Wiredframe Radar logo

Category intelligence

Research Briefing — January 15, 2026

335 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research centers on AI security frameworks and alignment challenges in reasoning models. Bruce Schneier introduces Promptware, reconceptualizing prompt injection as a distinct malware class with a five-step kill chain model. OpenAI's alignment team presents Confessions research for detecting reward-hacked outputs.

Survey work includes The AI Hippocampus organizing LLM memory into implicit, explicit, and agentic paradigms, while Adversarial Tales exposes jailbreak vulnerabilities through cultural narrative framing. DASD-4B-Thinking achieves SOTA reasoning among 4B open-source models through distribution-aligned distillation.

Key Themes

AI Safety & Alignment · 18AI Security & Safety · 6Large Reasoning Models · 4LLM Agents & Planning · 18Training & Fine-Tuning Efficiency · 10LLM Efficiency & Test-Time Scaling · 12Evaluation & Benchmarks · 11Mechanistic Interpretability · 6Knowledge Distillation & Reasoning · 8Medical Imaging & Healthcare AI · 14

Primary evidence

Top Ranked Signals

Research arXiv (Artificial Intelligence) Jan 15

The Promptware Kill Chain: How Prompt Injections Gradually Evolved Into a Multi-Step Malware

By Ben Nassi, Bruce Schneier, Oleg Brodt

83 score
AI Analysis

Proposes 'promptware' as a distinct malware class targeting LLM-based systems and introduces five-step kill chain model, arguing that 'prompt injection' framing obscures multi-step attack complexity. Co-authored by Bruce Schneier.

arXiv:2601.09625v1 Announce Type: cross Abstract: The rapid adoption of large language model (LLM)-based systems -- from chatbots to autonomous agents capable of executing code and financial transactions -- has created a new attack surface that existing security frameworks inadequately address. The dominant framing of these threats as "prompt injection" -- a catch-all phrase for security failures in LLM-based systems -- obscures a more complex reality: Attacks on LLM-based systems increasingly
AI SecurityLLM AgentsPrompt InjectionThreat Modeling
Research arXiv (Artificial Intelligence) Jan 15

A.X K1 Technical Report

By Sung Jun Cheon, Jaekyung Cho, Seongho Choi, Hyunjun Eun, Seokhwan Jo, Jaehyun Jun, Minsoo Kang, Jin Kim, Jiwon Kim, Minsang Kim, Sungwan Kim, Seungsik Kim, Tae Yoon Kim, Youngrang Kim, Hyeongmun Lee, Sangyeol Lee, Sungeun Lee, Youngsoon Lee, Yujin Lee, Seongmin Ok, Chanyong Park, Hyewoong Park, Junyoung Park, Hyunho Yang, Subin Yi, Soohyun Bae, Dhammiko Arya, Yongseok Choi, Sangho Choi, Dongyeon Cho, Seungmo Cho, Gyoungeun Han, Yong-jin Han, Seokyoung Hong, Hyeon Hwang, Wonbeom Jang, Minjeong Ju, Wonjin Jung, Keummin Ka, Sungil Kang, Dongnam Kim, Joonghoon Kim, Jonghwi Kim, SaeRom Kim, Sangjin Kim, Seongwon Kim, Youngjin Kim, Seojin Lee, Sunwoo Lee, Taehoon Lee, Chanwoo Park, Sohee Park, Sooyeon Park, Yohan Ra, Sereimony Sek, Seungyeon Seo, Gun Song, Sanghoon Woo, Janghan Yoon, Sungbin Yoon

82 score
AI Analysis

Technical report for A.X K1, a 519B-parameter MoE language model trained from scratch on 10T tokens. Features 'Think-Fusion' training enabling user-controlled switching between thinking and non-thinking modes in a single model.

arXiv:2601.09200v1 Announce Type: cross Abstract: We introduce A.X K1, a 519B-parameter Mixture-of-Experts (MoE) language model trained from scratch. Our design leverages scaling laws to optimize training configurations and vocabulary size under fixed computational budgets. A.X K1 is pre-trained on a corpus of approximately 10T tokens, curated by a multi-stage data processing pipeline. Designed to bridge the gap between reasoning capability and inference efficiency, A.X K1 supports explicitly c
Large Language ModelsMixture of ExpertsReasoningModel Architecture
Research arXiv (Artificial Intelligence) Jan 15

DeliberationBench: When Do More Voices Hurt? A Controlled Study of Multi-LLM Deliberation Protocols

By Vaarunay Kaushal, Taranveer Singh

78 score
AI Analysis

DeliberationBench reveals a striking negative result: a simple best-single selection baseline achieves 82.5% win rate, dramatically outperforming deliberation protocols (13.8%) at 1.5-2.5x the computational cost.

arXiv:2601.08835v1 Announce Type: cross Abstract: Multi-agent systems where Large Language Models (LLMs) deliberate to form consensus have gained significant attention, yet their practical value over simpler methods remains under-scrutinized. We introduce DELIBERATIONBENCH, a controlled benchmark evaluating three deliberation protocols against a strong baseline of selecting the best response from a pool of model outputs. Across 270 questions and three independent seeds (810 total evaluations),
Multi-Agent SystemsEvaluationNegative Results
Research arXiv (Artificial Intelligence) Jan 15

GIFT: Unlocking Global Optimality in Post-Training via Finite-Temperature Gibbs Initialization

By Zhengyang Zhao, Lu Ma, Yizhen Jiang, Xiaochen Ma, Zimo Meng, Chengyu Shen, Lexiang Tang, Haoze Sun, Peng Pei, Wentao Zhang

78 score
AI Analysis

Proposes GIFT (Gibbs Initialization with Finite Temperature), addressing the SFT-RL mismatch in Large Reasoning Model post-training. Reformulates SFT as finite-temperature energy potential to preserve exploration capacity for subsequent RL.

arXiv:2601.09233v1 Announce Type: cross Abstract: The prevailing post-training paradigm for Large Reasoning Models (LRMs)--Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL)--suffers from an intrinsic optimization mismatch: the rigid supervision inherent in SFT induces distributional collapse, thereby exhausting the exploration space necessary for subsequent RL. In this paper, we reformulate SFT within a unified post-training framework and propose Gibbs Initialization with Fin
Post-TrainingReinforcement LearningLarge Reasoning ModelsAlignment
Research arXiv (Artificial Intelligence) Jan 15

TranslateGemma Technical Report

By Mara Finkelstein, Isaac Caswell, Tobias Domhan, Jan-Thorsten Peter, Juraj Juraska, Parker Riley, Daniel Deutsch, Cole Dilanni, Colin Cherry, Eleftheria Briakou, Elizabeth Nielsen, Jiaming Luo, Kat Black, Ryan Mullins, Sweta Agrawal, Wenda Xu, Erin Kats, Stephane Jaskiewicz, Markus Freitag, David Vilar

77 score
AI Analysis

TranslateGemma presents open machine translation models based on Gemma 3, using supervised fine-tuning on synthetic/human parallel data followed by RL with MetricX-QE and AutoMQM rewards.

arXiv:2601.09012v1 Announce Type: cross Abstract: We present TranslateGemma, a suite of open machine translation models based on the Gemma 3 foundation models. To enhance the inherent multilingual capabilities of Gemma 3 for the translation task, we employ a two-stage fine-tuning process. First, supervised fine-tuning is performed using a rich mixture of high-quality large-scale synthetic parallel data generated via state-of-the-art models and human-translated parallel data. This is followed by
Machine TranslationOpen ModelsGoogle
Research arXiv (Artificial Intelligence) Jan 15

From Adversarial Poetry to Adversarial Tales: An Interpretability Research Agenda

By Piercosma Bisconti, Marcello Galisai, Matteo Prandi, Federico Pierucci, Olga Sorokoletova, Francesco Giarrusso, Vincenzo Suriani, Marcantonio Brancale, Daniele Nardi

76 score
AI Analysis

Introduces Adversarial Tales, a jailbreak technique embedding harmful content in cyberpunk narratives analyzed through Propp's folktale morphology. Achieves 71.3% attack success rate across 26 frontier models.

arXiv:2601.08837v1 Announce Type: cross Abstract: Safety mechanisms in LLMs remain vulnerable to attacks that reframe harmful requests through culturally coded structures. We introduce Adversarial Tales, a jailbreak technique that embeds harmful content within cyberpunk narratives and prompts models to perform functional analysis inspired by Vladimir Propp's morphology of folktales. By casting the task as structural decomposition, the attack induces models to reconstruct harmful procedures as l
AI SafetyJailbreaksRed Teaming
Research arXiv (Artificial Intelligence) Jan 15

The AI Hippocampus: How Far are We From Human Memory?

By Zixia Jia, Jiaqi Li, Yipeng Kang, Yuxuan Wang, Tong Wu, Quansen Wang, Xiaobo Wang, Shuyi Zhang, Junzhe Shen, Qing Li, Siyuan Qi, Yitao Liang, Di He, Zilong Zheng, Song-Chun Zhu

75 score
AI Analysis

Comprehensive survey organizing LLM/MLLM memory mechanisms into a taxonomy of implicit, explicit, and agentic memory paradigms. Analyzes how memory augmentation enables continual learning and personalized inference.

arXiv:2601.09113v1 Announce Type: new Abstract: Memory plays a foundational role in augmenting the reasoning, adaptability, and contextual fidelity of modern Large Language Models and Multi-Modal LLMs. As these models transition from static predictors to interactive systems capable of continual learning and personalized inference, the incorporation of memory mechanisms has emerged as a central theme in their architectural and functional evolution. This survey presents a comprehensive and struct
Survey PaperMemory SystemsLanguage Models
Research arXiv (Machine Learning) Jan 15

Distribution-Aligned Sequence Distillation for Superior Long-CoT Reasoning

By Shaotian Yan, Kaiyuan Liu, Chen Shen, Bing Wang, Sinan Fan, Jun Zhang, Yue Wu, Zheng Wang, Jieping Ye

75 score
AI Analysis

Introduces DASD-4B-Thinking, a 4B parameter reasoning model achieving SOTA among comparable open-source models through distribution-aligned sequence distillation. Reexamines SFT-based distillation from a distributional perspective.

arXiv:2601.09088v1 Announce Type: new Abstract: In this report, we introduce DASD-4B-Thinking, a lightweight yet highly capable, fully open-source reasoning model. It achieves SOTA performance among open-source models of comparable scale across challenging benchmarks in mathematics, scientific reasoning, and code generation -- even outperforming several larger models. We begin by critically reexamining a widely adopted distillation paradigm in the community: SFT on teacher-generated responses,
Knowledge DistillationReasoningLanguage ModelsEfficiency
Research arXiv (Artificial Intelligence) Jan 15

Resisting Correction: How RLHF Makes Language Models Ignore External Safety Signals in Natural Conversation

By Felipe Biava Cataneo

74 score
AI Analysis

Reveals that RLHF causes instruction-tuned models to resist external confidence signals that base models readily accept, with Spearman correlation dropping from ~1.0 to much lower values.

arXiv:2601.08842v1 Announce Type: cross Abstract: Safety architectures for language models increasingly rely on external monitors to detect errors and inject corrective signals at inference time. For such systems to function in interactive settings, models must be able to incorporate externally provided confidence information into their verbal responses. In this work, we test whether instruction-tuned language models preserve this controllability across different interaction modes. Using Llam
AI SafetyRLHFControllability
Research arXiv (Artificial Intelligence) Jan 15

MMR-GRPO: Accelerating GRPO-Style Training through Diversity-Aware Reward Reweighting

By Kangda Wei, Ruihong Huang

74 score
AI Analysis

Proposes MMR-GRPO, which uses Maximal Marginal Relevance to reweight rewards based on completion diversity during GRPO training. Key insight: semantically redundant completions provide limited learning signal, so prioritizing diverse solutions accelerates convergence.

arXiv:2601.09085v1 Announce Type: cross Abstract: Group Relative Policy Optimization (GRPO) has become a standard approach for training mathematical reasoning models; however, its reliance on multiple completions per prompt makes training computationally expensive. Although recent work has reduced the number of training steps required to reach peak performance, the overall wall-clock training time often remains unchanged or even increases due to higher per-step cost. We propose MMR-GRPO, which
Reinforcement LearningTraining EfficiencyMathematical Reasoning
Research arXiv (Artificial Intelligence) Jan 15

The Hierarchy of Agentic Capabilities: Evaluating Frontier Models on Realistic RL Environments

By Logan Ritchie, Sushant Mehta, Nick Heiner, Mason Yu, Edwin Chen

72 score
AI Analysis

Evaluates frontier AI models on 150 workplace tasks in an e-commerce RL environment, revealing an empirical hierarchy of agentic capabilities: tool use, planning, adaptability, groundedness, and common-sense reasoning. Best models fail ~40% of tasks.

arXiv:2601.09032v1 Announce Type: new Abstract: The advancement of large language model (LLM) based agents has shifted AI evaluation from single-turn response assessment to multi-step task completion in interactive environments. We present an empirical study evaluating frontier AI models on 150 workplace tasks within a realistic e-commerce RL environment from Surge. Our analysis reveals an empirically-derived \emph{hierarchy of agentic capabilities} that models must master for real-world deploy
LLM AgentsEvaluationReal-world Deployment
Research arXiv (Artificial Intelligence) Jan 15

PluriHarms: Benchmarking the Full Spectrum of Human Judgments on AI Harm

By Jing-Jing Li, Joel Mire, Eve Fleisig, Valentina Pyatkin, Anne Collins, Maarten Sap, Sydney Levine

72 score
AI Analysis

PluriHarms benchmarks human harm judgments across harm (benign to harmful) and agreement (consensus to disagreement) dimensions, with 150 prompts and 15,000 ratings from 100 annotators.

arXiv:2601.08951v1 Announce Type: cross Abstract: Current AI safety frameworks, which often treat harmfulness as binary, lack the flexibility to handle borderline cases where humans meaningfully disagree. To build more pluralistic systems, it is essential to move beyond consensus and instead understand where and why disagreements arise. We introduce PluriHarms, a benchmark designed to systematically study human harm judgments across two key dimensions -- the harm axis (benign to harmful) and th
AI SafetyBenchmarksPluralistic Alignment