Honglin Bao

Yiyou Sun, Xinyang Han, Weichen Zhang et al.

Recent AI systems have achieved strong results on a wide range of benchmarks, yet these gains have not translated into economically meaningful deployment across many professional domains. We argue that this gap is largely an evaluation problem: widely used benchmarks lack sustained performance measurement on real and economically valuable workflows. This paper introduces Agents' Last Exam (ALE), a benchmark designed to evaluate AI agents on long-horizon, economically valuable, real-world tasks with verifiable outcomes. Developed in collaboration with 250+ industry experts, ALE covers non-physical industries defined with reference to O*NET / SOC 2018 (the U.S. federal occupational taxonomy). It is organized around a task taxonomy with 55 subfields grouped into 13 industry clusters covering 1K+ tasks. Current results show that the hardest tier remains far from saturated: across mainstream harness and backbone configurations, the average full pass rate is 2.6%. ALE is designed as a living benchmark: its task pool grows continuously as new workflows and industries are onboarded. More broadly, ALE is intended not merely as another leaderboard, but as an instrument for closing the gap between benchmark success and GDP-relevant impact.

Zehan Li, Yutong Zhu, Siyang Wu et al.

Large language models produce fluent fiction, yet their creative output is widely seen as flat. We ask where this quality originates in the training and whether it affects different domains of human fiction equally. We construct a matched story-continuation paradigm across StoryStar (public-platform), TMAS (prompt-guided), and The New Yorker (professional literary)-and compare continuations from four OLMo 32B checkpoints (Base, SFT, DPO, RLVR) against matched human text. Because these checkpoints share architecture, scale, tokenizer, and pretraining, the design isolates the post-training effect. We measure each continuation along three sentence-level dimensions: thematic motion, affective prevalence, and linguistic diversity. Across all three, post-training compresses dynamic variation: thematic transitions become more uniform, high-intensity emotions give way to neutrality, and stylistic diversity across stories shrinks. We term this progressive loss narrative flattening. The effect is directionally stable across story domains but gap size depends on the human baseline: professional literary fiction is compressed most, while public-platform and prompt-guided stories show smaller gaps, consistent with their human baselines sitting closer to the model's default rhythm. Post-trained endpoints converge across domains, suggesting alignment produces a continuation regime largely insensitive to the source domain's narrative texture.

Jiawei Zhang, Honglin Bao, Pengda Wang et al.

Social and behavioral science runs thousands of experiments each year, yet their findings rarely accumulate into a coherent map of what is known, what conflicts, and what remains missing. We introduce ExAtlas, a framework for turning an archive of experiments into an atlas: a structured map in which studies link, conflict, or leave bridgeable gaps. Given a target study, ExAtlas searches for prior studies that are locally close in treatment and outcome space and asks whether their observed effects can be composed to predict the target effect. This yields three cases. If the composition succeeds and agrees with the observed result, ExAtlas links the target to consistent prior evidence. If composition succeeds but disagrees, ExAtlas reconciles the conflict and proposes candidate moderators or higher-level theories that could explain it. If composition fails, ExAtlas proposes bridge experiments to close the gap. We provide an error bound for composition under local smoothness of the treatment-effect surface. On held-out targets certified as locally supported, ExAtlas recovers effect direction in 98.6% of cases. Human evaluations further suggest that its proposed bridge experiments are plausible and exhibit connectedness, and that its conflict explanations are useful for theory generation. These results suggest that the archive of social experiments contains more latent structure than current practice extracts -- and that making this structure explicit can guide both future theory and future experimentation.

Honglin Bao, Siyang Wu, Jiwoong Choi et al.

This paper calls on the research community not only to investigate how human biases are inherited by large language models (LLMs) but also to explore how these biases in LLMs can be leveraged to make society's "unwritten code" - such as implicit stereotypes and heuristics - visible and accessible for critique. We introduce a conceptual framework through a case study in science: uncovering hidden rules in peer review - the factors that reviewers care about but rarely state explicitly due to normative scientific expectations. The idea of the framework is to push LLMs to speak out their heuristics through generating self-consistent hypotheses - why one paper appeared stronger in reviewer scoring - among paired papers submitted to 45 academic conferences, while iteratively searching deeper hypotheses from remaining pairs where existing hypotheses cannot explain. We observed that LLMs' normative priors about the internal characteristics of good science extracted from their self-talk, e.g., theoretical rigor, were systematically updated toward posteriors that emphasize storytelling about external connections, such as how the work is positioned and connected within and across literatures. Human reviewers tend to explicitly reward aspects that moderately align with LLMs' normative priors (correlation = 0.49) but avoid articulating contextualization and storytelling posteriors in their review comments (correlation = -0.14), despite giving implicit reward to them with positive scores. These patterns are robust across different models and out-of-sample judgments. We discuss the broad applicability of our proposed framework, leveraging LLMs as diagnostic tools to amplify and surface the tacit codes underlying human society, enabling public discussion of revealed values and more precisely targeted responsible AI.

Siyang Wu, Honglin Bao, Sida Li et al.

We develop signatures of capacity familiarity to characterize large language model (LLM) benchmarks and their meaningful overlaps. Benchmark signatures probe the capacity required for benchmark performance. We formally define them as a set of salient tokens drawn from in-the-wild, naturally authored corpora, where LLM token perplexity, reflecting more or less pre-training exposure, becomes highly predictive of LLM benchmark performance. Through a large-scale meta-evaluation, we extract benchmark signatures via stepwise forward selection with linear regressions across 32 LLMs and 88 benchmarks spanning diverse knowledge, coding, logic, instruction following, math, language, reasoning, and world modeling. Our analysis situates signatures in relation to both the semantic similarity of benchmark questions and the correlation of model performance. While performance overlaps are universally high and semantic overlaps remain confined to a narrow mid-range, benchmark signatures prove highly informative in capturing variation, overlap, and divergence. We observe overlap in knowledge and reasoning subtasks, whereas multilingual and cultural benchmarks exhibit less similarity, even compared to cross-task overlap. Notably, performance-level results are strongly influenced by benchmark-orthogonal factors such as question format, highlighting limitations in LLM generalization, the conflation of performance with ability, and issues inherent in current mainstream benchmark agreement studies. Benchmark signatures, however, remain robust to such effects. Ultimately, we identify cross-functional overlaps across logic, math, language, instruction following, and world modeling, with coding emerging as the least overlapping domain. Together, these findings provide mechanistic insights into benchmark validity and LLM sensitivities, and sketch the underlying landscape of interconnected LLM capabilities.

Honglin Bao, Kai Li

Persistence is often regarded as a virtue in science. In this paper, however, we challenge this conventional view by highlighting its contextual nature, particularly how persistence can become a liability during periods of paradigm shift. We focus on the deep learning revolution catalyzed by AlexNet in 2012. Analyzing the 20-year career trajectories of over 5,000 scientists who were active in top machine learning venues during the preceding decade, we examine how their research focus and output evolved. We first uncover a dynamic period in which leading venues increasingly prioritized cutting-edge deep learning developments that displaced relatively traditional statistical learning methods. Scientists responded to these changes in markedly different ways. Those who were previously successful or affiliated with old teams adapted more slowly, experiencing what we term a rigidity penalty - a reluctance to embrace new directions leading to a decline in scientific impact, as measured by citation percentile rank. In contrast, scientists who pursued strategic adaptation - selectively pivoting toward emerging trends while preserving weak connections to prior expertise - reaped the greatest benefits. Taken together, our macro- and micro-level findings show that scientific breakthroughs act as mechanisms that reconfigure power structures within a field.

Siyang Wu, Honglin Bao, Nadav Kunievsky et al.

Large language models (LLMs) are rapidly becoming core tools for science, engineering, and innovation. Their promise lies not just in remembering facts, but in putting knowledge to work. Despite their impressive ability to answer increasingly difficult questions, it remains unclear whether LLMs truly use their knowledge when confronted with new and challenging tasks. We address this question with a patent classification task that requires deep conceptual understanding: distinguishing objectively different but semantically similar patents. To evaluate this approach, we introduce a challenging new benchmark of 1.3 million post-2015 computer science patent pairs, characterized by dense technical jargon and strategically complex writing. We find that LLMs often fail our benchmark and struggle to distinguish among semantically similar patents. To probe this failure, we introduce a novel framework that decomposes model errors into two sources: missing and unused knowledge. Our approach asks models to generate clarifying questions to improve their understanding, and then compares three settings: raw performance, self-answered questions, and externally supplied answers. This decomposition reveals that LLMs often possess the relevant knowledge internally but fail to deploy it, while a smaller share of errors arises from genuine knowledge gaps. We then ask whether the ability of models to construct a task-specific database of questions and answers differs across models. We find that smaller models generate simpler, broadly transferable questions, while larger models propose more complex but less generalizable ones. This suggests new strategies for combining strengths across models. Our findings highlight a critical limitation of current LLMs and their evaluation: models often know more than they can use. LLM evaluation should shift from recall of static facts to application of dynamic knowledge.

Honglin Bao, Wolfgang Banzhaf

The complexity of cooperative behavior is a crucial issue in multiagent-based social simulation. In this paper, an agent-based model is proposed to study the evolution of cooperative hunting behaviors in an artificial society. In this model, the standard hunting game of stag is modified into a new situation with social hierarchy and penalty. The agent society is divided into multiple layers with supervisors and subordinates. In each layer, the society is divided into multiple clusters. A supervisor controls all subordinates in a cluster locally. Subordinates interact with rivals through reinforcement learning, and report learning information to their corresponding supervisor. Supervisors process the reported information through repeated affiliation-based aggregation and by information exchange with other supervisors, then pass down the reprocessed information to subordinates as guidance. Subordinates, in turn, update learning information according to guidance, following the "win stay, lose shift" strategy. Experiments are carried out to test the evolution of cooperation in this closed-loop semi-supervised emergent system with different parameters. We also study the variations and phase transitions in this game setting.