AI

arXiv:2505.16834v3 Announce Type: replace Abstract: Retrieval-augmented generation (RAG) systems have advanced large language models (LLMs) in complex deep search scenarios requiring multi-step reasoning and iterative information retrieval. However, existing approaches face critical limitations that lack high-quality training trajectories or suffer from the distributional mismatches in simulated environments and prohibitive computational costs for real-world deployment. This paper introduces SimpleDeepSearcher, a lightweight yet effective framework that bridges this gap through strategic data engineering rather than complex training paradigms. Our approach synthesizes high-quality training data by simulating realistic user interactions in live web search environments, coupled with a multi-criteria curation strategy that optimizes the diversity and quality of input and output side. Experiments on five benchmarks across diverse domains demonstrate that SFT on only 871 curated samples yields significant improvements over RL-based baselines. Our work establishes SFT as a viable pathway by systematically addressing the data-scarce bottleneck, offering practical insights for efficient deep search systems. Our code is available at https://github.com/RUCAIBox/SimpleDeepSearcher.

arXiv:2510.06811v1 Announce Type: new Abstract: The Circuit Localization track of the Mechanistic Interpretability Benchmark (MIB) evaluates methods for localizing circuits within large language models (LLMs), i.e., subnetworks responsible for specific task behaviors. In this work, we investigate whether ensembling two or more circuit localization methods can improve performance. We explore two variants: parallel and sequential ensembling. In parallel ensembling, we combine attribution scores assigned to each edge by different methods-e.g., by averaging or taking the minimum or maximum value. In the sequential ensemble, we use edge attribution scores obtained via EAP-IG as a warm start for a more expensive but more precise circuit identification method, namely edge pruning. We observe that both approaches yield notable gains on the benchmark metrics, leading to a more precise circuit identification approach. Finally, we find that taking a parallel ensemble over various methods, including the sequential ensemble, achieves the best results. We evaluate our approach in the BlackboxNLP 2025 MIB Shared Task, comparing ensemble scores to official baselines across multiple model-task combinations.

arXiv:2509.12508v3 Announce Type: replace Abstract: In recent years, automatic speech recognition (ASR) has witnessed transformative advancements driven by three complementary paradigms: data scaling, model size scaling, and deep integration with large language models (LLMs). However, LLMs are prone to hallucination, which can significantly degrade user experience in real-world ASR applications. In this paper, we present Fun-ASR, a large-scale, LLM-based ASR system that synergistically combines massive data, large model capacity, LLM integration, and reinforcement learning to achieve state-of-the-art performance across diverse and complex speech recognition scenarios. Moreover, Fun-ASR is specifically optimized for practical deployment, with enhancements in streaming capability, noise robustness, code-switching, hotword customization, and satisfying other real-world application requirements. Experimental results show that while most LLM-based ASR systems achieve strong performance on open-source benchmarks, they often underperform on real industry evaluation sets. Thanks to production-oriented optimizations, Fun-ASR achieves state-of-the-art performance on real application datasets, demonstrating its effectiveness and robustness in practical settings.

arXiv:2510.01367v3 Announce Type: replace-cross Abstract: Reward hacking, where a reasoning model exploits loopholes in a reward function to achieve high rewards without solving the intended task, poses a significant threat. This behavior may be explicit, i.e. verbalized in the model’s chain-of-thought (CoT), or implicit, where the CoT appears benign thus bypasses CoT monitors. To detect implicit reward hacking, we propose TRACE (Truncated Reasoning AUC Evaluation). Our key observation is that hacking occurs when exploiting the loophole is easier than solving the actual task. This means that the model is using less ‘effort’ than required to achieve high reward. TRACE quantifies effort by measuring how early a model’s reasoning becomes sufficient to obtain the reward. We progressively truncate a model’s CoT at various lengths, force the model to answer, and estimate the expected reward at each cutoff. A hacking model, which takes a shortcut, will achieve a high expected reward with only a small fraction of its CoT, yielding a large area under the accuracy-vs-length curve. TRACE achieves over 65% gains over our strongest 72B CoT monitor in math reasoning, and over 30% gains over a 32B monitor in coding. We further show that TRACE can discover unknown loopholes during training. Overall, TRACE offers a scalable unsupervised approach for oversight where current monitoring methods prove ineffective.

arXiv:2510.04417v1 Announce Type: cross Abstract: The study of multimodality has garnered significant interest in fields where the analysis of interactions among multiple information sources can enhance predictive modeling, data fusion, and interpretability. Partial information decomposition (PID) has emerged as a useful information-theoretic framework to quantify the degree to which individual modalities independently, redundantly, or synergistically convey information about a target variable. However, existing PID methods depend on optimizing over a joint distribution constrained by estimated pairwise probability distributions, which are costly and inaccurate for continuous and high-dimensional modalities. Our first key insight is that the problem can be solved efficiently when the pairwise distributions are multivariate Gaussians, and we refer to this problem as Gaussian PID (GPID). We propose a new gradient-based algorithm that substantially improves the computational efficiency of GPID based on an alternative formulation of the underlying optimization problem. To generalize the applicability to non-Gaussian data, we learn information-preserving encoders to transform random variables of arbitrary input distributions into pairwise Gaussian random variables. Along the way, we resolved an open problem regarding the optimality of joint Gaussian solutions for GPID. Empirical validation in diverse synthetic examples demonstrates that our proposed method provides more accurate and efficient PID estimates than existing baselines. We further evaluate a series of large-scale multimodal benchmarks to show its utility in real-world applications of quantifying PID in multimodal datasets and selecting high-performing models.

arXiv:2505.18350v2 Announce Type: replace-cross Abstract: As Large Language Models (LLMs) are increasingly deployed for narrow tasks in resource-constrained settings, a central question arises: how much of an LLM is truly necessary for a given task? We present LLM-Sieve, a framework that prunes LLMs down to the minimal parameter subset needed to preserve task performance. Our approach introduces two innovations: (i) output-aligned non-orthogonal projections, which yield more faithful low-rank approximations than traditional PCA/SVD by aligning directly with layer outputs; and (ii) adaptive pruning via a Genetic Algorithm, which automatically discovers matrix-specific pruning levels and exposes the uneven distribution of task-relevant knowledge. Across models from 3.8B to 70B parameters, LLM-Sieve removes 20-75% of weights with only 1-5% accuracy loss-substantially ahead of prior pruning methods. Beyond efficiency, our framework reveals bottleneck matrices that concentrate critical knowledge, suggesting architectural implications for future LLM design. LLM-Sieve integrates seamlessly with LoRA fine-tuning and quantization, enabling both efficient deployment and deeper understanding of knowledge organization in LLMs.

What if an AI agent could localize a root cause, prove a candidate fix via automated analysis and testing, and proactively rewrite related code to eliminate the entire vulnerability class—then open an upstream patch for review? Google DeepMind introduces CodeMender, an AI agent that generates, validates, and upstreams fixes for real-world vulnerabilities using Gemini “Deep Think” reasoning and a tool-augmented workflow. In six months of internal deployment, CodeMender contributed 72 security patches across open-source projects, including codebases up to ~4.5M lines, and is designed to act both reactively (patching known issues) and proactively (rewriting code to remove vulnerability classes). Understanding the Architecture The agent couples large-scale code reasoning with program-analysis tooling: static and dynamic analysis, differential testing, fuzzing, and satisfiability-modulo-theory (SMT) solvers. A multi-agent design adds specialized “critique” reviewers that inspect semantic diffs and trigger self-corrections when regressions are detected. These components let the system localize root causes, synthesize candidate patches, and automatically regression-test changes before surfacing them for human review. https://deepmind.google/discover/blog/introducing-codemender-an-ai-agent-for-code-security/? Validation Pipeline and Human Gate DeepMind emphasizes automatic validation before any human touches a patch: the system tests for root-cause fixes, functional correctness, absence of regressions, and style compliance; only high-confidence patches are proposed for maintainer review. This workflow is explicitly tied to Gemini Deep Think’s planning-centric reasoning over debugger traces, code search results, and test outcomes. Proactive Hardening: Compiler-Level Guards Beyond patching, CodeMender applies security-hardening transforms at scale. Example: automated insertion of Clang’s -fbounds-safety annotations in libwebp to enforce compiler-level bounds checks—an approach that would have neutralized the 2023 libwebp heap overflow (CVE-2023-4863) exploited in a zero-click iOS chain and similar buffer over/underflows where annotations are applied. Case Studies DeepMind details two non-trivial fixes: (1) a crash initially flagged as a heap overflow traced to incorrect XML stack management; and (2) a lifetime bug requiring edits to a custom C-code generator. In both cases, agent-generated patches passed automated analysis and an LLM-judge check for functional equivalence before proposal. https://deepmind.google/discover/blog/introducing-codemender-an-ai-agent-for-code-security/? Deployment Context and Related Initiatives Google’s broader announcement frames CodeMender as part of a defensive stack that includes a new AI Vulnerability Reward Program (consolidating AI-related bounties) and the Secure AI Framework 2.0 for agent security. The post reiterates the motivation: as AI-powered vulnerability discovery scales (e.g., via BigSleep and OSS-Fuzz), automated remediation must scale in tandem. Our Comments CodeMender operationalizes Gemini Deep Think plus program-analysis tools (static/dynamic analysis, fuzzing, SMT) to localize root causes and propose patches that pass automated validation before human review. Reported early data: 72 upstreamed security fixes across open-source projects over six months, including codebases on the order of ~4.5M lines. The system also applies proactive hardening (e.g., compiler-enforced bounds via Clang -fbounds-safety) to reduce memory-safety bug classes rather than only patching instances. No latency or throughput benchmarks are published yet, so impact is best measured by validated fixes and scope of hardened code. Check out the TECHNICAL DETAILS. Feel free to check out our GitHub Page for Tutorials, Codes and Notebooks. Also, feel free to follow us on Twitter and don’t forget to join our 100k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. The post Google DeepMind Introduces CodeMender: A New AI Agent that Uses Gemini Deep Think to Automatically Patch Critical Software Vulnerabilities appeared first on MarkTechPost.

arXiv:2510.04391v1 Announce Type: cross Abstract: What is the computational objective of imagination? While classical interpretations suggest imagination is useful for maximizing rewards, recent findings challenge this view. In this study, we propose that imagination serves to access an internal world model (IWM) and use psychological network analysis to explore IWMs in humans and large language models (LLMs). Specifically, we assessed imagination vividness ratings using two questionnaires and constructed imagination networks from these reports. Imagination networks from human groups showed correlations between different centrality measures, including expected influence, strength, and closeness. However, imagination networks from LLMs showed a lack of clustering and lower correlations between centrality measures under different prompts and conversational memory conditions. Together, these results indicate a lack of similarity between IWMs in human and LLM agents. Overall, our study offers a novel method for comparing internally-generated representations in humans and AI, providing insights for developing human-like imagination in artificial intelligence.

arXiv:2510.04400v1 Announce Type: new Abstract: In this work, we explore the relevance of textual semantics to Large Language Models (LLMs), extending previous insights into the connection between distributional semantics and structural semantics. We investigate whether LLM-generated texts preserve semantic isotopies. We design a story continuation experiment using 10,000 ROCStories prompts completed by five LLMs. We first validate GPT-4o’s ability to extract isotopies from a linguistic benchmark, then apply it to the generated stories. We then analyze structural (coverage, density, spread) and semantic properties of isotopies to assess how they are affected by completion. Results show that LLM completion within a given token horizon preserves semantic isotopies across multiple properties.