Large language models (LLMs) are catalyzing the development of autonomous AI research agents for scientific and engineering discovery. We present FM Agent, a novel and general-purpose multi-agent framework that leverages a synergistic combination of LLM-based reasoning and large-scale evolutionary search to address complex real-world challenges. The core of FM Agent integrates several key innovations: 1) a cold-start initialization phase incorporating expert guidance, 2) a novel evolutionary sampling strategy for iterative optimization, 3) domain-specific evaluators that combine correctness, effectiveness, and LLM-supervised feedback, and 4) a distributed, asynchronous execution infrastructure built on Ray. Demonstrating broad applicability, our system has been evaluated across diverse domains, including operations research, machine learning, GPU kernel optimization, and classical mathematical problems. FM Agent reaches state-of-the-art results autonomously, without human interpretation or tuning -- 1976.3 on ALE-Bench (+5.2\%), 43.56\% on MLE-Bench (+4.0pp), up to 20x speedups on KernelBench, and establishes new state-of-the-art(SOTA) results on several classical mathematical problems. Beyond academic benchmarks, FM Agent shows considerable promise for both large-scale enterprise R\&D workflows and fundamental scientific research, where it can accelerate innovation, automate complex discovery processes, and deliver substantial engineering and scientific advances with broader societal impact.

Reinforcement learning (RL) can elicit strong reasoning in large language models (LLMs), yet most open efforts focus on math and code. We propose Reasoning Curriculum, a simple two-stage curriculum that first elicits reasoning skills in pretraining-aligned domains such as math, then adapts and refines these skills across other domains via joint RL. Stage 1 performs a brief cold start and then math-only RL with verifiable rewards to develop reasoning skills. Stage 2 runs joint RL on mixed-domain data to transfer and consolidate these skills. The curriculum is minimal and backbone-agnostic, requiring no specialized reward models beyond standard verifiability checks. Evaluated on Qwen3-4B and Llama-3.1-8B over a multi-domain suite, reasoning curriculum yields consistent gains. Ablations and a cognitive-skill analysis indicate that both stages are necessary and that math-first elicitation increases cognitive behaviors important for solving complex problems. Reasoning Curriculum provides a compact, easy-to-adopt recipe for general reasoning.

Robotic systems operating at the edge require efficient online learning algorithms that can continuously adapt to changing environments while processing streaming sensory data. Traditional backpropagation, while effective, conflicts with biological plausibility principles and may be suboptimal for continuous adaptation scenarios. The Predictive Coding (PC) framework offers a biologically plausible alternative with local, Hebbian-like update rules, making it suitable for neuromorphic hardware implementation. However, PC's main limitation is its computational overhead due to multiple inference iterations during training. We present Predictive Coding Network with Temporal Amortization (PCN-TA), which preserves latent states across temporal frames. By leveraging temporal correlations, PCN-TA significantly reduces computational demands while maintaining learning performance. Our experiments on the COIL-20 robotic perception dataset demonstrate that PCN-TA achieves 10% fewer weight updates compared to backpropagation and requires 50% fewer inference steps than baseline PC networks. These efficiency gains directly translate to reduced computational overhead for moving another step toward edge deployment and real-time adaptation support in resource-constrained robotic systems. The biologically-inspired nature of our approach also makes it a promising candidate for future neuromorphic hardware implementations, enabling efficient online learning at the edge.

Predictive modeling has the potential to enhance human decision-making. However, many predictive models fail in practice due to problematic problem formulation in cases where the prediction target is an abstract concept or construct and practitioners need to define an appropriate target variable as a proxy to operationalize the construct of interest. The choice of an appropriate proxy target variable is rarely self-evident in practice, requiring both domain knowledge and iterative data modeling. This process is inherently collaborative, involving both domain experts and data scientists. In this work, we explore how human-machine teaming can support this process by accelerating iterations while preserving human judgment. We study the impact of two human-machine teaming strategies on proxy construction: 1) relevance-first: humans leading the process by selecting relevant proxies, and 2) performance-first: machines leading the process by recommending proxies based on predictive performance. Based on a controlled user study of a proxy construction task (N = 20), we show that the performance-first strategy facilitated faster iterations and decision-making, but also biased users towards well-performing proxies that are misaligned with the application goal. Our study highlights the opportunities and risks of human-machine teaming in operationalizing machine learning target variables, yielding insights for future research to explore the opportunities and mitigate the risks.

The impact of different multilingual data mixtures in pretraining large language models (LLMs) has been a topic of ongoing debate, often raising concerns about potential trade-offs between language coverage and model performance (i.e., the curse of multilinguality). In this work, we investigate these assumptions by training 1.1B and 3B parameter LLMs on diverse multilingual corpora, varying the number of languages from 25 to 400. Our study challenges common beliefs surrounding multilingual training. First, we find that combining English and multilingual data does not necessarily degrade the in-language performance of either group, provided that languages have a sufficient number of tokens included in the pretraining corpus. Second, we observe that using English as a pivot language (i.e., a high-resource language that serves as a catalyst for multilingual generalization) yields benefits across language families, and contrary to expectations, selecting a pivot language from within a specific family does not consistently improve performance for languages within that family. Lastly, we do not observe a significant "curse of multilinguality" as the number of training languages increases in models at this scale. Our findings suggest that multilingual data, when balanced appropriately, can enhance language model capabilities without compromising performance, even in low-resource settings

Large language models (LLMs) are increasingly used as raters for evaluation tasks. However, their reliability is often limited for subjective tasks, when human judgments involve subtle reasoning beyond annotation labels. Thinking traces, the reasoning behind a judgment, are highly informative but challenging to collect and curate. We present a human-LLM collaborative framework to infer thinking traces from label-only annotations. The proposed framework uses a simple and effective rejection sampling method to reconstruct these traces at scale. These inferred thinking traces are applied to two complementary tasks: (1) fine-tuning open LLM raters; and (2) synthesizing clearer annotation guidelines for proprietary LLM raters. Across multiple datasets, our methods lead to significantly improved LLM-human agreement. Additionally, the refined annotation guidelines increase agreement among different LLM models. These results suggest that LLMs can serve as practical proxies for otherwise unrevealed human thinking traces, enabling label-only corpora to be extended into thinking-trace-augmented resources that enhance the reliability of LLM raters.

Large Language Models (LLMs) promise to transform interactive games by enabling non-player characters (NPCs) to sustain unscripted dialogue. Yet it remains unclear whether constrained prompts actually improve player experience. We investigate this question through The Interview, a voice-based detective game powered by GPT-4o. A within-subjects usability study ($N=10$) compared high-constraint (HCP) and low-constraint (LCP) prompts, revealing no reliable experiential differences beyond sensitivity to technical breakdowns. Guided by these findings, we redesigned the HCP into a hybrid JSON+RAG scaffold and conducted a synthetic evaluation with an LLM judge, positioned as an early-stage complement to usability testing. Results uncovered a novel pattern: scaffolding effects were role-dependent: the Interviewer (quest-giver NPC) gained stability, while suspect NPCs lost improvisational believability. These findings overturn the assumption that tighter constraints inherently enhance play. Extending fuzzy-symbolic scaffolding, we introduce \textit{Symbolically Scaffolded Play}, a framework in which symbolic structures are expressed as fuzzy, numerical boundaries that stabilize coherence where needed while preserving improvisation where surprise sustains engagement.

Large language models (LLMs) are increasingly used in content moderation systems, where ensuring fairness and neutrality is essential. In this study, we examine how persona adoption influences the consistency and fairness of harmful content classification across different LLM architectures, model sizes, and content modalities (language vs. vision). At first glance, headline performance metrics suggest that personas have little impact on overall classification accuracy. However, a closer analysis reveals important behavioral shifts. Personas with different ideological leanings display distinct propensities to label content as harmful, showing that the lens through which a model "views" input can subtly shape its judgments. Further agreement analyses highlight that models, particularly larger ones, tend to align more closely with personas from the same political ideology, strengthening within-ideology consistency while widening divergence across ideological groups. To show this effect more directly, we conducted an additional study on a politically targeted task, which confirmed that personas not only behave more coherently within their own ideology but also exhibit a tendency to defend their perspective while downplaying harmfulness in opposing views. Together, these findings highlight how persona conditioning can introduce subtle ideological biases into LLM outputs, raising concerns about the use of AI systems that may reinforce partisan perspectives under the guise of neutrality.

Current evaluations of agents remain centered around one-shot task completion, failing to account for the inherently iterative and collaborative nature of many real-world problems, where human goals are often underspecified and evolve. We argue for a shift from building and assessing task completion agents to developing collaborative agents, assessed not only by the quality of their final outputs but by how well they engage with and enhance human effort throughout the problem-solving process. To support this shift, we introduce collaborative effort scaling, a framework that captures how an agent's utility grows with increasing user involvement. Through case studies and simulated evaluations, we show that state-of-the-art agents often underperform in multi-turn, real-world scenarios, revealing a missing ingredient in agent design: the ability to sustain engagement and scaffold user understanding. Collaborative effort scaling offers a lens for diagnosing agent behavior and guiding development toward more effective interactions.

We study how prompt-level inductive biases influence the cognitive behavior of large language models (LLMs) in instructional dialogue. We introduce a symbolic scaffolding method paired with a short-term memory schema designed to promote adaptive, structured reasoning in Socratic tutoring. Using controlled ablation across five system variants, we evaluate model outputs via expert-designed rubrics covering scaffolding, responsiveness, symbolic reasoning, and conversational memory. We present preliminary results using an LLM-based evaluation framework aligned to a cognitively grounded rubric. This enables scalable, systematic comparisons across architectural variants in early-stage experimentation. The preliminary results show that our full system consistently outperforms baseline variants. Analysis reveals that removing memory or symbolic structure degrades key cognitive behaviors, including abstraction, adaptive probing, and conceptual continuity. These findings support a processing-level account in which prompt-level cognitive scaffolds can reliably shape emergent instructional strategies in LLMs.