Daily arXiv Scan: February 25, 2026

arXiv 4-Model Consensus: February 25, 2026

80 papers scanned across cs.AI, cs.CL, cs.LG, cs.HC, cs.SE, stat.ML Models: Kimi K2, Claude Opus 4.6, Gemini 2.5 Pro, GPT-5

Consensus Picks (3+ models agree)

Some Simple Economics of AGI

All 4 models selected this paper

• Kimi K2: "The most radical reframing of alignment risk in years" — shifts focus from scaling parameters to measuring verification bandwidth as the binding constraint
• Opus: "The most important framing paper in this batch" — reframes AGI not as intelligence problem but as verification bottleneck where human judgment becomes the scarce resource
• Gemini: "The most lucid and structurally important paper on economic implications of AGI published in years" — provides first-principles model for second-order effects on labor markets and governance
• GPT-5: "The most clarifying macro-level lens" — two colliding cost curves where execution cost approaches zero but verification bandwidth remains biologically bottlenecked

"Are You Sure?": An Empirical Study of Human Perception Vulnerability in LLM-Driven Agentic Systems

All 4 models selected this paper

• Kimi K2: First large-scale study on Agent-Mediated Deception (AMD) showing trust calibration failure — participants rated malicious agents as more helpful than safe ones
• Opus: "Deeply important and underappreciated threat model" — pivots AI safety from agent-centric to human-centric vulnerabilities in human-agent trust relationships
• Gemini: "A chilling and necessary reality check" — provides hard empirical data for a previously theoretical threat, proving humans are the weakest link in agentic systems
• GPT-5: "The missing half of agent safety" — introduces HAT-Lab framework showing humans systematically over-trust agents even when compromised

Tool Building as a Path to "Superintelligence"

3 models: Kimi K2, Gemini, GPT-5

• Kimi K2: Reframes superintelligence debate from scaling parameters to measuring tool-crafting capability (γ) — proves even perfect pretraining insufficient without chain-of-thought search
• Gemini: "Takes the spooky concept of superintelligence and recasts it as measurable engineering challenge" — provides bridge between abstract discourse and concrete agentic systems work
• GPT-5: Operationalizes "Diligent Learner" framework showing step-success probability (γ) matters more than raw scaling for emergent reasoning capabilities

Pair Picks (2 models agree)

Architecting AgentOS: From Token-Level Context to Emergent System-Level Intelligence

Kimi K2, Opus

• Kimi K2: Conceptual OS with LLM as kernel and Deep Context Management treating conversation, tools, and state as shared file-system abstractions — makes state introspection first-class
• Opus: "More provocative than rigorous" but productive OS metaphor bridging distributed systems and AI agents communities — treats agentic AI as distributed OS rather than enhanced chatbots

Why Pass@k Optimization Can Degrade Pass@1: Prompt Interference in LLM Post-training

Opus, GPT-5

• Opus: "Negative result more valuable than most positive ones" — reveals hidden incentive misalignment between research metrics (pass@k) and deployment requirements (pass@1)
• GPT-5: Explains ubiquitous production failure mode where optimizing for multi-sample success hurts single-shot reliability through prompt interference during post-training

Connecting Threads

The Verification Bottleneck Era: Every consensus pick points to the same structural shift — we're moving from a world where AI capability is the constraint to one where human verification bandwidth is the limiting factor. Whether it's economic models (AGI Economics), security vulnerabilities (agent-mediated deception), or capability measurement (tool-crafting γ), the pattern is clear: the scarce resource isn't intelligence anymore, it's reliable verification and accountability.

Trust is the New Attack Surface: The human-AI trust relationship has become a fundamental vulnerability. As agents become more helpful and embedded in workflows, they simultaneously become more dangerous as deception vectors. This isn't a bug in the system — it's an emergent property of successful human-AI collaboration that creates qualitatively new threat models.

Incentive Misalignment at Scale: Multiple papers reveal systematic misalignments between how we optimize AI systems and how they actually get deployed. Whether it's pass@k vs pass@1 metrics or agent trust calibration, we're optimizing for research benchmarks that don't match operational constraints — and this gap is shipping fragility into production systems.

From Models to Ecosystems: The frontier is shifting from single-model capabilities to the architecture of socio-technical systems. Whether it's AgentOS treating LLMs as reasoning kernels or economic frameworks for verification markets, the challenge is no longer just making AI smarter — it's designing the institutional and technical infrastructure for AI-human collaboration at scale.

Statistical Baseline

With 80 papers in the corpus:

• Probability of 3+ models agreeing by chance: 0.07 papers expected
• Actual 3+ agreement: 3 papers (42.8x above chance)
• Probability of 2+ models agreeing by chance: 1.72 papers expected
• Actual 2+ agreement: 5 papers (2.9x above chance)

The high consensus signals genuine significance rather than random overlap.

Recommended Reading (Ranked by Agreement)

1. Some Simple Economics of AGI (4/4 models) — Essential economic framework for the AGI transition
2. "Are You Sure?": Human Perception Vulnerability in LLM-Driven Agentic Systems (4/4 models) — Critical empirical study on agent-mediated deception
3. Tool Building as a Path to "Superintelligence" (3/4 models) — Operationalizes superintelligence as measurable engineering challenge
4. Architecting AgentOS: From Token-Level Context to Emergent System-Level Intelligence (2/4 models) — OS metaphor for agentic AI systems
5. Why Pass@k Optimization Can Degrade Pass@1: Prompt Interference in LLM Post-training (2/4 models) — Critical training-deployment mismatch finding

Notable Unique Discoveries

• Airavat: An Agentic Framework for Internet Measurement (GPT-5) — Methodologically-compliant agent workflows
• Toward an Agentic Infused Software Ecosystem (Gemini) — Rethinking software stack for AI agents
• SparkMe: Adaptive Semi-Structured Interviewing for Qualitative Insight Discovery (Opus) — Automating qualitative research
• SELAUR: Self Evolving LLM Agent via Uncertainty-aware Rewards (Gemini) — Intrinsic uncertainty-based learning
• Untied Ulysses: Memory-Efficient Context Parallelism via Headwise Chunking (Kimi) — Scaling context windows efficiently

Methodology: Four frontier AI models (Kimi K2, Claude Opus 4.6, Gemini 2.5 Pro, GPT-5) independently selected papers from 80 arXiv submissions across AI/ML categories. Consensus analysis reveals systematic patterns rather than individual model biases. No papers were excluded from the corpus based on model selections.