PreprintNot peer reviewedFrom Eigen AIA preprint reports a multi-agent system that lifts scientific reasoning scores on a hard exam benchmarkOriginal title: Eigen-Agent: Adaptive Multi-Agent Scientific Reasoning with Monitor-Based RAGResearchers describe Eigen-Agent, combining implicit retrieval and structured multi-agent refinement to boost scores on expert-level science questions.Added 19 Oct 2026
PreprintNot peer reviewedSurvey maps how reinforcement learning is reshaping diffusion modelsOriginal title: Reinforcement Learning-driven Diffusion Models: Language, Vision, and BeyondA preprint surveys how reinforcement learning post-training is being used to improve diffusion models across language, vision, robotics and science.Zhiyuan Ma et al. · Added 19 Oct 2026
PreprintNot peer reviewedA preprint reports gradient-based pruning for LLMs without retrainingOriginal title: Beyond Size: How Gradients Shape Pruning Decisions in Large Language ModelsResearchers propose GBLM-Pruner, a training-free method using gradients to decide which LLM weights to remove.Rocktim Jyoti Das et al. · Added 19 Oct 2026
PreprintNot peer reviewedFrom WorkatoA preprint proposes token-level credit fix for reasoning RL trainingOriginal title: Right Answer, Wrong Tokens: Tail-Aware Credit Calibration for RLVRA preprint reports a method that softens reward credit for rare, risky tokens within otherwise correct RL training trajectories.Added 19 Oct 2026
PreprintNot peer reviewedNew benchmark finds agentic LLMs remain widely exploitable via indirect prompt injectionOriginal title: INFUSE: A Unified Surface Benchmark for Indirect Prompt Injection in Agentic SystemsA preprint introduces INFUSE, a 40-task benchmark showing many agentic LLMs can be hijacked by instructions hidden in files, logs and other surfaces.Thuy Dung Nguyen et al. · Added 19 Oct 2026
PreprintNot peer reviewedStudy finds LLM stereotyping varies by prompt language, not just cultureOriginal title: MBBQ: A Dataset for Cross-Lingual Comparison of Stereotypes in Generative LLMsA preprint reports that several LLMs show more biased question-answering behaviour in some languages than in English, even after accounting for cultural differences.Vera Neplenbroek et al. · Added 19 Oct 2026
Conference paperReview status unknownFrom AXAResearchers compare energy use of diffusion and autoregressive language modelsOriginal title: Energy Consumption Analysis of Discrete Diffusion and Autoregressive Language ModelsA conference paper finds diffusion language models use less energy per token than autoregressive ones on long outputs, but more on short ones.Added 19 Oct 2026 · Released Apr 2026
PreprintNot peer reviewedA preprint reports sparse autoencoders that bind concepts to dedicated feature slotsOriginal title: AlignSAE: Concept-Aligned Sparse AutoencodersResearchers describe AlignSAE, a method that supervises sparse autoencoder features to align with chosen concepts, improving causal control in GPT-2 tests.Minglai Yang et al. · Added 19 Oct 2026
Conference paperReview status unknownSoftware tool locates bit-flip faults in large language models and aids recoveryOriginal title: BitFlipScope: Scalable Fault Localization and Recovery for Bit-Flip Corruptions in LLMsA conference paper presents BitFlipScope, a software method to find and partly fix bit-flip corruption in transformer models.Muhammad Zeeshan Karamat et al. · Added 18 Oct 2026 · Released Apr 2026
PreprintNot peer reviewedFrom AIM IntelligenceNew benchmark probes why multimodal AI models hallucinate across multiple imagesOriginal title: Fine-Grained Multi Image Object Hallucination BenchmarkA preprint introduces MIOH, a benchmark testing 29 models to find out where multi-image object hallucination actually originates.Added 18 Oct 2026
PreprintNot peer reviewedFrom Palo Alto NetworksA preprint reports bug-fixing benchmarks may be inflated by training data contaminationOriginal title: BenchChecker: Assessing the Credibility of Bug-Fixing Benchmarks for LLMsA preprint presents BenchChecker, a tool that detects contaminated test samples in LLM bug-fixing benchmarks using only text outputs.Added 18 Oct 2026
PreprintNot peer reviewedReview maps 114 studies on LLM-based multi-agent code generation systemsOriginal title: LLM-Based Multi-Agent Systems for Code Generation: A Multi-Vocal Literature ReviewA literature review pulls together academic and industry work on multi-agent LLM systems for code generation, covering benchmarks, challenges and open questions.Zeeshan Rasheeda et al. · Added 18 Oct 2026 · Released Feb 2026
PreprintNot peer reviewedA preprint describes a multi-agent system that auto-converts jailbreak papers into runnable attack codeOriginal title: Jailbreak Foundry: From Papers to Runnable Attacks for Reproducible BenchmarkingResearchers report an automated pipeline that turns jailbreak attack papers into executable, benchmark-ready modules with close fidelity to original results.Zhicheng Fang et al. · Added 18 Oct 2026 · Released Jul 2026
PreprintNot peer reviewedA preprint reports a prompt-aware tuning method that beats LoRA with fewer bottlenecksOriginal title: PEDRO: Parameter-Efficient Fine-tuning with Prompt DEpenDent Representation MOdificationA preprint describes PEDRO, a parameter-efficient fine-tuning method that the authors say outperforms LoRA-style baselines while running faster in multi-tenant settings.Tianfang Xie et al. · Added 18 Oct 2026 · Released Sep 2024
Conference paperReview status unknownFrom GoogleResearchers detail Google’s PLUM framework for LLM-based YouTube recommendationsOriginal title: PLUM: Adapting Pre-trained Language Models for Industrial-scale Generative RecommendationsA conference paper describes PLUM, which adapts pre-trained language models to generate video recommendations for YouTube, replacing large embedding tables.Added 18 Oct 2026 · Released Oct 2025
Conference paperReview status unknownResearchers combine LLM task planning and reward shaping for sparse-reward robot controlOriginal title: Progress Reward Model for Reinforcement Learning via Large Language ModelsA conference paper describes a framework that uses large language models to both plan subtasks and build dense rewards for reinforcement learning on robotics tasks.Xiuhui Zhang et al. · Added 18 Oct 2026 · Released Jan 2025
PreprintNot peer reviewedA preprint reports LLM agents inventing their own symbolic languages to reason fasterOriginal title: When LLMs Develop Languages: Symbolic Communication for Efficient Multi-Agent ReasoningA preprint describes LLM agents that evolve shared symbolic shorthand protocols, cutting reasoning tokens while keeping accuracy steady.Zhengqi Pei et al. · Added 18 Oct 2026
PreprintNot peer reviewedFrom Western DigitalA preprint reports a mobile KV-cache system that extends on-device LLM context lengthOriginal title: TierKV: Long-Context On-Device LLMs via Predictive Multi-Tier KV CachingA preprint describes TierKV, a mobile LLM inference system that predicts and tiers KV-cache storage to support longer contexts on phones.Added 18 Oct 2026
PreprintNot peer reviewedFrom TencentResearchers report long-context LLMs often fail to spot unanswerable questionsOriginal title: Towards Reliable Long-Context Reasoning: Detecting Unanswerable Questions via FactGuardA preprint introduces a benchmark and training method to help long-context language models recognise when a question cannot be answered from the given text.Added 18 Oct 2026
PreprintNot peer reviewedPreprint traces PPO training collapses to critic instability and proposes fixesOriginal title: EasyPPO: Stabilizing the Critic Is KeyA preprint reports that two critic failure modes destabilise PPO for LLM reasoning, and proposes EasyPPO to fix them.Xuanyi Zhou et al. · Added 18 Oct 2026
