Chain-of-thought reasoning unfolds in discrete token space: each step is committed as text, errors propagate, and eliciting good traces presupposes traces to imitate. Reasoning instead in a model’s continuous representation space – where intermediate states are vectors rather than words – sidesteps these constraints, but leaves open how those latent states should be computed. We approach this along two axes. First, we keep a large language model (LLM) frozen and use it for what it is already good at – modeling and decoding sequences – while a small auxiliary network supplies continuous latent thoughts as input. Second, we produce those latents by recurrence: a tiny recurrent reasoner refines them over many steps, decoupling the depth of computation from the size of the model, so that the latents are a product of iterative processing rather than a single forward pass. We instantiate this as Latent Recurrent Thoughts (LRT): a task-dedicated proposer supplies base latents, a recurrent reasoner refines them through bounded residual corrections, and the frozen LLM decodes the answer. On symbolic reasoning with answer supervision but no reasoning traces (Countdown-4, Sudoku) and on natural-language reasoning (HumanEval, MBPP, StrategyQA), LRT substantially outperforms prior frozen-decoder continuous-space reasoning methods under an identical decoder, prompt, data, and training budget, and outperforms non-thinking-mode chain-of-thought prompting on the same backbone at a small fraction of its inference compute.

Hypernetworks are a class of meta-networks that generate weights for main neural networks. Their unique parameter spaces necessitate exploring suitable optimization strategies to enhance performance, especially for language models. However, a comprehensive investigation into optimization strategies for hypernetworks remains absent. To address this gap, we analyze the loss landscape of hypernetworks and propose that restart optimization strategies can improve their performance for language models. We find that hypernetworks have inherently more complicated loss landscapes compared to conventional networks due to their distinct parameter spaces. Consequently, a restart strategy that periodically resets the learning rate can facilitate better convergence for hypernetworks. Through experiments on instruction tuning and multi-task training, we demonstrate that the restart strategy consistently enhances the performance of hypernetworks for language models, often more effectively than for conventional deep neural networks. Our findings highlight the importance of tailored optimization techniques to unlock the full potential of hypernetworks in natural language processing tasks.

Remarkable progress on large language models (LLMs), particularly in English, has facilitated impressive capabilities in following human instructions. However, there remains a noticeable gap in instruction fine-tuning for Chinese, where the complex linguistic features pose significant challenges. Existing datasets, generally distilled from English-centric LLMs, are not well-aligned with Chinese users’ interaction patterns. To bridge this gap, we introduce COIG-CQIA, a new Chinese instruction tuning dataset derived from various real-world data resources and undergoing comprehensive human verification. We conduct extensive experiments on COIG-CQIA, and compare them with strong baseline models and datasets. The experimental results show that models trained on COIG-CQIA achieve highly competitive performance in diverse benchmarks. Additionally, our findings offer several insights for designing effective Chinese instruction-tuning datasets and data mixing strategies. Our dataset are available at https://huggingface.co/datasets/m-a-p/COIG-CQIA.

Mixture-of-experts (MoE) is gaining increasing attention due to its unique properties and remarkable performance, especially for language tasks. By sparsely activating a subset of parameters for each token, MoE architecture could increase the model size without sacrificing computational efficiency, achieving a better trade-off between performance and training costs. However, the underlying mechanism of MoE still lacks further exploration, and its modularization degree remains questionable. In this paper, we make an initial attempt to understand the inner workings of MoE-based large language models. Concretely, we comprehensively study the parametric and behavioral features of four popular MoE-based models and reveal some intriguing observations, including 1) Neurons act like fine-grained experts; 2) The router of MoE usually selects experts with larger output norms; 3) The expert diversity increases as the layer increases, while the last layer is an outlier, which is further validated by an initial experiment. Based on the observations, we also provide suggestions for a broad spectrum of MoE practitioners, such as router design and expert allocation. We hope this work could shed light on future research on the MoE framework and other modular architectures. Code is available at https://github.com/kamanphoebe/Look-into-MoEs.

Large Language Models infuse newfound vigor into the advancement of the medical domain, yet the scarcity of data poses a significant bottleneck hindering community progress. In this paper, we release the largest ever medical Question Answering (QA) dataset with 26 Million QA pairs named Huatuo-26M. We benchmark many existing approaches in our dataset in terms of both retrieval and generation. We also experimentally show the benefit of the proposed dataset in many aspects: (i) it serves as a fine-tuning data for training medical Large Language Models (LLMs); (ii) it works as an external knowledge source for retrieval-augmented generation (RAG); (iii) it demonstrates transferability by enhancing zero-shot performance on other QA datasets; and (iv) it aids in training biomedical model as a pre-training corpus. Our empirical findings substantiate the dataset’s utility in these domains, thereby confirming its significance as a resource in the medical QA landscape.

In this paper, we introduce MIO, a novel foundation model built on multimodal tokens, capable of understanding and generating speech, text, images, and videos in an end-to-end, autoregressive manner. While the emergence of large language models (LLMs) and multimodal large language models (MM-LLMs) propels advancements in artificial general intelligence through their versatile capabilities, they still lack true any-to-any understanding and generation. Recently, the release of GPT-4o has showcased the remarkable potential of any-to-any LLMs for complex real-world tasks, enabling omnidirectional input and output across images, speech, and text. However, it is closed-source and does not support the generation of multimodal interleaved sequences. To address this gap, we present MIO, which is trained on a mixture of discrete tokens across four modalities using causal multimodal modeling. MIO undergoes a four-stage training process: (1) alignment pre-training, (2) interleaved pre-training, (3) speech-enhanced pre-training, and (4) comprehensive supervised fine-tuning on diverse textual, visual, and speech tasks. Our experimental results indicate that MIO exhibits competitive, and in some cases superior, performance compared to previous dual-modal baselines, any-to-any model baselines, and even modality-specific baselines. Moreover, MIO demonstrates advanced capabilities inherent to its any-to-any feature, such as interleaved video-text generation, chain-of-visual-thought reasoning, visual guideline generation, instructional image editing, etc.

It is a longstanding challenge to understand how neural models perform mathematical reasoning. Recent mechanistic interpretability work indicates that large language models (LLMs) use a “bag of heuristics” in middle to late-layer MLP neurons for arithmetic, where each heuristic promotes logits for specific numerical patterns. Building on this, we aim for fine-grained manipulation of these heuristic neurons to causally steer model predictions towards specific arithmetic outcomes, moving beyond simply disrupting accuracy. This paper presents a methodology that enables the systematic identification and causal manipulation of heuristic neurons, which is applied to the addition task in this study. We train a linear classifier to predict heuristics based on activation values, achieving over 90% classification accuracy. The trained classifier also allows us to rank neurons by their importance to a given heuristic. By targeting a small set of top-ranked neurons (K=50), we demonstrate high success rates—over 80% for the ones place and nearly 70% for the tens place—in controlling addition outcomes. This manipulation is achieved by transforming the activation of identified neurons into specific target heuristics by zeroing out source-heuristic neurons and adjusting target-heuristic neurons towards their class activation centroids. We explain these results by hypothesizing that high-ranking neurons possess ‘cleaner channels’ for their heuristics, supported by Signal-to-Noise Ratio (SNR) analysis where these neurons show higher SNR scores. Our work offers a robust approach to dissect, causally test, and precisely influence LLM arithmetic, advancing understanding of their internal mechanisms.

Chain-of-thought (CoT) significantly enhances the performance of large language models (LLMs) across a wide range of tasks, and prior research shows that CoT can theoretically increase expressiveness. However, there is limited mechanistic understanding of the algorithms that Transformer+CoT can learn. Our key contributions are: (1) We evaluate the state tracking capabilities of Transformer+CoT and its variants, confirming the effectiveness of CoT. (2) Next, we identify the circuit (a subset of model components, responsible for tracking the world state), indicating that late-layer MLP neurons play a key role. We propose two metrics, compression and distinction, and show that the neuron sets for each state achieve nearly 100% accuracy, providing evidence of an implicit finite state automaton (FSA) embedded within the model. (3) Additionally, we explore three challenging settings: skipping intermediate steps, introducing data noises, and testing length generalization. Our results demonstrate that Transformer+CoT learns robust algorithms (FSAs), highlighting its resilience in challenging scenarios. Our code is available at https://github.com/IvanChangPKU/FSA.

Code LLMs have been widely used in various domains, including code generation, logical reasoning, and agent systems. However, open-access code LLMs mostly only release weights, lacking key features such as reproducible data pipelines and transparent training protocols, which are crucial for advancing deeper, more reliable investigations. To address the gap, we introduce OpenCoder, a top-tier code LLM that not only achieves performance comparable to leading models but also serves as an “open cookbook” for the research community. Unlike most prior efforts, we release not only model weights and inference code, but also the reproducible training data, complete data processing pipeline, rigorous experimental ablation results, and detailed training protocols for open scientific research. Our work identifies the key ingredients for building a top-tier code LLM: optimized heuristic rules for data cleaning and deduplication, effective recall of code-related text corpus, and high-quality synthetic data for both annealing and supervised fine-tuning stages. By offering this level of openness, we aim to broaden access to all aspects of a top-tier code LLM, with OpenCoder serving as both a powerful model and an open foundation to accelerate research and enable reproducible advancements in code intelligence. The released resource is available at https://opencoder-llm.github.io.

Continual pretraining enables Large Language Models (LLMs) to adapt to specialized domains like medicine and law. However, we observe a consistent phenomenon across different model sizes and domains: a temporary performance drop at the start of the continual pretraining process, followed by a performance recovery phase. To gain a deeper understanding of this issue, we use the stability gap— a concept adapted from the visual domain—which explains this initial drop arises from instability in the model’s general abilities. We validate this hypothesis through a series of experiments. To address this initial instability and enhance LLM performance within a fixed compute budget, we propose a training strategy that mitigates instability by increasing the number of epochs, alongside two data sampling strategies targeting data domain relevance and corpus distribution. We conduct experiments on Llama-family models to validate the effectiveness of our strategies for continual pretraining and instruction tuning in medical and legal domains. Our strategies improve the average medical task performance of the OpenLlama-3B model from 36.2% to 40.7% using only 40% of the original training budget, while also enhancing general task performance without causing forgetting. Furthermore, we aPPLy our strategies to continually pre-train and instruction-tune the Llama-3-8B model. The resulting model, Llama-3-Physician, achieves the best medical performance among open-source models on several benchmarks and rivals GPT-4 on specific tasks. We release our models at https://huggingface.co/YiDuo1999/Llama-3-Physician-8B-Instruct.

Alignment of large language models (LLMs) involves training models on preference-contrastive output pairs to adjust their responses according to human preferences. To obtain such contrastive pairs, traditional methods like RLHF and RLAIF rely on limited contrasting patterns, such as varying model variants or decoding temperatures. This singularity leads to two issues: (1) alignment is not comprehensive; and thereby (2) models are susceptible to harmful response tendencies. To address these issues, we investigate how to construct more comprehensive and diversified contrasting patterns to enhance preference data (RQ1) and verify the impact of the diversification of contrasting patterns on model alignment (RQ2). For RQ1, we propose PopAlign, a framework that integrates diversified contrasting patterns across the prompt, model, and pipeline levels, introducing six contrasting strategies that do not require additional feedback labeling procedures. Regarding RQ2, we conduct thorough experiments demonstrating that PopAlign significantly outperforms existing methods, leading to more comprehensive alignment.

Large Language Models (LLMs) demonstrate impressive capabilities across various domains, including role-playing, creative writing, mathematical reasoning, and coding. Despite these advancements, LLMs still encounter challenges with length control, frequently failing to adhere to specific length constraints due to their token-level operations and insufficient training on data with strict length limitations. We identify this issue as stemming from a lack of positional awareness and propose novel approaches—PositionID Prompting and PositionID Fine-Tuning—to address it. These methods enhance the model’s ability to continuously monitor and manage text length during generation. Additionally, we introduce PositionID CP Prompting to enable LLMs to perform copy and paste operations accurately. Furthermore, we develop two benchmarks for evaluating length control and copy-paste abilities. Our experiments demonstrate that our methods significantly improve the model’s adherence to length constraints and copy-paste accuracy without compromising response quality.

The advent of Large Language Models (LLMs) has paved the way for complex tasks such as role-playing, which enhances user interactions by enabling models to imitate various characters. However, the closed-source nature of state-of-the-art LLMs and their general-purpose training limit role-playing optimization. In this paper, we introduce RoleLLM, a framework to benchmark, elicit, and enhance role-playing abilities in LLMs. RoleLLM comprises four stages: (1) Role Profile Construction for 100 roles; (2) Context-Based Instruction Generation (Context-Instruct) for role-specific knowledge extraction; (3) Role Prompting using GPT (RoleGPT) for speaking style imitation; and (4) Role-Conditioned Instruction Tuning (RoCIT) for fine-tuning open-source models along with role customization. By Context-Instruct and RoleGPT, we create RoleBench, the first systematic and fine-grained character-level benchmark dataset for role-playing with 168,093 samples. Moreover, RoCIT on RoleBench yields RoleLLaMA (English) and RoleGLM (Chinese), significantly enhancing role-playing abilities and even achieving comparable results with RoleGPT (using GPT-4).

The introduction of large language models has significantly advanced code generation. However, open-source models often lack the execution capabilities and iterative refinement of advanced systems like the GPT-4 Code Interpreter. To address this, we introduce OpenCodeInterpreter, a family of open-source code systems designed for generating, executing, and iteratively refining code. Supported by Code Feedback, a dataset featuring 68K multi-turn interactions, OpenCodeInterpreter integrates execution and human feedback for dynamic code refinement. Our comprehensive evaluation of OpenCodeInterpreter across key benchmarks such as HumanEval, MBPP, and their enhanced versions from EvalPlus reveals its exceptional performance. Notably, OpenCodeInterpreter-33B achieves an accuracy of 83.2 (76.4) on the average (and plus versions) of HumanEval and MBPP, closely rivaling GPT-4’s 84.2 (76.2) and further elevates to 91.6 (84.6) with synthesized human feedback from GPT-4. OpenCodeInterpreterbrings the gap between open-source code generation models and proprietary systems like GPT-4 Code Interpreter.

Multi-modal information retrieval (MMIR) is a rapidly evolving field where significant progress has been made through advanced representation learning and cross-modality alignment research, particularly in image-text pairing.However, current benchmarks for evaluating MMIR performance on image-text pairings overlook the scientific domain, which has a notable gap with the generic data since the caption of scientific charts and tables usually describes the analysis of experimental results or scientific principles in contrast to human activity or scenery depicted in generic images.To bridge this gap, we develop a scientific domain-specific MMIR benchmark (SciMMIR) by leveraging open-access research paper corpora to extract data relevant to the scientific domain. This benchmark comprises 530K meticulously curated image-text pairs, extracted from figures and tables with detailed captions from scientific documents.We further annotate the image-text pairs with a two-level subset-subcategory hierarchy to facilitate a more comprehensive evaluation of the baselines. We conduct zero-shot and fine-tuned evaluations on prominent multi-modal image-captioning and visual language models, such as CLIP, BLIP, and BLIP-2.Our findings offer critical insights for MMIR in the scientific domain, including the impact of pre-training and fine-tuning settings and the effects of different visual and textual encoders.

The advancement of large language models (LLMs) has enhanced the ability to generalize across a wide range of unseen natural language processing (NLP) tasks through instruction-following.Yet, their effectiveness often diminishes in low-resource languages like Chinese, exacerbated by biased evaluations from data leakage, casting doubt on their true generalizability to new linguistic territories. In response, we introduce the Chinese Instruction-Following Benchmark (CIF-Bench), designed to evaluate the zero-shot generalizability of LLMs to the Chinese language. CIF-Bench comprises 150 tasks and 15,000 input-output pairs, developed by native speakers to test complex reasoning and Chinese cultural nuances across 20 categories. To mitigate data contamination, we release only half of the dataset publicly, with the remainder kept private, and introduce diversified instructions to minimize score variance, totaling 45,000 data instances.Our evaluation of 28 selected LLMs reveals a noticeable performance gap, with the best model scoring only 52.9%, highlighting the limitations of LLMs in less familiar language and task contexts.This work not only uncovers the current limitations of LLMs in handling Chinese language tasks but also sets a new standard for future LLM generalizability research, pushing towards the development of more adaptable, culturally informed, and linguistically diverse models.

While LLMs demonstrate impressive capabilities in musical knowledge, we find that music reasoning is still an unsolved task.We introduce ChatMusician, an open-source large language model (LLM) that integrates intrinsic musical abilities. It is based on continual pre-training and finetuning LLaMA2 on a text-compatible music representation, ABC notation, and the music is treated as a second language.ChatMusician can understand and generate music with a pure text tokenizer without external multi-modal neural structures or tokenizers. Interestingly, endowing musical abilities does not harm language abilities, even achieving a slightly higher MMLU score.ChatMusician is capable of composing well-structured, full-length music, condition on texts, chords, melodies, motifs, musical forms, etc.On our meticulously curated college-level music understanding benchmark, MusicTheoryBench, ChatMusician surpasses LLaMA2 and GPT-3.5 by a noticeable margin. We show that ChatMusician preserves or even surpasses the original LLaMA2 7B’s language abilities by evaluating on MMLU benchmark.Our work reveals that LLMs can be an excellent compressor for music, which can be seen as humanity’s creative language, but there remains significant territory to be conquered.We release our 5B token music-language corpora MusicPiles, the collected MusicTheoryBench, code, model and demo.

Training Large Language Models (LLMs) to process extensive context lengths incurs prohibitive computational costs. Prevailing techniques for extending context capabilities in LLMs typically require not only additional training procedures but also access to datasets with long context (e.g., sequences of 32K tokens), presupposing substantial GPU expenditures. To address the aforementioned issues, we introduce a novel solution named Efficient and Extreme length extension for Large Language Models (E2-LLM). E2-LLM entails a singular training process over considerably short sequences (e.g., 4K tokens), which greatly mitigates the cost of continual-pretraining or fine-tuning. Within the training phase, we incorporate a dual augmentation strategy with Rotary Position Embeddings (RoPE) that adjusts the scale and position indices across distinct training samples. E 2 -LLM is meticulously designed to enhance the model’s robustness to diverse relative positions. The experimental results on multiple benchmark datasets demonstrate the superior performance of E 2 -LLM on demanding tasks of processing long contexts.

Claim: This work is not advocating the use of LLMs for paper (meta-)reviewing. Instead, wepresent a comparative analysis to identify and distinguish LLM activities from human activities. Two research goals: i) Enable better recognition of instances when someone implicitly uses LLMs for reviewing activities; ii) Increase community awareness that LLMs, and AI in general, are currently inadequate for performing tasks that require a high level of expertise and nuanced judgment.This work is motivated by two key trends. On one hand, large language models (LLMs) have shown remarkable versatility in various generative tasks such as writing, drawing, and question answering, significantly reducing the time required for many routine tasks. On the other hand, researchers, whose work is not only time-consuming but also highly expertise-demanding, face increasing challenges as they have to spend more time reading, writing, and reviewing papers. This raises the question: how can LLMs potentially assist researchers in alleviating their heavy workload?This study focuses on the topic of LLMs as NLP Researchers, particularly examining the effectiveness of LLMs in assisting paper (meta-)reviewing and its recognizability. To address this, we constructed the ReviewCritique dataset, which includes two types of information: (i) NLP papers (initial submissions rather than camera-ready) with both human-written and LLM-generated reviews, and (ii) each review comes with “deficiency” labels and corresponding explanations for individual segments, annotated by experts. Using ReviewCritique, this study explores two threads of research questions: (i) “LLMs as Reviewers”, how do reviews generated by LLMs compare with those written by humans in terms of quality and distinguishability? (ii) “LLMs as Metareviewers”, how effectively can LLMs identify potential issues, such as Deficient or unprofessional review segments, within individual paper reviews? To our knowledge, this is the first work to provide such a comprehensive analysis.