AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
MACI
This paper argues that LLMs are not fundamentally limited as pattern matchers - the real bottleneck is the lack of a System-2 coordination layer. The authors propose MACI, an architecture implementing three mechanisms: baiting (behavior-modulated debate), filtering (Socratic judging), and persistence (transactional memory) to enable goal-directed reasoning on top of LLM substrates.
Monitoring Monitorability
OpenAI introduces a framework for measuring how well we can detect misbehavior in AI systems by monitoring their chain-of-thought reasoning. The paper proposes three evaluation archetypes and a new metric (g-mean2) to track monitorability across different models and training regimes.
Test-Time Training for Long-Context LLMs
This paper shows that long-context LLMs can access millions of tokens but often fail to meaningfully use that information. The authors propose query-only test-time training (qTTT), which adapts models during inference through targeted gradient updates rather than generating more thinking tokens.
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.
JustRL
JustRL challenges the assumption that complex RL pipelines are necessary for training small language models. Using single-stage training with fixed hyperparameters, the authors achieve state-of-the-art math reasoning performance on two 1.5B models while using 2x less compute than sophisticated multi-stage approaches.
Sophia
Sophia introduces System 3, a meta-layer beyond traditional dual-process theory that enables LLM agents to maintain persistent identity and align short-term actions with long-term goals. The framework achieves 80% reduction in reasoning steps for recurring operations and 40% performance improvement on high-complexity tasks.
Budget Aware Test-time Scaling
Researchers discover that simply expanding tool-call budgets without proper awareness fails to improve agent performance. They introduce BATS (Budget Aware Test-time Scaling), a framework that makes web search agents budget-aware, enabling more strategic resource allocation and pushing the cost-performance Pareto frontier.
FrontierScience
OpenAI introduced FrontierScience, a new benchmark measuring AI capabilities for expert-level scientific reasoning across physics, chemistry, and biology. The benchmark consists of over 700 questions created and verified by domain experts, including international olympiad medalists and PhD scientists.
Vision-Language Synergy Reasoning
Researchers propose Vision-Language Synergy Reasoning (VLSR), a method that combines visual and textual reasoning to improve performance on ARC-AGI abstract reasoning tasks. The key insight is that vision excels at global pattern abstraction while language specializes in symbolic rule formulation.
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.
Pre-Training, Mid-Training, and RL Interplay
CMU researchers develop a controlled experimental framework using synthetic reasoning tasks to isolate how pre-training, mid-training, and RL-based post-training each contribute to reasoning capabilities in language models. The study reconciles conflicting views on whether RL truly extends reasoning beyond what models learn during pre-training.
Reasoning Models Ace the CFA Exams
Researchers evaluate state-of-the-art reasoning models on mock CFA exams consisting of 980 questions across all three certification levels. While previous studies reported that LLMs performed poorly on these exams, the latest reasoning models now pass all three levels, with Gemini 3.0 Pro achieving a record 97.6% on Level I.
Nanbeige4-3B
Nanbeige4-3B is a 3B parameter model pretrained on 23T tokens and fine-tuned on over 30M instructions using a Fine-Grained Warmup-Stable-Decay scheduler, Dual Preference Distillation, and multi-stage reinforcement learning. Despite its compact size, it outperforms Qwen3-8B and Qwen3-14B on reasoning benchmarks and rivals much larger models on WritingBench, demonstrating that well-engineered small models can match far larger counterparts.
DeepSeek-V3.2
DeepSeek releases V3.2, an open model that matches GPT-5 on reasoning benchmarks while introducing significant architectural and training innovations. The high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 and achieves gold-medal performance in both the 2025 IMO and IOI competitions.
Evolving Multi-Agent Orchestration
OpenBMB researchers propose a “puppeteer-style” paradigm for multi-agent LLM collaboration, where a centralized orchestrator dynamically directs agents based on evolving task states. Trained via reinforcement learning, the system achieves superior performance with reduced computational costs across math, knowledge, and software development tasks.
FINDER and DEFT
OPPO AI introduces FINDER, a fine-grained benchmark with 100 expert-curated research tasks and 419 structured checklist items for evaluating deep research agents, along with DEFT, the first failure taxonomy categorizing 14 failure modes across reasoning, retrieval, and generation dimensions.
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.
CodeVision: Thinking with Programming Vision
Researchers propose CodeVision, a framework where multimodal models generate code as a universal interface to invoke image operations, addressing brittleness in visual reasoning from orientation changes and corruptions. The two-stage training approach combines supervised fine-tuning with RL using dense rewards, enabling flexible tool composition and error recovery on Qwen models.
INTELLECT-3
INTELLECT-3 is a 106B-parameter Mixture-of-Experts model (12B active) trained with large-scale reinforcement learning, achieving state-of-the-art performance for its size across math, code, science, and reasoning benchmarks. Built on top of GLM-4.5-Air base, it outperforms many larger frontier models, including DeepSeek R1-0528, and matches GLM-4.6 (which has over 3x the parameters) on key benchmarks.
LatentMAS
LatentMAS introduces a framework enabling language model agents to collaborate directly within a continuous latent space rather than relying on text-based communication. By using last-layer hidden embeddings and a shared latent working memory, agents preserve and transfer internal representations without information loss from text serialization.
Evolution Strategies at Hyperscale
EGGROLL (Evolution Guided General Optimization via Low-rank Learning) is an evolution strategies algorithm designed to scale backprop-free optimization to large population sizes for billion-parameter neural networks. By using low-rank matrix perturbations instead of full-rank ones, EGGROLL achieves a hundredfold increase in training throughput while nearly matching pure batch inference speed.
Training LLMs with Reasoning Traces
This study investigates how reasoning traces from frontier models like DeepSeek-R1 and GPT-OSS can improve smaller language models through post-training, offering a practical pathway to distill advanced reasoning capabilities without expensive human annotation.
Cognitive Foundations for Reasoning in LLMs
Researchers develop a taxonomy of 28 cognitive elements and evaluate 192K reasoning traces from 18 models plus human think-aloud traces, finding that LLMs under-utilize cognitive elements correlated with success while relying on surface-level enumeration rather than human-like abstraction. Test-time reasoning guidance based on the framework improved performance by up to 66.7% on complex problems.
GPT-5 for Science Acceleration
OpenAI and collaborators present early case studies demonstrating GPT-5’s capabilities in accelerating scientific research across mathematics, physics, biology, computer science, astronomy, and materials science. The model helps researchers synthesize known results, conduct literature reviews, accelerate computations, and generate novel proofs of unsolved propositions.