Prompt-tuning (PT) for large language models (LLMs) can facilitate the performance on various conventional NLP tasks with significantly fewer trainable parameters. However, our investigation reveals that PT provides limited improvement and may even degrade the primitive performance of LLMs on complex reasoning tasks. Such a phenomenon suggests that soft prompts can positively impact certain instances while negatively affecting others, particularly during the later phases of reasoning. To address these challenges, We first identify an information accumulation within the soft prompts. Through detailed analysis, we demonstrate that this phenomenon is often accompanied by erroneous information flow patterns in the deeper layers of the model, which ultimately lead to incorrect reasoning outcomes. we propose a novel method called \textbf{D}ynamic \textbf{P}rompt \textbf{C}orruption (DPC) to take better advantage of soft prompts in complex reasoning tasks, which dynamically adjusts the influence of soft prompts based on their impact on the reasoning process. Specifically, DPC consists of two stages: Dynamic Trigger and Dynamic Corruption. First, Dynamic Trigger measures the impact of soft prompts, identifying whether beneficial or detrimental. Then, Dynamic Corruption mitigates the negative effects of soft prompts by selectively masking key tokens that interfere with the reasoning process. We validate the proposed approach through extensive experiments on various LLMs and reasoning tasks, including GSM8K, MATH, and AQuA. Experimental results demonstrate that DPC can consistently enhance the performance of PT, achieving 4\%-8\% accuracy gains compared to vanilla prompt tuning, highlighting the effectiveness of our approach and its potential to enhance complex reasoning in LLMs.

As large language models (LLMs) become increasingly capable, ensuring that their self-generated explanations are faithful to their internal decision-making process is critical for safety and oversight. In this work, we conduct a comprehensive counterfactual faithfulness analysis across 62 models from 8 families, encompassing both pretrained and instruction-tuned variants and significantly extending prior studies of counterfactual tests. We introduce phi-CCT, a simplified variant of the Correlational Counterfactual Test, which avoids the need for token probabilities while explaining most of the variance of the original test. Our findings reveal clear scaling trends: larger models are consistently more faithful on our metrics. However, when comparing instruction-tuned and human-imitated explanations, we find that observed differences in faithfulness can often be attributed to explanation verbosity, leading to shifts along the true-positive/false-positive Pareto frontier. While instruction-tuning and prompting can influence this trade-off, we find limited evidence that they fundamentally expand the frontier of explanatory faithfulness beyond what is achievable with pretrained models of comparable size. Our analysis highlights the nuanced relationship between instruction-tuning, verbosity, and the faithful representation of model decision processes.

The assumption across nearly all language model (LM) tokenization schemes is that tokens should be subwords, i.e., contained within word boundaries. While providing a seemingly reasonable inductive bias, is this common practice limiting the potential of modern LMs? Whitespace is not a reliable delimiter of meaning, as evidenced by multi-word expressions (e.g., "by the way"), crosslingual variation in the number of words needed to express a concept (e.g., "spacesuit helmet" in German is "raumanzughelm"), and languages that do not use whitespace at all (e.g., Chinese). To explore the potential of tokenization beyond subwords, we introduce a "superword" tokenizer, SuperBPE, which incorporates a simple pretokenization curriculum into the byte-pair encoding (BPE) algorithm to first learn subwords, then superwords that bridge whitespace. This brings dramatic improvements in encoding efficiency: when fixing the vocabulary size to 200k, SuperBPE encodes a fixed piece of text with up to 33% fewer tokens than BPE on average. In experiments, we pretrain 8B transformer LMs from scratch while fixing the model size, vocabulary size, and train compute, varying *only* the algorithm for learning the vocabulary. Our model trained with SuperBPE achieves an average +4.0% absolute improvement over the BPE baseline across 30 downstream tasks (including +8.2% on MMLU), while simultaneously requiring 27% less compute at inference time. In analysis, we find that SuperBPE results in segmentations of text that are more uniform in per-token difficulty. Qualitatively, this may be because SuperBPE tokens often capture common multi-word expressions that function semantically as a single unit. SuperBPE is a straightforward, local modification to tokenization that improves both encoding efficiency and downstream performance, yielding better language models overall.

Recent advances in large language models (LLMs) have introduced the novel paradigm of using LLMs as judges, where an LLM evaluates and scores the outputs of another LLM, which often correlates highly with human preferences. However, the use of LLM-as-a-judge has been primarily studied in English. In this paper, we evaluate this framework in Russian by introducing the Russian Error tyPes Annotation dataset (REPA), a dataset of 1k user queries and 2k LLM-generated responses. Human annotators labeled each response pair expressing their preferences across ten specific error types, as well as selecting an overall preference. We rank six generative LLMs across the error types using three rating systems based on human preferences. We also evaluate responses using eight LLM judges in zero-shot and few-shot settings. We describe the results of analyzing the judges and position and length biases. Our findings reveal a notable gap between LLM judge performance in Russian and English. However, rankings based on human and LLM preferences show partial alignment, suggesting that while current LLM judges struggle with fine-grained evaluation in Russian, there is potential for improvement.

Multimodal large language models (MLLMs) excel at 2D visual understanding but remain limited in their ability to reason about 3D space. In this work, we leverage large-scale high-quality 3D scene data with open-set annotations to introduce 1) a novel supervised fine-tuning dataset and 2) a new evaluation benchmark, focused on indoor scenes. Our Cubify Anything VQA (CA-VQA) data covers diverse spatial tasks including spatial relationship prediction, metric size and distance estimation, and 3D grounding. We show that CA-VQA enables us to train MM-Spatial, a strong generalist MLLM that also achieves state-of-the-art performance on 3D spatial understanding benchmarks, including our own. We show how incorporating metric depth and multi-view inputs (provided in CA-VQA) can further improve 3D understanding, and demonstrate that data alone allows our model to achieve depth perception capabilities comparable to dedicated monocular depth estimation models. We will publish our SFT dataset and benchmark.

Knowledge editing aims to update outdated information in Large Language Models (LLMs). A representative line of study is locate-then-edit methods, which typically employ causal tracing to identify the modules responsible for recalling factual knowledge about entities. However, we find these methods are often sensitive only to changes in the subject entity, leaving them less effective at adapting to changes in relations. This limitation results in poor editing locality, which can lead to the persistence of irrelevant or inaccurate facts, ultimately compromising the reliability of LLMs. We believe this issue arises from the insufficient precision of knowledge localization. To address this, we propose a Fine-grained Neuron-level Knowledge Editing (FiNE) method that enhances editing locality without affecting overall success rates. By precisely identifying and modifying specific neurons within feed-forward networks, FiNE significantly improves knowledge localization and editing. Quantitative experiments demonstrate that FiNE efficiently achieves better overall performance compared to existing techniques, providing new insights into the localization and modification of knowledge within LLMs. Recently, various methods for the precise editing of outdated or wrong knowledge within Large Language Models (LLMs) (Touvron et al., 2023a;b; Jiang et al., 2024; Dubey et al., 2024) have been proposed (Mazzia et al., 2023; Yao et al., 2023; Wang et al., 2023). This paper primarily focuses on locate-then-edit methods, which have emerged as a promising and mainstream approach for knowledge editing in LLMs. A key representative of these approaches is ROME (Meng et al., 2022), which employs causal tracing to identify specific modules responsible for recalling facts about subject entities.

After more than a year of Israeli bombing, tens of thousands of Palestinian deaths, and a humanitarian catastrophe in Gaza, the world was largely united in saying "enough is enough". United Nations General Assembly (UNGA) resolution 12667 in December was clear in its demand: An immediate ceasefire in Gaza. Countries as diverse as Vietnam, Zimbabwe and Colombia echoed that call. And yet, bucking that consensus were nine "no" votes – chief among them, as is typical when it comes to resolutions calling for Israel to adhere to international law or human rights, was the United States. The US has provided unwavering support to Israel throughout its war on Gaza, even as Israel faces accusations of genocide at the International Court of Justice (ICJ) and its prime minister has an International Criminal Court (ICC) arrest warrant to his name.

The use of soft and compliant manipulators in marine environments represents a promising paradigm shift for subsea inspection, with devices better suited to tasks owing to their ability to safely conform to items during contact. However, limitations driven by material characteristics often restrict the reach of such devices, with the complexity of obtaining state estimations making control non-trivial. Here, a detailed analysis of a 1m long compliant manipulator prototype for subsea inspection tasks is presented, including its mechanical design, state estimation technique, closed-loop control strategies, and experimental performance evaluation in underwater conditions. Results indicate that both the configuration-space and task-space controllers implemented are capable of positioning the end effector to desired locations, with deviations of <5% of the manipulator length spatially and to within 5^{o} of the desired configuration angles. The manipulator was also tested when subjected to various disturbances, such as loads of up to 300g and random point disturbances, and was proven to be able to limit displacement and restore the desired configuration. This work is a significant step towards the implementation of compliant manipulators in real-world subsea environments, proving their potential as an alternative to classical rigid-link designs.

The early 2020s has seen the rise of two strange and potentially quite impactful social phenomena, namely pseudolaw, where users rely upon pseudolegal arguments that mimic the form and ritual of legal argumentation but fundamentally distort the content of law, and generative AI/LLMs, which generate content that uses probabilistic calculations to create outputs that look like human generated text. This article argues that the juxtaposition of the two phenomena helps to reveal that they both share two fundamental traits as both elevate form and appearance over substance and content, and users of both routinely mistake the form for the substance. In drawing upon legal theory, computer science, linguistics and cognitive psychology, the article argues that both phenomena rely upon creating illusions of meaning that users mistake for the underlying primary phenomenon. I then explore four implications of this conception of both phenomena. Firstly, both rely on human tendencies of conceptual pareidolia resulting in the erroneous perception of meaningful linguistic legal patterns from nebulous inputs. Secondly, both rely upon the confidence heuristic, the human cognitive bias for treating confidence as a proxy for competence. Thirdly, both succeed when the primary concern is with the form of the output and not its content. Fourthly, both rely heavily upon the magical thinking of users and the desire for the promise of the approach to be real. The article argues that the legal context helps to reveal a solution for the problems caused by both phenomena as it is only where users possess sufficient legal and technological literacy that it becomes possible to reveal to them the illusionary nature of the phenomena.

While it is commonly accepted that maintaining common ground plays a role in conversational success, little prior research exists connecting conversational grounding to success in task-oriented conversations. We study failures of grounding in the Ubuntu IRC dataset, where participants use text-only communication to resolve technical issues. We find that disruptions in conversational flow often stem from a misalignment in common ground, driven by a divergence in beliefs and assumptions held by participants. These disruptions, which we call conversational friction, significantly correlate with task success. We find that although LLMs can identify overt cases of conversational friction, they struggle with subtler and more context-dependent instances requiring pragmatic or domain-specific reasoning.