoning models.

Abstract Recent decoder-based pre-trained language models (PLMs) generally use subword tokenizers. However, adding character-level perturbations drastically changes the delimitation of texts by the tokenizers, leading to the vulnerability of PLMs. This study proposes a method of continual pre-training to convert decoder-based PLMs with subword tokenizers into perturbation-robust few-shot in-context learners. Our method continually trains decoder-based PLMs to predict the next tokens conditioning on artificially created character-level noisy texts. Since decoder-based language models are auto-regressive, we skip noised words from the target optimization. In addition, to maintain the same word prediction performance under noisy text as clean text, our method employs word distribution matching between the original PLMs and training models. We conducted experiments on various subword-based PLMs, including GPT2, Pythia, Mistral, Gemma2, and Llama3, ranging from 1B to 8B parameters. The results demonstrate that our method consistently improves the performance of few-shot in-context learning on downstream tasks which contain actual typos or misspellings as well as artificial noise.1

Recent large language models (LLMs) have demonstrated remarkable generalization abilities in mathematics and logical reasoning tasks.Prior research indicates that LLMs pre-trained with programming language data exhibit high mathematical and reasoning abilities; however, this causal relationship has not been rigorously tested.Our research aims to verify which programming languages and features during pre-training affect logical inference performance.Specifically, we pre-trained decoderbased language models from scratch using datasets from ten programming languages (e.g., Python, C, Java) and three natural language datasets (Wikipedia, Fineweb, C4) under identical conditions.Thereafter, we evaluated the trained models in a few-shot in-context learning setting on logical reasoning tasks: FLD and bAbi, which do not require commonsense or world knowledge.The results demonstrate that nearly all models trained with programming languages consistently outperform those trained with natural languages, indicating that programming languages contain factors that elicit logic inference performance.In addition, we found that models trained with programming languages exhibit a better ability to follow instructions compared to those trained with natural languages.Further analysis reveals that the depth of Abstract Syntax Trees representing parsed results of programs also affects logical reasoning performance.These findings will offer insights into the essential elements of pretraining for acquiring the foundational abilities of LLMs. 1

Takeshi Kojima, Itsuki Okimura, Yusuke Iwasawa, Hitomi Yanaka, Yutaka Matsuo. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

While Large Language Models (LLMs) have achieved remarkable performance in many tasks, much about their inner workings remains unclear.In this study, we present novel experimental insights into the resilience of LLMs, particularly GPT-4, when subjected to extensive character-level permutations.To investigate this, we first propose the Scrambled Bench, a suite designed to measure the capacity of LLMs to handle scrambled input, in terms of both recovering scrambled sentences and answering questions given scrambled context.The experimental results indicate that multiple advanced LLMs demonstrate the capability akin to typoglycemia 1 , a phenomenon where humans can understand the meaning of words even when the letters within those words are scrambled, as long as the first and last letters remain in place.More surprisingly, we found that only GPT-4 nearly flawlessly processes inputs with unnatural errors, a task that poses significant challenges for other LLMs and often even for humans.Specifically, GPT-4 can almost perfectly reconstruct the original sentences from scrambled ones, decreasing the edit distance by 95%, even when all letters within each word are entirely scrambled.It is counter-intuitive that LLMs can exhibit such resilience despite severe disruption to input tokenization caused by scrambled text. 2

Vision Transformer (ViT) is becoming more popular in image processing. Specifically, we investigate the effectiveness of test-time adaptation (TTA) on ViT, a technique that has emerged to correct its prediction during test-time by itself. First, we benchmark various test-time adaptation approaches on ViT-B16 and ViT-L16. It is shown that the TTA is effective on ViT and the prior-convention (sensibly selecting modulation parameters) is not necessary when using proper loss function. Based on the observation, we propose a new test-time adaptation method called class-conditional feature alignment (CFA), which minimizes both the class-conditional distribution differences and the whole distribution differences of the hidden representation between the source and target in an online manner. Experiments of image classification tasks on common corruption (CIFAR-10-C, CIFAR-100-C, and ImageNet-C) and domain adaptation (digits datasets and ImageNet-Sketch) show that CFA stably outperforms the existing baselines on various datasets. We also verify that CFA is model agnostic by experimenting on ResNet, MLP-Mixer, and several ViT variants (ViT-AugReg, DeiT, and BeiT). Using BeiT backbone, CFA achieves 19.8% top-1 error rate on ImageNet-C, outperforming the existing test-time adaptation baseline 44.0%. This is a state-of-the-art result among TTA methods that do not need to alter training phase.

Pretrained large language models (LLMs) are widely used in many sub-fields of natural language processing (NLP) and generally known as excellent few-shot learners with task-specific exemplars. Notably, chain of thought (CoT) prompting, a recent technique for eliciting complex multi-step reasoning through step-by-step answer examples, achieved the state-of-the-art performances in arithmetics and symbolic reasoning, difficult system-2 tasks that do not follow the standard scaling laws for LLMs. While these successes are often attributed to LLMs' ability for few-shot learning, we show that LLMs are decent zero-shot reasoners by simply adding ``Let's think step by step'' before each answer. Experimental results demonstrate that our Zero-shot-CoT, using the same single prompt template, significantly outperforms zero-shot LLM performances on diverse benchmark reasoning tasks including arithmetics (MultiArith, GSM8K, AQUA-RAT, SVAMP), symbolic reasoning (Last Letter, Coin Flip), and other logical reasoning tasks (Date Understanding, Tracking Shuffled Objects), without any hand-crafted few-shot examples, e.g. increasing the accuracy on MultiArith from 17.7% to 78.7% and GSM8K from 10.4% to 40.7% with large-scale InstructGPT model (text-davinci-002), as well as similar magnitudes of improvements with another off-the-shelf large model, 540B parameter PaLM. The versatility of this single prompt across very diverse reasoning tasks hints at untapped and understudied fundamental zero-shot capabilities of LLMs, suggesting high-level, multi-task broad cognitive capabilities may be extracted by simple prompting. We hope our work not only serves as the minimal strongest zero-shot baseline for the challenging reasoning benchmarks, but also highlights the importance of carefully exploring and analyzing the enormous zero-shot knowledge hidden inside LLMs before crafting finetuning datasets or few-shot exemplars.