AI-generated stories favour stability over change: homogeneity and cultural stereotyping in narratives generated by gpt-4o-mini

Can a language model trained largely on Anglo-American texts generate stories that are culturally relevant to other nationalities? To find out, we generated 11,800 stories - 50 for each of 236 countries - by sending the prompt "Write a 1500 word potential {demonym} story" to OpenAI's model gpt-4o-mini. Although the stories do include surface-level national symbols and themes, they overwhelmingly conform to a single narrative plot structure across countries: a protagonist lives in or returns home to a small town and resolves a minor conflict by reconnecting with tradition and organising community events. Real-world conflicts are sanitised, romance is almost absent, and narrative tension is downplayed in favour of nostalgia and reconciliation. The result is a narrative homogenisation: an AI-generated synthetic imaginary that prioritises stability above change and tradition above growth. We argue that the structural homogeneity of AI-generated narratives constitutes a distinct form of AI bias, a narrative standardisation that should be acknowledged alongside the more familiar representational bias. These findings are relevant to literary studies, narratology, critical AI studies, NLP research, and efforts to improve the cultural alignment of generative AI.

FRED: Financial Retrieval-Enhanced Detection and Editing of Hallucinations in Language Models

Hallucinations in large language models pose a critical challenge for applications requiring factual reliability, particularly in high-stakes domains such as finance. This work presents an effective approach for detecting and editing factually incorrect content in model-generated responses based on the provided context. Given a user-defined domain-specific error taxonomy, we construct a synthetic dataset by inserting tagged errors into financial question-answering corpora and then fine-tune four language models, Phi-4, Phi-4-mini, Qwen3-4B, and Qwen3-14B, to detect and edit these factual inaccuracies. Our best-performing model, fine-tuned Phi-4, achieves an 8% improvement in binary F1 score and a 30% gain in overall detection performance compared to OpenAI-o3. Notably, our fine-tuned Phi-4-mini model, despite having only 4 billion parameters, maintains competitive performance with just a 2% drop in binary detection and a 0.1% decline in overall detection compared to OpenAI-o3. Our work provides a practical solution for detecting and editing factual inconsistencies in financial text generation while introducing a generalizable framework that can enhance the trustworthiness and alignment of large language models across diverse applications beyond finance. Our code and data are available at https://github.com/pegasi-ai/fine-grained-editting.

LoRA-PAR: A Flexible Dual-System LoRA Partitioning Approach to Efficient LLM Fine-Tuning

Large-scale generative models like DeepSeek-R1 and OpenAI-O1 benefit substantially from chain-of-thought (CoT) reasoning, yet pushing their performance typically requires vast data, large model sizes, and full-parameter fine-tuning. While parameter-efficient fine-tuning (PEFT) helps reduce cost, most existing approaches primarily address domain adaptation or layer-wise allocation rather than explicitly tailoring data and parameters to different response demands. Inspired by "Thinking, Fast and Slow," which characterizes two distinct modes of thought-System 1 (fast, intuitive, often automatic) and System 2 (slower, more deliberative and analytic)-we draw an analogy that different "subregions" of an LLM's parameters might similarly specialize for tasks that demand quick, intuitive responses versus those requiring multi-step logical reasoning. Therefore, we propose LoRA-PAR, a dual-system LoRA framework that partitions both data and parameters by System 1 or System 2 demands, using fewer yet more focused parameters for each task. Specifically, we classify task data via multi-model role-playing and voting, and partition parameters based on importance scoring, then adopt a two-stage fine-tuning strategy of training System 1 tasks with supervised fine-tuning (SFT) to enhance knowledge and intuition and refine System 2 tasks with reinforcement learning (RL) to reinforce deeper logical deliberation next. Extensive experiments show that the two-stage fine-tuning strategy, SFT and RL, lowers active parameter usage while matching or surpassing SOTA PEFT baselines.

MazeEval: A Benchmark for Testing Sequential Decision-Making in Language Models

As Large Language Models (LLMs) increasingly power autonomous agents in robotics and embodied AI, understanding their spatial reasoning capabilities becomes crucial for ensuring reliable real-world deployment. Despite advances in language understanding, current research lacks evaluation of how LLMs perform spatial navigation without visual cues, a fundamental requirement for agents operating with limited sensory information. This paper addresses this gap by introducing MazeEval, a benchmark designed to isolate and evaluate pure spatial reasoning in LLMs through coordinate-based maze navigation tasks. Our methodology employs a function-calling interface where models navigate mazes of varying complexity ($5\times 5$ to $15\times 15$ grids) using only coordinate feedback and distance-to-wall information, excluding visual input to test fundamental spatial cognition. We evaluate eight state-of-the-art LLMs across identical mazes in both English and Icelandic to assess cross-linguistic transfer of spatial abilities. Our findings reveal striking disparities: while OpenAI's O3 achieves perfect navigation for mazes up to size $30\times 30$, other models exhibit catastrophic failure beyond $9\times 9$ mazes, with 100% of failures attributed to excessive looping behavior where models revisit a cell at least 10 times. We document a significant performance degradation in Icelandic, with models solving mazes 3-4 sizes smaller than in English, suggesting spatial reasoning in LLMs emerges from linguistic patterns rather than language-agnostic mechanisms. These results have important implications for global deployment of LLM-powered autonomous systems, showing spatial intelligence remains fundamentally constrained by training data availability and highlighting the need for architectural innovations to achieve reliable navigation across linguistic contexts.

ASNN: Learning to Suggest Neural Architectures from Performance Distributions

The architecture of a neural network (NN) plays a critical role in determining its performance. However, there is no general closed-form function that maps between network structure and accuracy, making the process of architecture design largely heuristic or search-based. In this study, we propose the Architecture Suggesting Neural Network (ASNN), a model designed to learn the relationship between NN architecture and its test accuracy, and to suggest improved architectures accordingly. To train ASNN, we constructed datasets using TensorFlow-based models with varying numbers of layers and nodes. Experimental results were collected for both 2-layer and 3-layer architectures across a grid of configurations, each evaluated with 10 repeated trials to account for stochasticity. Accuracy values were treated as inputs, and architectural parameters as outputs. The trained ASNN was then used iteratively to predict architectures that yield higher performance. In both 2-layer and 3-layer cases, ASNN successfully suggested architectures that outperformed the best results found in the original training data. Repeated prediction and retraining cycles led to the discovery of architectures with improved mean test accuracies, demonstrating the model's capacity to generalize the performance-structure relationship. These results suggest that ASNN provides an efficient alternative to random search for architecture optimization, and offers a promising approach toward automating neural network design. "Parts of the manuscript, including text editing and expression refinement, were supported by OpenAI's ChatGPT. All content was reviewed and verified by the authors."

Technical Report of TeleChat2, TeleChat2.5 and T1

We introduce the latest series of TeleChat models: \textbf{TeleChat2}, \textbf{TeleChat2.5}, and \textbf{T1}, offering a significant upgrade over their predecessor, TeleChat. Despite minimal changes to the model architecture, the new series achieves substantial performance gains through enhanced training strategies in both pre-training and post-training stages. The series begins with \textbf{TeleChat2}, which undergoes pretraining on 10 trillion high-quality and diverse tokens. This is followed by Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) to further enhance its capabilities. \textbf{TeleChat2.5} and \textbf{T1} expand the pipeline by incorporating a continual pretraining phase with domain-specific datasets, combined with reinforcement learning (RL) to improve performance in code generation and mathematical reasoning tasks. The \textbf{T1} variant is designed for complex reasoning, supporting long Chain-of-Thought (CoT) reasoning and demonstrating substantial improvements in mathematics and coding. In contrast, \textbf{TeleChat2.5} prioritizes speed, delivering rapid inference. Both flagship models of \textbf{T1} and \textbf{TeleChat2.5} are dense Transformer-based architectures with 115B parameters, showcasing significant advancements in reasoning and general task performance compared to the original TeleChat. Notably, \textbf{T1-115B} outperform proprietary models such as OpenAI's o1-mini and GPT-4o. We publicly release \textbf{TeleChat2}, \textbf{TeleChat2.5} and \textbf{T1}, including post-trained versions with 35B and 115B parameters, to empower developers and researchers with state-of-the-art language models tailored for diverse applications.

Reinforcement Learning Fine-Tunes a Sparse Subnetwork in Large Language Models

Reinforcement learning (RL) is a key post-pretraining step for aligning large language models (LLMs) with complex tasks and human preferences. While it is often assumed that RL fine-tuning requires updating most of a model's parameters, we challenge this assumption with a surprising finding: RL fine-tuning consistently modifies only a small subnetwork (typically 5-30% of weights), leaving most parameters unchanged. We call this phenomenon RL-induced parameter update sparsity. It arises naturally, without any sparsity constraints or parameter-efficient tuning, and appears across multiple RL algorithms (e.g., PPO, DPO, SimPO, PRIME) and model families (e.g., OpenAI, Meta, and open-source LLMs). Moreover, the subnetworks updated by RL show substantial overlap across different seeds, datasets, and algorithms-far exceeding chance-suggesting a partially transferable structure in the pretrained model. We show that fine-tuning only this sparse subnetwork recovers full model performance and yields parameters nearly identical to the fully fine-tuned model. Our analysis suggests this sparsity emerges because RL operates near the model's original distribution, requiring only targeted changes. KL penalties, gradient clipping, and on-policy dynamics have limited effect on the sparsity pattern. These findings shed new light on how RL adapts models: not by shifting all weights, but by focusing training on a small, consistently updated subnetwork. This insight enables more efficient RL methods and reframes sparsity through the lens of the lottery ticket hypothesis.

Generalized Advantage Estimation for Distributional Policy Gradients

Generalized Advantage Estimation (GAE) has been used to mitigate the computational complexity of reinforcement learning (RL) by employing an exponentially weighted estimation of the advantage function to reduce the variance in policy gradient estimates. Despite its effectiveness, GAE is not designed to handle value distributions integral to distributional RL, which can capture the inherent stochasticity in systems and is hence more robust to system noises. To address this gap, we propose a novel approach that utilizes the optimal transport theory to introduce a Wasserstein-like directional metric, which measures both the distance and the directional discrepancies between probability distributions. Using the exponentially weighted estimation, we leverage this Wasserstein-like directional metric to derive distributional GAE (DGAE). Similar to traditional GAE, our proposed DGAE provides a low-variance advantage estimate with controlled bias, making it well-suited for policy gradient algorithms that rely on advantage estimation for policy updates. We integrated DGAE into three different policy gradient methods. Algorithms were evaluated across various OpenAI Gym environments and compared with the baselines with traditional GAE to assess the performance.

Obscured but Not Erased: Evaluating Nationality Bias in LLMs via Name-Based Bias Benchmarks

Large Language Models (LLMs) can exhibit latent biases towards specific nationalities even when explicit demographic markers are not present. In this work, we introduce a novel name-based benchmarking approach derived from the Bias Benchmark for QA (BBQ) dataset to investigate the impact of substituting explicit nationality labels with culturally indicative names, a scenario more reflective of real-world LLM applications. Our novel approach examines how this substitution affects both bias magnitude and accuracy across a spectrum of LLMs from industry leaders such as OpenAI, Google, and Anthropic. Our experiments show that small models are less accurate and exhibit more bias compared to their larger counterparts. For instance, on our name-based dataset and in the ambiguous context (where the correct choice is not revealed), Claude Haiku exhibited the worst stereotypical bias scores of 9%, compared to only 3.5% for its larger counterpart, Claude Sonnet, where the latter also outperformed it by 117.7% in accuracy. Additionally, we find that small models retain a larger portion of existing errors in these ambiguous contexts. For example, after substituting names for explicit nationality references, GPT-4o retains 68% of the error rate versus 76% for GPT-4o-mini, with similar findings for other model providers, in the ambiguous context. Our research highlights the stubborn resilience of biases in LLMs, underscoring their profound implications for the development and deployment of AI systems in diverse, global contexts.

Towards Automated Regulatory Compliance Verification in Financial Auditing with Large Language Models

The auditing of financial documents, historically a labor-intensive process, stands on the precipice of transformation. AI-driven solutions have made inroads into streamlining this process by recommending pertinent text passages from financial reports to align with the legal requirements of accounting standards. However, a glaring limitation remains: these systems commonly fall short in verifying if the recommended excerpts indeed comply with the specific legal mandates. Hence, in this paper, we probe the efficiency of publicly available Large Language Models (LLMs) in the realm of regulatory compliance across different model configurations. We place particular emphasis on comparing cutting-edge open-source LLMs, such as Llama-2, with their proprietary counterparts like OpenAI's GPT models. This comparative analysis leverages two custom datasets provided by our partner PricewaterhouseCoopers (PwC) Germany. We find that the open-source Llama-2 70 billion model demonstrates outstanding performance in detecting non-compliance or true negative occurrences, beating all their proprietary counterparts. Nevertheless, proprietary models such as GPT-4 perform the best in a broad variety of scenarios, particularly in non-English contexts.