Jiaming Ji, Donghai Hong, Borong Zhang, Boyuan Chen, Josef Dai, Boren Zheng, Tianyi Alex Qiu, Jiayi Zhou, Kaile Wang, Boxun Li, Sirui Han, Yike Guo, Yaodong Yang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Jiaming Ji, Kaile Wang, Tianyi Alex Qiu, Boyuan Chen, Jiayi Zhou, Changye Li, Hantao Lou, Josef Dai, Yunhuai Liu, Yaodong Yang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The training and deployment of large language models (LLMs) create a feedback loop with human users: models learn human beliefs from data, reinforce these beliefs with generated content, reabsorb the reinforced beliefs, and feed them back to users again and again. This dynamic resembles an echo chamber. We hypothesize that this feedback loop entrenches the existing values and beliefs of users, leading to a loss of diversity in human ideas and potentially the lock-in of false beliefs. We formalize this hypothesis and test it empirically with agent-based LLM simulations and real-world GPT usage data. Analysis reveals sudden but sustained drops in diversity after the release of new GPT iterations, consistent with the hypothesized human-AI feedback loop. Website: https://thelockinhypothesis.com

Recent advances in reasoning techniques have substantially improved the performance of large language models (LLMs), raising expectations for their ability to provide accurate, truthful, and reliable information. However, emerging evidence suggests that iterative reasoning may foster belief entrenchment, rather than enhancing truth-seeking behavior. In this study, we propose a systematic evaluation framework for *belief entrenchment* in LLM reasoning by leveraging the Martingale property from Bayesian statistics. This property implies that, under rational belief updating, the expected value of future beliefs should remain equal to the current belief, i.e., belief updates cannot be predicted from solely the current belief. We propose the unsupervised, regression-based *Martingale Score* to measure violations of this property, signaling a deviation from the Bayesian ability of updating on new evidence. In open-ended problem domains, including event forecasting, value-laden questions, and academic paper review, we found such violations to be widespread across models, reasoning paradigms, problem domains, and system prompts, where the future beliefs are consistently predictable from the model's current belief, a phenomenon which we term *belief entrenchment*. Through comprehensive experiments, we identify the models (e.g., GPT-4o), reasoning techniques (e.g., chain of thought), and domains (e.g., forecasting) more prone to belief entrenchment. Finally, we validate the Martingale Score by showing that it predicts ground-truth accuracy on problem domains where ground truth labels are available. This indicates that, while designed as an unsupervised metric that operates even in domains without access to ground truth, the Martingale Score is a useful proxy of the truth-seeking ability of the LLM reasoning process.

Vaccine development is one of the most promising fields, and multi-epitope vaccine, which does not need laborious culture processes, is an attractive alternative to classical vaccines with the advantage of safety, and efficiency. The rapid development of algorithms and the accumulation of immune data have facilitated the advancement of computer-aided vaccine design. Here we systemically reviewed the in silico data and algorithms resource, for different steps of computational vaccine design, including immunogen selection, epitope prediction, vaccine construction, optimization, and evaluation. The performance of different available tools on epitope prediction and immunogenicity evaluation was tested and compared on benchmark datasets. Finally, we discuss the future research direction for the construction of a multiepitope vaccine.

Nuclear receptors (NRs) are a class of essential proteins that regulate the expression of specific genes and are associated with multiple diseases. In silico methods for prescreening potential NR binders with predictive binding ability are highly desired for NR-related drug development but are rarely reported. Here, we present the PbsNRs (Predicting binders and scaffolds for Nuclear Receptors), a user-friendly web server designed to predict the potential NR binders and scaffolds through proteochemometric modeling. The utility of PbsNRs was systemically evaluated using both chemical compounds and natural products. Results indicated that PbsNRs achieved a good prediction performance for chemical compounds on internal (ROC-AUC = 0.906, where ROC is Receiver-Operating Characteristic curve and AUC is the Area Under the Curve) and external (ROC-AUC = 0.783) datasets, outperforming both compound-ligand interaction tools and NR-specific predictors. PbsNRs also successfully identified bioactive chemical scaffolds for NRs by screening massive natural products. Moreover, the predicted bioactive and inactive natural products for NR2B1 were experimentally validated using biosensors. PbsNRs not only aids in screening potential therapeutic NR binders but also helps discover the essential molecular scaffold and guide the drug discovery for multiple NR-related diseases. The PbsNRs web server is available at http://pbsnrs.badd-cao.net.

With the rapid development of large language models (LLMs) and ever-evolving practical requirements, finding an efficient and effective alignment method has never been more critical. However, the tension between the complexity of current alignment methods and the need for rapid iteration in deployment scenarios necessitates the development of a model-agnostic alignment approach that can operate under these constraints. In this paper, we introduce Aligner, a novel and simple alignment paradigm that learns the correctional residuals between preferred and dispreferred answers using a small model. Designed as a model-agnostic, plug-and-play module, Aligner can be directly applied to various open-source and API-based models with only one-off training, making it suitable for rapid iteration. Notably, Aligner can be applied to any powerful, large-scale upstream models. Moreover, it can even iteratively bootstrap the upstream models using corrected responses as synthetic human preference data, breaking through the model's performance ceiling. Our experiments demonstrate performance improvements by deploying the same Aligner model across 11 different LLMs, evaluated on the 3H dimensions (helpfulness, harmlessness, and honesty). Specifically, Aligner-7B has achieved an average improvement of 68.9% in helpfulness and 22.8% in harmlessness across the tested LLMs while also effectively reducing hallucination. In the Alpaca-Eval leaderboard, stacking Aligner-2B on GPT-4 Turbo improved its LC Win Rate from 55.0% to 58.3%, surpassing GPT-4 Omni's 57.5% Win Rate (community report).

Frontier AI systems, including large language models (LLMs), hold increasing influence over the epistemology of human users. Such influence can reinforce prevailing societal values, potentially contributing to the lock-in of misguided moral beliefs and, consequently, the perpetuation of problematic moral practices on a broad scale. We introduce progress alignment as a technical solution to mitigate this imminent risk. Progress alignment algorithms learn to emulate the mechanics of human moral progress, thereby addressing the susceptibility of existing alignment methods to contemporary moral blindspots. To empower research in progress alignment, we introduce ProgressGym, an experimental framework allowing the learning of moral progress mechanics from history, in order to facilitate future progress in real-world moral decisions. Leveraging 9 centuries of historical text and 18 historical LLMs, ProgressGym enables codification of real-world progress alignment challenges into concrete benchmarks. Specifically, we introduce three core challenges: tracking evolving values (PG-Follow), preemptively anticipating moral progress (PG-Predict), and regulating the feedback loop between human and AI value shifts (PG-Coevolve). Alignment methods without a temporal dimension are inapplicable to these tasks. In response, we present lifelong and extrapolative algorithms as baseline methods of progress alignment, and build an open leaderboard soliciting novel algorithms and challenges.

Phages can specifically recognize and kill bacteria, which lead to important application value of bacteriophage in bacterial identification and typing, livestock aquaculture and treatment of human bacterial infection. Considering the variety of human-infected bacteria and the continuous discovery of numerous pathogenic bacteria, screening suitable therapeutic phages that are capable of infecting pathogens from massive phage databases has been a principal step in phage therapy design. Experimental methods to identify phage-host interaction (PHI) are time-consuming and expensive; high-throughput computational method to predict PHI is therefore a potential substitute. Here, we systemically review bioinformatic methods for predicting PHI, introduce reference databases and in silico models applied in these methods and highlight the strengths and challenges of current tools. Finally, we discuss the application scope and future research direction of computational prediction methods, which contribute to the performance improvement of prediction models and the development of personalized phage therapy.

Bacteriophages can help the treatment of bacterial infections yet require in-silico models to deal with the great genetic diversity between phages and bacteria. Despite the tolerable prediction performance, the application scope of current approaches is limited to the prediction at the species level, which cannot accurately predict the relationship of phages across strain mutants. This has hindered the development of phage therapeutics based on the prediction of phage-bacteria relationships. In this paper, we present, PB-LKS, to predict the phage-bacteria interaction based on local K-mer strategy with higher performance and wider applicability. The utility of PB-LKS is rigorously validated through (i) large-scale historical screening, (ii) case study at the class level and (iii) in vitro simulation of bacterial antiphage resistance at the strain mutant level. The PB-LKS approach could outperform the current state-of-the-art methods and illustrate potential clinical utility in pre-optimized phage therapy design.

Foot-and-mouth disease virus (FMDV) is an antigenic-variable RNA virus that is responsible for the recurrence of foot-and-mouth disease in livestock and can be prevented and controlled using a vaccine with broad-spectrum protection. Current anti-genicity evaluation methods, which involve animal immunity experiments and serum preparation, are unable to fulfill the needs of high-throughput antigenicity measurements. This study designed an antigenicity scoring model to rapidly predict the antigenicity of FMDV. Antigenic-dominant sites were initially determined on the VP1 protein, a position-specific scoring matrix and physical chemical indexes were integrated to generate antigenicity descriptors. Independent tests showed a high accuracy of 0.848 and an AUC value of 0.889, indicating the good performance of the model in antigenicity measurement. When applying this model to historical data, annual antigenicity coverage of widely used vaccine strains was successfully evaluated, this was also supported by previous experiments. Furthermore, the utility of this model was extended to select potential broad-spectrum vaccines among 1,201 historical non-redundant strains to recommend potential univalent, bivalent and trivalent vaccine candidates. The results suggested that the computational model designed in this study could be used for the high-throughput antigenicity measurement of FMDV and could aid in vaccine development for preventing FMDV epidemics.

As an extension of the conventional quantitative structure activity relationship models, proteochemometric (PCM) modelling is a computational method that can predict the bioactivity relations between multiple ligands and multiple targets. Traditional PCM modelling includes three essential elements: descriptors (including target descriptors, ligand descriptors and cross-term descriptors), bioactivity data and appropriate learning functions that link the descriptors to the bioactivity data. Since its appearance, PCM modelling has developed rapidly over the past decade by taking advantage of the progress of different descriptors and machine learning techniques, along with the increasing amounts of available bioactivity data. Specifically, the new emerging target descriptors and cross-term descriptors not only significantly increased the performance of PCM modelling but also expanded its application scope from traditional protein-ligand interaction to more abundant interactions, including protein-peptide, protein-DNA and even protein-protein interactions. In this review, target descriptors and cross-term descriptors, as well as the corresponding application scope, are intensively summarized. Additionally, we look forward to seeing PCM modelling extend into new application scopes, such as Target-Catalyst-Ligand systems, with the further development of descriptors, machine learning techniques and increasing amounts of available bioactivity data.