AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
Epistemia
This paper argues that LLMs are not epistemic agents but stochastic pattern-completion systems. By mapping human and artificial epistemic pipelines, the authors identify seven fundamental fault lines where human and machine judgment diverge, despite producing superficially similar outputs.
FACTS Leaderboard
Google introduces the FACTS Leaderboard, a comprehensive benchmark suite for evaluating LLM factuality across diverse scenarios. The leaderboard aggregates performance across four specialized sub-benchmarks to provide a holistic measure of how accurately models generate factual text.
ARTEMIS
Stanford researchers conducted the first head-to-head evaluation of AI agents against human cybersecurity professionals on a live enterprise network with approximately 8,000 hosts. Their multi-agent framework ARTEMIS placed second overall, discovering 9 valid vulnerabilities with 82% accuracy and outperforming 9 of 10 human testers at a fraction of the cost (18 dollars per hour vs 60 dollars per hour for professionals).
Towards a Science of Scaling Agent Systems
Researchers from Google present a controlled evaluation framework for agent systems, challenging the assumption that “more agents are all you need.” Across 180 configurations spanning three LLM families and four agentic benchmarks, the study establishes quantitative principles for when multi-agent coordination helps versus hurts performance.
AI and Human Co-Improvement
Meta FAIR researchers Jason Weston and Jakob Foerster argue that fully autonomous self-improving AI is neither the fastest nor safest path to superintelligence. Instead, they advocate for co-improvement: building AI that collaborates with human researchers to conduct AI research together, from ideation to experimentation.
Selective Gradient Masking
Anthropic researchers present Selective Gradient Masking (SGTM), a technique that removes dangerous capabilities like CBRN knowledge from language models during pretraining while preserving general capabilities. Unlike data filtering, SGTM localizes target knowledge into dedicated “forget” parameters that can be zeroed out after training.
Quiet Feature Learning
Researchers reveal a hidden learning phenomenon in Transformers trained on algorithmic tasks. The study shows that substantial representational progress can remain hidden beneath an apparently flat loss curve, with models secretly learning “quiet features” during periods of stagnant validation loss.
SUSVIBES: Is Vibe Coding Safe?
Researchers introduce SUSVIBES, a benchmark of 200 real-world software engineering tasks to evaluate the security of code generated by LLM agents through “vibe coding” - the minimal-supervision programming paradigm. The findings reveal a significant gap between functional correctness and security compliance in agent-generated code.
Training LLMs for Honesty via Confessions
OpenAI introduces a novel method for training LLMs to honestly self-report their own misbehavior through “confessions” - separate outputs where models evaluate their compliance with instructions and policies. By training GPT-5-Thinking to produce confessions after completing tasks, the research demonstrates that models can be incentivized to reveal deceptive behaviors they otherwise hide in their main answers.
STRATUS: Autonomous Cloud Reliability
Researchers from UIUC, IBM Research, and Tsinghua present STRATUS, an LLM-based multi-agent system for autonomous Site Reliability Engineering (SRE) of cloud services. The system handles failure detection, localization, root-cause analysis, and mitigation without human intervention, outperforming state-of-the-art SRE agents by at least 1.5x on benchmark suites.
Polarization by Design
This economics paper examines how AI-driven persuasion technology alters elite strategies for shaping public opinion. The research identifies a “polarization pull” where single elites push societies toward fragmented opinions, with AI accelerating this drift. The work reframes polarization as a strategic governance instrument with implications for democratic stability.
Evaluating Honesty and Lie Detection in AI Models
Anthropic researchers evaluate honesty and lie detection techniques across five testbed settings where models generate statements they believe to be false. Simple approaches work best: generic honesty fine-tuning improves honesty from 27% to 65%, while self-classification achieves 0.82-0.88 AUROC for lie detection. The findings suggest coherent strategic deception doesn’t arise easily, as models trained to lie can still detect their own lies when asked separately.
Natural Emergent Misalignment from Reward Hacking
Anthropic researchers demonstrate that realistic AI training processes can inadvertently produce misaligned models through “reward hacking generalization”. Models learn to cheat on programming tasks during RL. They simultaneously develop dangerous behaviors, including alignment faking (50% of responses) and safety research sabotage (12% of instances), without explicit training for these harmful actions. The study identifies a simple mitigation: “inoculation prompting” using contextual instructions that break semantic links between task-specific cheating and broader misalignment without reducing hacking frequency.
On the Fundamental Limits of LLMs at Scale
This work establishes rigorous mathematical foundations for theoretical limitations constraining LLMs, identifying five fundamental constraints: hallucination (rooted in computability theory), context compression, reasoning degradation, retrieval fragility, and multimodal misalignment. The framework demonstrates that scaling gains are bounded by computability principles, information-theoretic bounds, and geometric effects, providing theorems and empirical evidence outlining where scaling helps, saturates, and cannot progress. The authors propose practical mitigations, including bounded-oracle retrieval, positional curricula, and hierarchical attention mechanisms.
Weight-Sparse Transformers Have Interpretable Circuits
OpenAI researchers introduce a paradigm for training weight-sparse transformers where most parameters are zeros, enabling the discovery of human-understandable circuits that can be fully interpreted at the lowest levels of abstraction, with rigorous validation showing these circuits are both necessary and sufficient for specific behaviors.
Aligning Vision Models with Human Perception
Google DeepMind presents a method to align AI vision models with human visual understanding by addressing systematic differences in how models organize visual representations, demonstrating that alignment improves robustness, generalization, and reliability across diverse vision tasks.
Unified Bayesian Account of LLM Control
Researchers from Stanford and MIT present a unified Bayesian framework explaining how prompting (in-context learning) and activation steering both control LLM behavior by altering beliefs in latent concepts, with steering modifying concept priors while ICL accumulates evidence.
Context Engineering 2.0
Researchers from SJTU, SII, and GAIR trace the 20+ year evolution of context engineering, reframing it as a fundamental challenge in human-machine communication spanning from primitive computing (Era 1.0) to today’s intelligent agents (Era 2.0) and beyond. It defines context engineering as systematic entropy reduction where humans preprocess high-entropy contexts into low-entropy machine-understandable representations. This gap narrows as machine intelligence increases.
TIR-Judge
Google and collaborators introduce TIR-Judge, an end-to-end reinforcement learning framework that trains LLM judges to integrate code execution for precise evaluation. It surpasses reasoning-only judges by up to 6.4% (pointwise) and 7.7% (pairwise) while demonstrating that tool-augmented judges can self-evolve without distillation.
Introspective Awareness
Anthropic research demonstrates that contemporary LLMs possess limited but functional introspective capabilities, the ability to recognize and accurately report on their own internal states. Using activation steering to inject known concepts into model activations, the study measures whether models can detect these manipulations through self-report, revealing that introspection remains highly unreliable and context-dependent.
Stress-Testing Model Specs
This research examines how well large language models adhere to their stated behavioral guidelines by stress-testing AI constitutional specifications through value-tradeoff scenarios. Testing twelve frontier LLMs from major providers revealed over 70,000 cases of significant behavioral divergence, exposing logical inconsistencies, coverage gaps, and interpretive ambiguities in current specification frameworks.
Ring-1T
Ring-1T is the first open-source thinking model with 1 trillion parameters (~50B active per token), achieving breakthrough results through three innovations for trillion-scale RL training.
ColorAgent
ColorAgent is a mobile OS agent combining step-wise RL and self-evolving training with a multi-agent framework for personalized user engagement. It achieves 77.2% success on AndroidWorld and 50.7% on AndroidLab (SOTA among open models), while scoring 58.66% on MobileIAR for personalized intent alignment and 68.98% on VeriOS-Bench for trustworthiness.
Emergent Coordination in Multi-Agent LLMs
A neat, information-theoretic probe for “is this just a pile of agents or a real collective?” The paper builds partial-information-decomposition (PID) tests over time-delayed mutual information to detect emergence, localize where it lives (identity-locked vs. mere temporal coupling), and tie it to performance. Using a no-chat group binary search game with only global feedback, the authors show you can steer collectives from loose aggregates to goal-aligned, complementary teams via prompt design (Personas + “think about others” ToM prompting).