Large Language Models (LLMs) exhibit a troubling duality, capable of both remarkable generalization and brittle, verbatim memorization of their training data. This unpredictability undermines their reliability in high-stakes applications. In this work, we propose a unified framework to understand, identify, and control these distinct reasoning modes. First, we introduce a theoretical model based on the Information Bottleneck (IB) principle, formalizing generalization as the learning of a compressed, task-relevant representation and memorization as a failure to compress. Building on this theory, we develop Dynamic Mode Steering (DMS), a novel inference-time algorithm which comprises two components: (1) a lightweight, causally-grounded linear probe that identifies the model's instantaneous reliance on memorization, and (2) a dynamic activation steering mechanism that nudges the model's computation towards pre-identified generalization circuits. We frame DMS as a form of adaptive, self-contrastive decoding. Experiments on reasoning and faithfulness tasks demonstrate that DMS significantly improves logical consistency and factual accuracy, thereby offering a principled approach to enhancing LLM reliability.

We present a novel approach for controllable mathematical reasoning that leverages self-optimizing thought vectors with entropy minimization. Our method introduces learnable thought vectors that dynamically modulate the internal reasoning process of large language models. Using Gemma-2-9B on GSM8K, we achieve 90.1% accuracy with a controllability score of 0.42, demonstrating that entropy-based rewards effectively guide focused reasoning patterns without requiring external reward annotations. Our analysis reveals distinct thought vector clusters and consistent low-entropy distributions across control conditions, validating our framework for controllable AI reasoning.

Computer-assisted pronunciation training (CAPT) manages to facilitate second-language (L2) learners to practice pronunciation skills by offering timely and instructive feedback. To examine pronunciation proficiency from multiple facets, existing methods for CAPT broadly fall into two categories: mispronunciation detection and diagnosis (MDD) as well as automatic pronunciation assessment (APA). The former aims to pinpoint phonetic pronunciation errors and provide diagnostic feedback, while the latter seeks instead to quantify pronunciation proficiency pertaining to various aspects. Despite the natural complementarity between MDD and APA, researchers and practitioners, however, often treat them as independent tasks with disparate modeling paradigms. In light of this, we in this paper first introduce MuFFIN, a Multi-Faceted pronunciation Feedback model with an Interactive hierarchical Neural architecture, to jointly address the tasks of MDD and APA. To better capture the nuanced distinctions between phonemes in the feature space, a novel phoneme-contrastive ordinal regularization mechanism is then put forward to optimize the proposed model to generate more phoneme-discriminative features while factoring in the ordinality of the aspect scores. In addition, to address the intricate data imbalance problem in MDD, we design a simple yet effective training objective, which is specifically tailored to perturb the outputs of a phoneme classifier with the phoneme-specific variations, so as to better render the distribution of predicted phonemes meanwhile considering their mispronunciation characteristics. A series of experiments conducted on the Speechocean762 benchmark dataset demonstrates the efficacy of our method in relation to several cutting-edge baselines, showing state-of-the-art performance on both the APA and MDD tasks.

When writing and talking, people sometimes pause to think. Although reasoning-focused works have often framed reasoning as a method of answering questions or completing agentic tasks, reasoning is implicit in almost all written text. For example, this applies to the steps not stated between the lines of a proof or to the theory of mind underlying a conversation. In the Self-Taught Reasoner (STaR, Zelikman et al. 2022), useful thinking is learned by inferring rationales from few-shot examples in question-answering and learning from those that lead to a correct answer. This is a highly constrained setting -- ideally, a language model could instead learn to infer unstated rationales in arbitrary text. We present Quiet-STaR, a generalization of STaR in which LMs learn to generate rationales at each token to explain future text, improving their predictions. We address key challenges, including 1) the computational cost of generating continuations, 2) the fact that the LM does not initially know how to generate or use internal thoughts, and 3) the need to predict beyond individual next tokens. To resolve these, we propose a tokenwise parallel sampling algorithm, using learnable tokens indicating a thought's start and end, and an extended teacher-forcing technique. Encouragingly, generated rationales disproportionately help model difficult-to-predict tokens and improve the LM's ability to directly answer difficult questions. In particular, after continued pretraining of an LM on a corpus of internet text with Quiet-STaR, we find zero-shot improvements on GSM8K (5.9%$\rightarrow$10.9%) and CommonsenseQA (36.3%$\rightarrow$47.2%) and observe a perplexity improvement of difficult tokens in natural text. Crucially, these improvements require no fine-tuning on these tasks. Quiet-STaR marks a step towards LMs that can learn to reason in a more general and scalable way.

An emotional support conversation system aims to alleviate users' emotional distress and assist them in addressing their challenges. To generate supportive responses, it is critical to consider multiple factors such as empathy, support strategies, and response coherence, as established in prior methods. Nonetheless, previous models occasionally generate unhelpful responses, which intend to provide support but display counterproductive effects. According to psychology and communication theories, poor performance in just one contributing factor might cause a response to be unhelpful. From the model training perspective, since these models have not been exposed to unhelpful responses during their training phase, they are unable to distinguish if the tokens they generate might result in unhelpful responses during inference. To address this issue, we introduce a novel model-agnostic framework named mitigating unhelpfulness with multifaceted AI feedback for emotional support (Muffin). Specifically, Muffin employs a multifaceted AI feedback module to assess the helpfulness of responses generated by a specific model with consideration of multiple factors. Using contrastive learning, it then reduces the likelihood of the model generating unhelpful responses compared to the helpful ones. Experimental results demonstrate that Muffin effectively mitigates the generation of unhelpful responses while slightly increasing response fluency and relevance.

Recent Multimodal Large Language Models (MLLMs) exhibit impressive abilities to perceive images and follow open-ended instructions. The capabilities of MLLMs depend on two crucial factors: the model architecture to facilitate the feature alignment of visual modules and large language models; the multimodal instruction tuning datasets for human instruction following. (i) For the model architecture, most existing models introduce an external bridge module to connect vision encoders with language models, which needs an additional feature-alignment pre-training. In this work, we discover that compact pre-trained vision language models can inherently serve as ``out-of-the-box'' bridges between vision and language. Based on this, we propose Muffin framework, which directly employs pre-trained vision-language models to act as providers of visual signals. (ii) For the multimodal instruction tuning datasets, existing methods omit the complementary relationship between different datasets and simply mix datasets from different tasks. Instead, we propose UniMM-Chat dataset which explores the complementarities of datasets to generate 1.1M high-quality and diverse multimodal instructions. We merge information describing the same image from diverse datasets and transforms it into more knowledge-intensive conversation data. Experimental results demonstrate the effectiveness of the Muffin framework and UniMM-Chat dataset. Muffin achieves state-of-the-art performance on a wide range of vision-language tasks, significantly surpassing state-of-the-art models like LLaVA and InstructBLIP. Our model and dataset are all accessible at https://github.com/thunlp/muffin.

Model fairness (a.k.a., bias) has become one of the most critical problems in a wide range of AI applications. An unfair model in autonomous driving may cause a traffic accident if corner cases (e.g., extreme weather) cannot be fairly regarded; or it will incur healthcare disparities if the AI model misdiagnoses a certain group of people (e.g., brown and black skin). In recent years, there have been emerging research works on addressing unfairness, and they mainly focus on a single unfair attribute, like skin tone; however, real-world data commonly have multiple attributes, among which unfairness can exist in more than one attribute, called 'multi-dimensional fairness'. In this paper, we first reveal a strong correlation between the different unfair attributes, i.e., optimizing fairness on one attribute will lead to the collapse of others. Then, we propose a novel Multi-Dimension Fairness framework, namely Muffin, which includes an automatic tool to unite off-the-shelf models to improve the fairness on multiple attributes simultaneously. Case studies on dermatology datasets with two unfair attributes show that the existing approach can achieve 21.05% fairness improvement on the first attribute while it makes the second attribute unfair by 1.85%. On the other hand, the proposed Muffin can unite multiple models to achieve simultaneously 26.32% and 20.37% fairness improvement on both attributes; meanwhile, it obtains 5.58% accuracy gain.

Even as the whole of Silicon Valley grapples with historic inflation, a bank crash, and mass layoffs, Google's woes stand apart. The explosion of ChatGPT and artificial intelligence more broadly has produced something of an existential crisis for the company, a "code red" moment for the business. Yes," Sundar Pichai, Google's CEO, told The New York Times. But Google employees are encountering another problem: "They took away the dried mango," says a project manager at Google's San Francisco office, whom I agreed not to name to protect the employee from reprisal. At least at that office, the project manager said, workers are seeing less of long-cherished food items--not just the mango, but also the Maui-onion chips and the fun-size bags of M&Ms.

The groundbreaking work of Neural Module Networks in prior years aimed to break down jobs into simpler modules. Through training from beginning to finish using modules that were reconfigured for various issues, each module would learn its true purpose and become reusable. Nevertheless, it took a lot of work to use this strategy in the actual world due to several problems. Program development, in particular, needed reinforcement learning from scratch or relied on hand-tuned natural language parsers, making them challenging to optimize. Program creation was severely domain-restricted in each scenario.

Multi-agent systems are used in domains where individual component autonomy and cooperation are necessary. The overall system performance requires that the diverse agents maintain quality interactions to facilitate cooperation. A complication to inter-agent interaction occurs when the agents learn (change their own functionality), when new agents are introduced, or existing agents are functionally modified. This research focuses on creating a general use multi-agent system, Middleware Unifying Framework for Independent Nodes System (MUFFINS), and implementing a mechanism, the Megagent, that addresses the interaction challenges. The Megagent provides the ability for agents to assess their performance per data source and to improve it with transformations based on feedback. Evaluation of the concept is tested on data mangled from the Digits dataset to represent learning and new agents and in all cases improves accuracy over a static agent.