Daily arXiv Scan: March 26, 2026

Daily arXiv 4-Model Comparison: March 26, 2026

80 papers scanned across cs.AI, cs.CL, cs.LG, cs.HC, cs.SE, stat.ML

Models participating: Kimi K2, Claude Opus 4.6, GPT-5 (3 succeeded) Failed: Gemini 2.5 Pro (HTTP 503 Service Unavailable)

Consensus Picks (3+ Models Agree)

Large Language Model Guided Incentive Aware Reward Design for Cooperative Multi-Agent Reinforcement Learning

Chosen by: Kimi K2, Claude Opus 4.6, GPT-5

• Kimi K2: Demonstrates LLMs can synthesize game-theoretically valid reward functions from natural language, bridging intention and mathematical incentive design. Creates automated pipeline from vague goals to rigorously tested incentive structures.
• Opus 4.6: Uses LLMs to automatically synthesize executable reward programs with formal validity constraints. Most directly relevant to incentive design in complex multi-agent coordination problems.
• GPT-5: Automates reward shaping by using LLMs to synthesize executable reward programs constrained by formal validity envelope. Turns incentive design into program search with testable mechanisms.

The Stochastic Gap: A Markovian Framework for Pre-Deployment Reliability and Oversight-Cost Auditing in Agentic Artificial Intelligence

Chosen by: Kimi K2, Claude Opus 4.6, GPT-5

• Kimi K2: Introduces measure-theoretic framework for quantifying "governability gap" in agentic systems. Formalizes oversight as first-class constraint in policy optimization, embedding human-in-the-loop constraints directly into agent utility functions.
• Opus 4.6: Tackles the most important unsolved infrastructure problem in agentic AI deployment: auditing reliability of stochastic agents before deployment. Introduces blind-spot mass and oversight cost metrics.
• GPT-5: Reframes agentic AI reliability as measure-theoretic Markov process, introducing actionable quantities for blind-spot mass and oversight costs. Provides "reliability/oversight SLO" framework for agents.

Pair Picks (2 Models Agree)

Claudini: Autoresearch Discovers State-of-the-Art Adversarial Attack Algorithms for LLMs

Chosen by: Claude Opus 4.6, GPT-5

• Opus 4.6: Demonstrates autonomous research pipeline discovering novel adversarial attacks that outperform 30+ existing methods. Shows AI systems exhibiting capabilities not explicitly programmed, with immediate safety implications.
• GPT-5: Uses autonomous research pipeline to invent new white-box adversarial attacks beating existing methods. Signals shift where security risk originates from scalable agentic research loops rather than human red teamers.

The Free-Market Algorithm: Self-Organizing Optimization for Open-Ended Complex Systems

Chosen by: Kimi K2, Claude Opus 4.6

• Kimi K2: Abandons fitness functions entirely, replacing them with distributed supply-demand dynamics where competition emerges from agent interactions. Creates bridge between AI optimization and economic governance.
• Opus 4.6: Introduces metaheuristic where fitness is emergent from distributed supply-demand dynamics. Allows agents to discover rules and compete with no centralized controller, relevant for incentive design.

Unique Finds (Single Model)

Kimi K2 Selections

• Evidence of an Emergent "Self" in Continual Robot Learning — Proposes empirical method for detecting self-awareness through cognitive invariants across learning episodes
• AI-Supervisor: Autonomous AI Research Supervision via a Persistent Research World Model — Shift to artificial research supervisors with persistent world models across projects

GPT-5 Selections

• ClawKeeper: Comprehensive Safety Protection for OpenClaw Agents Through Skills, Plugins, and Watchers — Full-stack safety architecture for agent runtime with high-risk privileges
• Comparing Developer and LLM Biases in Code Evaluation — Framework revealing systematic weighting differences between human and LLM code judges

Claude Opus 4.6 Selection

• Retrieval Improvements Do Not Guarantee Better Answers: A Study of RAG for AI Policy QA — Shows better retrieval doesn't reliably improve answer quality in policy analysis RAG systems

Connecting Threads: The Systems Turn

Today's papers reveal a field pivoting from component optimization to systems-level thinking:

The Governance Infrastructure Gap is Widening. We lack frameworks for auditing stochastic agent trajectories (Stochastic Gap) while autonomous agents discover novel attack vectors faster than humans can respond (Claudini). The capability curve is outpacing our governance infrastructure.

Emergence as the Default Mode. From emergent fitness in optimization (Free-Market Algorithm) to emergent attack discovery (Claudini) to emergent self-awareness (Robot Learning), complex AI systems exhibit emergence as their natural operating mode, not an edge case.

Component Improvements Don't Compose. The RAG policy paper shows better retrieval doesn't yield better answers. The MARL reward paper addresses the same compositional failure—better individual agents don't guarantee better coordination without proper incentive alignment.

LLMs as Meta-Designers. Both the reward design and attack discovery papers use LLMs not as end-user tools but as designers of other systems—reward functions and attack algorithms. This introduces new feedback loops that existing governance frameworks don't address.

Economic Thinking Pervades Everything. Whether it's oversight costs (Stochastic Gap), market dynamics (Free-Market Algorithm), or incentive compatibility (MARL rewards), economic reasoning is becoming central to AI system design and governance.

Statistical Baseline

• Total unique papers selected: 9
• Papers with 3+ model agreement: 2 (expected by chance: 0.02)
• Papers with 2+ model agreement: 4 (expected by chance: 0.90)

The consensus picks show 100x over-representation compared to random chance, indicating genuine signal detection by multiple models.

Recommended Reading (By Agreement Level)

1. Large Language Model Guided Incentive Aware Reward Design for Cooperative Multi-Agent Reinforcement Learning (3 models) — Most practical for incentive design
2. The Stochastic Gap: A Markovian Framework for Pre-Deployment Reliability and Oversight-Cost Auditing in Agentic Artificial Intelligence (3 models) — Most important for AI governance
3. Claudini: Autoresearch Discovers State-of-the-Art Adversarial Attack Algorithms for LLMs (2 models) — Most surprising capability result
4. The Free-Market Algorithm: Self-Organizing Optimization for Open-Ended Complex Systems (2 models) — Most intellectually ambitious

Methodology: Each model independently scanned 80 papers and selected 4-5 based on novelty, significance, and relevance to frontier AI research. Papers are ranked by multi-model agreement, with statistical baselines calculated assuming uniform random selection.