Search

We introduce a new benchmark for reward models that quantifies their ability to produce strong language models through RLHF (Reinforcement Learning from Human Feedback). The gold-standard approach is to run a full RLHF training pipeline and directly probe downstream LLM performance. However, this process is prohibitively expensive. To address this, we build a predictive model of downstream LLM performance by evaluating the reward model on proxy tasks. These proxy tasks consist of a large-scale human preference and a verifiable correctness preference dataset, in which we measure 12 metrics across 12 domains. To investigate which reward model metrics are most correlated to gold-standard RLHF outcomes, we launch an end-to-end RLHF experiment on a large-scale crowdsourced human preference platform to view real reward model downstream performance as ground truth. Ultimately, we compile our data and findings into Preference Proxy Evaluations (PPE), the first reward model benchmark explicitly linked to post-RLHF real-world human preference performance, which we open-source for public use and further development. Our code and evaluations can be found at https://github.com/lmarena/PPE .

We study the ground-state phase diagram of a non-Hermitian cluster-XY spin chain in the language of free fermions. By calculating the second derivative of ground-state energy density and various types of order parameters, we establish the global ground-state phase diagram of the model, exhibiting rich quantum phases and corresponding phase transitions. Specially, the results reveal that the non-Hermitian cluster-XY model contains five different phases and two critical regions, i.e., ferromagnetic (FM), antiferromagnetic (AFM), symmetry-protected topological (SPT), paramagnetic (PM), Luttinger liquid-like phase, as well as critical region I and II. The order parameters and critical behaviors are investigated and the correctness of the theory is confirmed.

Data annotation refers to the labeling or tagging of textual data with relevant information. A large body of works have reported positive results on leveraging LLMs as an alternative to human annotators. However, existing studies focus on classic NLP tasks, and the extent to which LLMs as data annotators perform in domains requiring expert knowledge remains underexplored. In this work, we investigate comprehensive approaches across three highly specialized domains and discuss practical suggestions from a cost-effectiveness perspective. To the best of our knowledge, we present the first systematic evaluation of LLMs as expert-level data annotators., Comment: Accepted to WiML @ NeurIPS 2024 (extended version)

We calculate the energy spectrum of the S-wave doubly heavy tetraquark systems, including the $ QQ^{(\prime)}\bar q\bar q$, $QQ^{(\prime)}\bar s\bar q$, and $ QQ^{(\prime)}\bar s\bar s$ ($Q^{(\prime)}=b,c$ and $q=u,d$) systems within the constituent quark model. We use the complex scaling method to obtain bound states and resonant states simultaneously, and the Gaussian expansion method to solve the complex-scaled four-body Schr\"odinger equation. With a novel definition of the root-mean-square radii, we are able to distinguish between meson molecules and compact tetraquark states. The compact tetraquarks are further classified into three different types with distinct spatial configurations: compact even tetraquarks, compact diquark-antidiquark tetraquarks and compact diquark-centered tetraquarks. In the $ I(J^P)=0(1^+) $ $QQ\bar q\bar q$ system, there exists the $ D^*D $ molecular bound state with a binding energy of $ -14 $ MeV, which is the candidate for $ T_{cc}(3875)^+ $. The shallow $\bar B^*\bar B$ molecular bound state is the bottom analogue of $T_{cc}(3875)^+$. Moreover, we identify two resonant states near the $D^*D^*$ and $\bar B^*\bar B^*$ thresholds. In the $ J^P=1^+ $ $bb\bar q\bar q\,(I=0)$ and $bb\bar s\bar q$ systems, we obtain deeply bound states with a compact diquark-centered tetraquark configuration and a dominant $\chi_{\bar 3_c\otimes 3_c}$ component, along with resonant states with similar configurations as their radial excitations. These states are the QCD analogue of the helium atom. We also obtain some other bound states and resonant states with ``QCD Hydrogen molecule" configurations. Moreover, we investigate the heavy quark mass dependence of the $ I(J^P)=0(1^+) $ $ QQ\bar q\bar q $ bound states. We strongly urge the experimental search for the predicted states., Comment: 17 pages, 13 figures

We determine the Dirac series of $U(n,2)$ and give related information about Dirac series: the spin lowest $K$-types are multiplicity free, and the Dirac cohomology are consistent with the Dirac index., Comment: 23 pages

We conduct systematic calculations of the S-wave fully strange systems with ``normal" $\left(J^{P C}=0^{++}, 1^{+-}, 2^{++}\right)$ and ``exotic" $\left(J^{P C}=0^{+-}, 1^{++}, 2^{+-}\right)$ C-parities, which are the strange analogue of the fully charmed tetraquark state $X(6900)$. Within a constituent quark potential model, we employ the Gaussian expansion method to solve the four-body Schr\"odinger equation and the complex scaling method to identify resonant states. We obtain a series of resonant states and zero-width states in the mass range of 2.7 to 3.3 GeV, with their widths ranging from less than 1 MeV to about 50 MeV. Their rms radii strongly indicate that they are compact tetraquark states. Among these states, the $T_{4s,2^{++}}(2714)$ may be the most likely one to be observed experimentally. We urge the experimental exploration of the $2^{++}$ $s s \bar{s} \bar{s}$ state around 2.7 GeV in the $\phi\phi$ channel. Since the lowest S-wave $s s \bar{s} \bar{s}$ state is around 2.7 GeV, the compact P-wave $s s \bar{s} \bar{s}$ states are expected to be heavier. Hence, $\phi(2170)$ and $X(2370)$ are unlikely to be compact tetraquark states., Comment: 9 pages, 2 figures. Comments are welcomed

The 1st Workshop on Data Contamination (CONDA 2024) focuses on all relevant aspects of data contamination in natural language processing, where data contamination is understood as situations where evaluation data is included in pre-training corpora used to train large scale models, compromising evaluation results. The workshop fostered a shared task to collect evidence on data contamination in current available datasets and models. The goal of the shared task and associated database is to assist the community in understanding the extent of the problem and to assist researchers in avoiding reporting evaluation results on known contaminated resources. The shared task provides a structured, centralized public database for the collection of contamination evidence, open to contributions from the community via GitHub pool requests. This first compilation paper is based on 566 reported entries over 91 contaminated sources from a total of 23 contributors. The details of the individual contamination events are available in the platform. The platform continues to be online, open to contributions from the community., Comment: https://huggingface.co/spaces/CONDA-Workshop/Data-Contamination-Database

Large language models (LLMs) exhibit impressive capabilities across a wide range of tasks, yet the choice of which model to use often involves a trade-off between performance and cost. More powerful models, though effective, come with higher expenses, while less capable models are more cost-effective. To address this dilemma, we propose several efficient router models that dynamically select between a stronger and a weaker LLM during inference, aiming to optimize the balance between cost and response quality. We develop a training framework for these routers leveraging human preference data and data augmentation techniques to enhance performance. Our evaluation on widely-recognized benchmarks shows that our approach significantly reduces costs-by over 2 times in certain cases-without compromising the quality of responses. Interestingly, our router models also demonstrate significant transfer learning capabilities, maintaining their performance even when the strong and weak models are changed at test time. This highlights the potential of these routers to provide a cost-effective yet high-performance solution for deploying LLMs.

We use the quark potential model to calculate the mass spectrum of the S-wave fully heavy tetraquark systems with different flavors, including the $ bc\bar b\bar c, bb\bar c\bar c, cc\bar c\bar b $ and $ bb\bar b\bar c $ systems. We employ the Gaussian expansion method to solve the four-body Schr\"odinger equation, and the complex scaling method to identify resonant states. The $ bc\bar b\bar c, bb\bar c\bar c, cc\bar c\bar b $ and $ bb\bar b\bar c $ resonant states are obtained in the mass regions of $ (13.2,13.5) $, $ (13.3,13.6) $, $ (10.0,10.3) $, $ (16.5,16.7) $ GeV, respectively. Among these states, the $ bc\bar b\bar c $ tetraquark states are the most promising ones to be discovered in the near future. We recommend the experimental exploration of the $ 1^{++} $ and $ 2^{++} $ $ bc\bar b\bar c $ states with masses near $ 13.3 $ GeV in the $ J/\psi\Upsilon $ channel. From the root-mean-square radii, we find that all the resonant states we have identified are compact tetraquark states., Comment: 10 pages,7 figures,8 tables

Retrieval-augmented generation (RAG) has shown promising potential to enhance the accuracy and factuality of language models (LMs). However, imperfect retrievers or noisy corpora can introduce misleading or even erroneous information to the retrieved contents, posing a significant challenge to the generation quality. Existing RAG methods typically address this challenge by directly predicting final answers despite potentially noisy inputs, resulting in an implicit denoising process that is difficult to interpret and verify. On the other hand, the acquisition of explicit denoising supervision is often costly, involving significant human efforts. In this work, we propose InstructRAG, where LMs explicitly learn the denoising process through self-synthesized rationales -- First, we instruct the LM to explain how the ground-truth answer is derived from retrieved documents. Then, these rationales can be used either as demonstrations for in-context learning of explicit denoising or as supervised fine-tuning data to train the model. Compared to standard RAG approaches, InstructRAG requires no additional supervision, allows for easier verification of the predicted answers, and effectively improves generation accuracy. Experiments show InstructRAG consistently outperforms existing RAG methods in both training-free and trainable scenarios, achieving a relative improvement of 8.3% over the best baseline method on average across five knowledge-intensive benchmarks. Extensive analysis indicates that InstructRAG scales well with increased numbers of retrieved documents and consistently exhibits robust denoising ability even in out-of-domain datasets, demonstrating strong generalizability., Comment: Code: https://github.com/weizhepei/InstructRAG

The rapid evolution of Large Language Models (LLMs) has outpaced the development of model evaluation, highlighting the need for continuous curation of new, challenging benchmarks. However, manual curation of high-quality, human-aligned benchmarks is expensive and time-consuming. To address this, we introduce BenchBuilder, an automated pipeline that leverages LLMs to curate high-quality, open-ended prompts from large, crowd-sourced datasets, enabling continuous benchmark updates without human in the loop. We apply BenchBuilder to datasets such as Chatbot Arena and WildChat-1M, extracting challenging prompts and utilizing LLM-as-a-Judge for automatic model evaluation. To validate benchmark quality, we propose new metrics to measure a benchmark's alignment with human preferences and ability to separate models. We release Arena-Hard-Auto, a benchmark consisting 500 challenging prompts curated by BenchBuilder. Arena-Hard-Auto provides 3x higher separation of model performances compared to MT-Bench and achieves 98.6% correlation with human preference rankings, all at a cost of $20. Our work sets a new framework for the scalable curation of automated benchmarks from extensive data.

The concept of persona, originally adopted in dialogue literature, has re-surged as a promising framework for tailoring large language models (LLMs) to specific context (e.g., personalized search, LLM-as-a-judge). However, the growing research on leveraging persona in LLMs is relatively disorganized and lacks a systematic taxonomy. To close the gap, we present a comprehensive survey to categorize the current state of the field. We identify two lines of research, namely (1) LLM Role-Playing, where personas are assigned to LLMs, and (2) LLM Personalization, where LLMs take care of user personas. Additionally, we introduce existing methods for LLM personality evaluation. To the best of our knowledge, we present the first survey for role-playing and personalization in LLMs under the unified view of persona. We continuously maintain a paper collection to foster future endeavors: https://github.com/MiuLab/PersonaLLM-Survey, Comment: Accepted to EMNLP 2024 Findings

Large Language Models (LLMs) require careful safety alignment to prevent malicious outputs. While significant research focuses on mitigating harmful content generation, the enhanced safety often come with the side effect of over-refusal, where LLMs may reject innocuous prompts and become less helpful. Although the issue of over-refusal has been empirically observed, a systematic measurement is challenging due to the difficulty of crafting prompts that appear harmful but are benign. This study proposes a novel method for automatically generating large-scale sets of "seemingly toxic prompts" (benign prompts likely rejected by LLMs). Leveraging this technique, we introduce OR-Bench, the first large-scale over-refusal benchmark. OR-Bench comprises 80,000 seemingly toxic prompts across 10 common rejection categories, a subset of around 1,000 hard prompts that are challenging even for state-of-the-art LLMs, and an additional 600 toxic prompts to prevent indiscriminate responses. We then conduct a comprehensive study to measure the over-refusal of 25 popular LLMs across 8 model families. Our datasets are available at https://huggingface.co/datasets/bench-llm/or-bench and the demo can be found at https://huggingface.co/spaces/bench-llm/or-bench. We hope this benchmark can help the community develop better safety aligned models., Comment: version 2, 10 pages main, 22 pages total

Large language models (LLMs) are increasingly integrated into many online services, yet they remain cost-prohibitive to deploy due to the requirement of expensive GPU instances. Prior work has addressed the high cost of LLM serving by improving the inference engine, but less attention has been given to selecting the most cost-efficient GPU type(s) for a specific LLM service. There is a large and growing landscape of GPU types and, within these options, higher cost does not always lead to increased performance. Instead, through a comprehensive investigation, we find that three key LLM service characteristics (request size, request rate, SLO) strongly influence GPU cost efficiency, and differing GPU types are most cost efficient for differing LLM service settings. As a result, the most cost-efficient allocation for a given service is typically a mix of heterogeneous GPU types. Based on this analysis, we introduce M\'elange, a GPU allocation framework that navigates these diverse LLM service characteristics and heterogeneous GPU option space to automatically and efficiently derive the minimal-cost GPU allocation for a given LLM service. We formulate the GPU allocation task as a cost-aware bin packing problem where GPUs are bins and items are slices of the service workload. Our formulation's constraints account for a service's unique characteristics, allowing M\'elange to be flexible to support diverse service settings and heterogeneity-aware to adapt the GPU allocation to a specific service. Compared to using only a single GPU type, M\'elange reduces deployment costs by up to 77% in conversational settings, 33% in document-based settings, and 51% in a mixed setting.

The evaluation of large language models (LLMs) has drawn substantial attention in the field recently. This work focuses on evaluating LLMs in a Chinese context, specifically, for Traditional Chinese which has been largely underrepresented in existing benchmarks. We present TMLU, a holistic evaluation suit tailored for assessing the advanced knowledge and reasoning capability in LLMs, under the context of Taiwanese Mandarin. TMLU consists of an array of 37 subjects across social science, STEM, humanities, Taiwan-specific content, and others, ranging from middle school to professional levels. In addition, we curate chain-of-thought-like few-shot explanations for each subject to facilitate the evaluation of complex reasoning skills. To establish a comprehensive baseline, we conduct extensive experiments and analysis on 24 advanced LLMs. The results suggest that Chinese open-weight models demonstrate inferior performance comparing to multilingual proprietary ones, and open-weight models tailored for Taiwanese Mandarin lag behind the Simplified-Chinese counterparts. The findings indicate great headrooms for improvement, and emphasize the goal of TMLU to foster the development of localized Taiwanese-Mandarin LLMs. We release the benchmark and evaluation scripts for the community to promote future research., Comment: Preprint. Under review

The long-range interaction can fundamentally alter properties in gapped topological phases such as emergent massive edge modes. However, recent research has shifted attention to topological nontrivial critical points or phases, and it is natural to explore how long-range interaction influences them. In this work, we investigate the topological behavior and phase transition of extended Kitaev chains with long-range interactions, which can be derived from the critical Ising model via the Jordan-Wigner transformation in the short-range limit. Specifically, we analytically find the critical edge modes at the critical point remain stable against long-range interaction. More importantly, we observe these critical edge modes remain massless even when long-range interactions become substantially strong. As a byproduct, we numerically find that the critical behavior of the long-range model belongs to the free Majorana fermion universality class, which is entirely different from the long-range universality class in usual long-range spin models. Our work could shed new light on the interplay between long-range interactions (frustrated) and the gapless topological phases of matter., Comment: 13 pages, 11 figures. Any comments or suggestions are welcome !

By constructing an exactly solvable spin model, we investigate the critical behaviors of transverse field Ising chains interpolated with cluster interactions, which exhibit various types of topologically distinct Ising critical points. Using fidelity susceptibility as an indicator, we establish the global phase diagram, including ferromagnetic, trivial paramagnetic, and symmetry-protected topological phases. Different types of critical points exist between these phases, encompassing both topologically trivial and non-trivial Ising critical points, as well as Gaussian critical points. Importantly, we demonstrate the existence of a Lifshitz transition between these topologically distinct Ising critical points, with central charge and critical exponents determined through finite-size scaling. This work serves as a valuable reference for further research on phase transitions within the gapless quantum phase of matter., Comment: 13 pages, 9 figures. Any comments or suggestions are welcome!

Large Language Models (LLMs) have unlocked new capabilities and applications; however, evaluating the alignment with human preferences still poses significant challenges. To address this issue, we introduce Chatbot Arena, an open platform for evaluating LLMs based on human preferences. Our methodology employs a pairwise comparison approach and leverages input from a diverse user base through crowdsourcing. The platform has been operational for several months, amassing over 240K votes. This paper describes the platform, analyzes the data we have collected so far, and explains the tried-and-true statistical methods we are using for efficient and accurate evaluation and ranking of models. We confirm that the crowdsourced questions are sufficiently diverse and discriminating and that the crowdsourced human votes are in good agreement with those of expert raters. These analyses collectively establish a robust foundation for the credibility of Chatbot Arena. Because of its unique value and openness, Chatbot Arena has emerged as one of the most referenced LLM leaderboards, widely cited by leading LLM developers and companies. Our demo is publicly available at \url{https://chat.lmsys.org}.

Affective computing, which aims to recognize, interpret, and understand human emotions, provides benefits in healthcare, such as improving patient care and enhancing doctor-patient communication. However, there is a noticeable absence of a comprehensive summary of recent advancements in affective computing for healthcare, which could pose difficulties for researchers entering this field. To address this, our paper aims to provide an extensive literature review of related studies published in the last five years. We begin by analyzing trends, benefits, and limitations of recent datasets and affective computing methods devised for healthcare. Subsequently, we highlight several healthcare application hotspots of current technologies that could be promising for real-world deployment. Through our analysis, we identify and discuss some ongoing challenges in the field as evidenced by the literature. Concluding with a thorough review, we further offer potential future research directions and hope our findings and insights could guide related researchers to make better contributions to the evolution of affective computing in healthcare.

We calculate the mass spectrum of the S-wave fully heavy tetraquark systems $ QQ\bar Q\bar Q~(Q=c,b) $ with both normal $ (J^{PC}=0^{++},1^{+-},2^{++}) $ and exotic $ (J^{PC}=0^{+-},1^{++},2^{+-}) $ C-parities using three different quark potential models (AL1, AP1, BGS). The exotic C-parity systems refer to the ones that cannot be composed of two S-wave ground heavy quarkonia. We incorporate the molecular dimeson and compact diquark-antidiquark spatial correlations simultaneously, thereby discerning the actual configurations of the states. We employ the Gaussian expansion method to solve the four-body Schr\"odinger equation, and the complex scaling method to identify the resonant states. The mass spectra in three different models qualitatively agree with each other. We obtain several resonant states with $ J^{PC} = 0^{++}, 1^{+-}, 2^{++}, 1^{++} $ in the mass region $(6.92,7.30)\, \mathrm{GeV}$, some of which are good candidates of the experimentally observed $X(6900)$ and $X(7200)$. We also obtain several exotic C-parity zero-width states with $ J^{PC}=0^{+-} $ and $ 2^{+-} $. These zero-width states have no corresponding S-wave diquarkonium threshold and can only decay strongly to final states with P-wave quarkonia. With the notation $T_{4Q,J(C)}(M)$, we deduce from the root mean square radii that the $ X(7200) $ candidates $ T_{4c,0(+)}(7173), T_{4c,2(+)}(7214) $ and the state $ T_{4c,1(-)}(7191) $ look like molecular states although most of the resonant and zero-width states are compact states., Comment: 16 pages, 6 figures, 10 tables. Version accepted by PRD

Large language models are increasingly trained on all the data ever produced by humans. Many have raised concerns about the trustworthiness of public benchmarks due to potential contamination in pre-training or fine-tuning datasets. While most data decontamination efforts apply string matching (e.g., n-gram overlap) to remove benchmark data, we show that these methods are insufficient, and simple variations of test data (e.g., paraphrasing, translation) can easily bypass these decontamination measures. Furthermore, we demonstrate that if such variation of test data is not eliminated, a 13B model can easily overfit a test benchmark and achieve drastically high performance, on par with GPT-4. We validate such observations in widely used benchmarks such as MMLU, GSK8k, and HumanEval. To address this growing risk, we propose a stronger LLM-based decontamination method and apply it to widely used pre-training and fine-tuning datasets, revealing significant previously unknown test overlap. For example, in pre-training sets such as RedPajama-Data-1T and StarCoder-Data, we identified that 8-18\% of the HumanEval benchmark overlaps. Interestingly, we also find such contamination in synthetic dataset generated by GPT-3.5/4, suggesting a potential risk of unintentional contamination. We urge the community to adopt stronger decontamination approaches when using public benchmarks. Moreover, we call for the community to actively develop fresh one-time exams to evaluate models accurately. Our decontamination tool is publicly available at https://github.com/lm-sys/llm-decontaminator.