A fundamental limitation of current AI agents is their inability to learn complex skills on the fly at test time, often behaving like "clever but clueless interns" in novel environments. This severely limits their practical utility. To systematically measure and drive progress on this challenge, we first introduce the Jericho Test-Time Learning (J-TTL) benchmark. J-TTL is a new evaluation setup where an agent must play the same game for several consecutive episodes, attempting to improve its performance from one episode to the next. On J-TTL, we find that existing adaptation methods like reflection, memory, or reinforcement learning struggle. To address the challenges posed by our benchmark, we present EvoTest, an evolutionary test-time learning framework that improves an agent without any fine-tuning or gradients-by evolving the entire agentic system after every episode. EvoTest has two roles: the Actor Agent, which plays the game, and the Evolver Agent, which analyzes the episode transcript to propose a revised configuration for the next run. This configuration rewrites the prompt, updates memory by logging effective state-action choices, tunes hyperparameters, and learns the tool-use routines. On our J-TTL benchmark, EvoTest consistently increases performance, outperforming not only reflection and memory-only baselines but also more complex online fine-tuning methods. Notably, our method is the only one capable of winning two games (Detective and Library), while all baselines fail to win any.

Modern IR systems are increasingly tasked with answering complex, multi-faceted queries that require deep reasoning rather than simple keyword or semantic matching. While LLM-based IR has shown great promise, the prevailing retrieve-then-rerank paradigm inherits the limitations of embedding-based retrieval; parametric generative approaches are difficult to update with new information; and long-context methods that place the entire corpus in context are computationally infeasible for large document collections. To address these challenges, we introduce LATTICE, a hierarchical retrieval framework that enables an LLM to reason over and navigate large corpora with logarithmic search complexity by imposing a semantic tree structure on the corpus. Our approach consists of two stages: (1) an offline phase that organizes the corpus into a semantic hierarchy via either a bottom-up agglomerative strategy or a top-down divisive strategy using multi-level summaries and (2) an online traversal phase where a search LLM navigates this tree. A central challenge in such LLM-guided search is that the model's relevance judgments are noisy, context-dependent, and unaware of the hierarchy, making cross-branch and cross-level comparisons difficult. To overcome this, we propose a traversal algorithm that estimates calibrated latent relevance scores from local LLM outputs and aggregates them into a global path relevance metric. Our training-free framework achieves state-of-the-art zero-shot performance on the reasoning-intensive BRIGHT benchmark, demonstrating up to 9% improvement in Recall@100 and 5% in nDCG@10 over the next best zero-shot baseline. Furthermore, compared to the fine-tuned SOTA method DIVER-v2, LATTICE attains comparable results on BRIGHT subsets that use a static corpus for evaluation.

Recent advances in Large Language Models (LLMs) demonstrate that chain-of-thought prompting and deep reasoning substantially enhance performance on complex tasks, and multi-agent systems can further improve accuracy by enabling model debates. However, applying deep reasoning to all problems is computationally expensive. To mitigate these costs, we propose a complementary agent system integrating small and large LLMs. The small LLM first generates an initial answer, which is then verified by the large LLM. If correct, the answer is adopted directly; otherwise, the large LLM performs in-depth reasoning. Experimental results show that, for simple problems, our approach reduces the computational cost of the large LLM by more than 50% with negligible accuracy loss, while consistently maintaining robust performance on complex tasks.

Large language model (LLM) alignment is typically achieved through learning from human preference comparisons, making the quality of preference data critical to its success. Existing studies often pre-process raw training datasets to identify valuable preference pairs using external reward models or off-the-shelf LLMs, achieving improved overall performance but rarely examining whether individual, selected data point is genuinely beneficial. We assess data quality through individual influence on validation data using our newly proposed truncated influence function (TIF), which mitigates the over-scoring present in traditional measures and reveals that preference data quality is inherently a property of the model. In other words, a data pair that benefits one model may harm another. This leaves the need to improve the preference data selection approaches to be adapting to specific models. To this end, we introduce two candidate scoring functions (SFs) that are computationally simpler than TIF and positively correlated with it. They are also model dependent and can serve as potential indicators of individual data quality for preference data selection. Furthermore, we observe that these SFs inherently exhibit errors when compared to TIF. To this end, we combine them to offset their diverse error sources, resulting in a simple yet effective data selection rule that enables the models to achieve a more precise selection of valuable preference data. We conduct experiments across diverse alignment benchmarks and various LLM families, with results demonstrating that better alignment performance can be achieved using less data, showing the generality of our findings and new methods.

This is the preprint version of the article "LLM - Guided Synthetic Augmentation (LGSA) for Mitigating Bias in AI Systems." This version is made available on arXiv for early dissemination. If accepted, the final authenticated version will be published in the respective venue. Dr. G opichand G School of Computer Science and Engineering Vellore Institute of Technology Vellore - 632014, TamilNadu, India gopichand.g@vit.ac.in Abstract -- Bias in Artificial Intelligence systems, especially those that rely on natural language data, brings up serious ethical and practical issues. When certain groups are underrepresented, it often leads to uneven performance across different demographics. Whil e traditional fairness methods like pre - processing, in - processing, and post - processing can be helpful, they usually depend on protected - attribute labels, create a trade - off between accuracy and fairness, and struggle to adapt across various datas ets. To tackle these challenges, this study presents LLM - Guided Synthetic Augmentation (LGSA), a process that leverages large language models to create counterfactual examples for underrepresented groups while keeping label integrity intact. We put LGSA to the test on a controlled dataset of short English sentences that included gendered pronouns, professions, and binary task labels. The process involved using structured prompts to a large language model to generate gender - swapped paraphrases, followed by a thorough quality control process. This included checking for semantic similarity, verifying attributes, screening for toxi city, and conducting human spot checks. The augmented dataset broadened training coverage and was utilized to train a classifier under consistent experimental conditions. The results showed that LGSA significantly lessens performance disparities without co mpromising accuracy. The baseline model achieved an impressive 96.7% accuracy but had a gender bias gap of 7.2%. A simple swap augmentation brought the gap down to 0.7% but also reduced accuracy to 95.6%. In contrast, LGSA achieved an overall accuracy of 9 9.1%, showing strong performance on female - labeled examples and a reduced gap of 1.9%. These results indicate that LGSA is a powerful and dependable strategy for mitigating bias. By generating diverse and semantically accurate counterfactuals, this method enhances the balance of subgroup performance, narrows bias gaps, and maintains high ove rall task accuracy and label fidelity, showcasing its potential as a practical framework for fairness - focused AI systems.

EmphST -Bench To guide algorithm exploration and evaluate the performance of our model, we design an evaluation pipeline for the emphasis-preserving speech-to-speech translation system. Given the lack of ready-to-use benchmarks for this important task, we leverage LLMs to translate the test set from the StressTest [21] corpus into the target language and then filter the results via human experts. This process creates a high-quality benchmark dataset, EmphST -Bench, with manually verified emphasis alignments between source and target utterances, ensuring reliable assessment of cross-lingual emphasis preservation. The human filtering step focuses on correcting any discrepancies in semantic equivalence, contrastive focus, and emotional intensity, resulting in a robust evaluation set that closely mirrors real-world linguistic nuances. EmphST -Bench consists of carefully selected parallel samples from English (source) to Chinese (target), providing a standardized resource for evaluating stress-aware S2ST systems. We report the statistics of EmphST -Bench in Table. 1. T able 1: Statistics of the EmphST -Bench dataset.Statistic V alue Number of Samples 218 Avg.

Large Vision-Language Models (LVLMs) unlock powerful multimodal reasoning but also expand the attack surface, particularly through adversarial inputs that conceal harmful goals in benign prompts. We propose SHIELD, a lightweight, model-agnostic preprocessing framework that couples fine-grained safety classification with category-specific guidance and explicit actions (Block, Reframe, Forward). Unlike binary moderators, SHIELD composes tailored safety prompts that enforce nuanced refusals or safe redirection without retraining. Across five benchmarks and five representative LVLMs, SHIELD consistently lowers jailbreak and non-following rates while preserving utility. Our method is plug-and-play, incurs negligible overhead, and is easily extendable to new attack types -- serving as a practical safety patch for both weakly and strongly aligned LVLMs.

Detoxification in large language models (LLMs) remains a significant research challenge. Existing decoding detoxification methods are all based on external constraints, which require additional resource overhead and lose generation fluency. This work proposes Detoxification with Self-Constrained Decoding (DSCD), a novel method for LLM detoxification without parameter fine-tuning. DSCD strengthens the inner next-token distribution of the safety layer while weakening that of hallucination and toxic layers during output generation. This effectively diminishes toxicity and enhances output safety. DSCD offers lightweight, high compatibility, and plug-and-play capabilities, readily integrating with existing detoxification methods for further performance improvement. Extensive experiments on representative open-source LLMs and public datasets validate DSCD's effectiveness, demonstrating state-of-the-art (SOTA) performance in both detoxification and generation fluency, with superior efficiency compared to existing methods. These results highlight DSCD's potential as a practical and scalable solution for safer LLM deployments.

Chain of thought (CoT) fine-tuning aims to endow large language models (LLMs) with reasoning capabilities by training them on curated reasoning traces. It leverages both supervised and reinforced fine-tuning to cultivate human-like reasoning skills in LLMs, including detailed planning, divergent thinking, intuitive judgment, timely reflection, internal thinking, and fact perception, etc. As CoT fine-tuning has advanced, LLMs have demonstrated substantial improvements in tasks such as mathematical reasoning and code generation. However, existing surveys about CoT fine-tuning primarily focus on technical aspects and overlook a systematic analysis from the perspective of human reasoning mechanisms. Given that the ultimate goal of CoT fine-tuning is to enable LLMs to reason like humans, it is crucial to investigate this technique through the lens of human cognition. To fill this gap, we present the first comprehensive survey of CoT fine-tuning grounded in human reasoning theory. Specifically, inspired by the well-known Six Thinking Hats framework, which systematically characterizes common human thinking modes using six metaphorical hats, we classify and examine CoT fine-tuning methods through this lens. Furthermore, building upon this theory, we outline potential directions for future research in CoT fine-tuning. In addition, we compile a comprehensive overview of existing datasets and model performances, and a real-time GitHub repository \footnote{https://github.com/AI-Chen/Awesome-CoT-Finetuning} that continuously tracks recent advances in this area is maintained. We hope this survey will serve as a valuable resource to inspire innovation and foster progress in this rapidly evolving field.

Most large language models (LLMs) today excel at generating raw, sequential code with minimal abstractions and custom structures. However, there has been little work on graph-based abstract code generation, where significant logic is encapsulated in predefined nodes and execution flow is determined by edges. This is relevant for visual programming languages, and in cases where raw source code is inaccessible to users and LLM training sets. In this work, we propose and evaluate JSON representations for graphs to enable high accuracy graph-based abstract code generation. We evaluate these representations on ScratchTest, a mini-benchmark based on our custom Python re-implementation of Scratch, which tests the LLM in code graph space. Our findings demonstrate that LLMs can indeed perform the aforementioned generation task in a single pass without relying on specialized or complex pipelines, given the correct graph representations. We also show that different representations induce significantly different accuracies, highlighting the instrumental role of representations in this generation task. All in all, this work establishes the first steps towards representation learning for graph-based abstract code generation.