Daily arXiv Scan: March 22, 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 (4 succeeded, 0 failed)

Consensus Picks (All 4 Models)

Behavioral Fingerprints for LLM Endpoint Stability and Identity

Selected by: Kimi K2, Claude Opus 4.6, Gemini 2.5 Pro, GPT-5

• Opus: "Addresses a problem that's obvious in retrospect but surprisingly under-studied: traditional reliability metrics don't capture when an LLM endpoint's behavior changes due to silent updates. This creates a new category of reliability engineering for AI-native applications."
• GPT-5: "The key structural move is to treat model identity as a measurable, contractible surface and to normalize drift detection as part of production hygiene, akin to canarying and anomaly detection."
• Gemini: "This isn't just an idea; it's a necessary utility. The fact that an endpoint can be 'up' but functionally broken is a nightmare scenario that this work provides a concrete, actionable framework to address."
• Kimi: "This will become the 'SSL certificate' equivalent for AI services — something enterprises refuse to deploy without, but also an enabling technology for liability frameworks."

Constitutive vs. Corrective: A Causal Taxonomy of Human Runtime Involvement in AI Systems

Selected by: Kimi K2, Claude Opus 4.6, Gemini 2.5 Pro, GPT-5

• Opus: "Attacks one of the most consequential terminological muddles in AI governance. The causal framework gives regulators and system designers a rigorous way to evaluate whether human involvement is meaningful or performative."
• GPT-5: "Moves orgs away from checkbox 'HITL' claims toward defensible, measurable runtime involvement. Regulators can demand evidence of constitutive vs corrective control."
• Gemini: "This simple reframing has profound implications. It forces system designers and regulators to be precise about the failure modes they are guarding against."
• Kimi: "This is the paper that regulators will cite in 2027 when they realize current oversight requirements are meaningless without specifying how human judgment must meaningfully constrain system behavior."

Consensus Picks (3 Models)

Regret Bounds for Competitive Resource Allocation with Endogenous Costs

Selected by: Kimi K2, Claude Opus 4.6, Gemini 2.5 Pro

• Opus: "Formalizes something that distributed systems designers deal with constantly: resource allocation where costs depend on the full allocation vector, not just your own choice. The multiplicative weights approach with interaction feedback is practically implementable."
• Gemini: "The concept of endogenous costs is fundamental. This work provides the mathematical bedrock for designing markets, social platforms, and other distributed systems where AIs will compete and cooperate."
• Kimi: "This answers why your distributed training runs keep hitting mysterious throughput cliffs — you're not accounting for how jobs change each other's execution environments."

Pair Picks (2 Models)

Towards Verifiable AI with Lightweight Cryptographic Proofs of Inference

Selected by: Claude Opus 4.6, GPT-5

• Opus: "Proposes a middle path between heavyweight zero-knowledge proofs and 'trust me' cloud claims. The statistical sampling approach exploits the fact that neural networks have structural regularities that can be checked without rerunning the full computation."
• GPT-5: "Not a cryptographic silver bullet, but exactly the kind of low-friction, high-leverage verification primitive the ecosystem needs to move beyond 'trust the endpoint.'"

Evaluating Counterfactual Strategic Reasoning in Large Language Models

Selected by: Gemini 2.5 Pro, GPT-5

• Gemini: "The counterfactual evaluation method is a powerful tool that should become standard practice for anyone claiming their model can 'reason' about incentives. The results are a sobering reminder that we are still far from understanding the true nature of LLM intelligence."
• GPT-5: "A needed cold shower. If your agent operates in environments with shifting incentives, don't trust solo LLM 'reasoning.' Add structure, or you're courting failure modes that look clever until they aren't."

Unique Finds

• Communication-Efficient and Robust Multi-Modal Federated Learning via Latent-Space Consensus (Kimi K2)
• SOL-ExecBench: Speed-of-Light Benchmarking for Real-World GPU Kernels Against Hardware Limits (GPT-5)
• Box Maze: A Process-Control Architecture for Reliable LLM Reasoning (Gemini 2.5 Pro)
• Exploring the Role of Interaction Data to Empower End-User Decision-Making In UI Personalization (Kimi K2)
• Online Learning and Equilibrium Computation with Ranking Feedback (Claude Opus 4.6)

Connecting Threads

The Trust and Verification Infrastructure Gap. The consensus papers reveal a critical blind spot: we lack basic infrastructure to verify that AI systems are doing what we think they're doing. Behavioral fingerprinting detects stability drift, while cryptographic proofs verify inference correctness. The causal taxonomy of human involvement provides a framework for meaningful oversight rather than theatrical checkboxes.

From Cardinal to Ordinal Information. Multiple papers grapple with the degradation of information quality as we move from precise feedback to rankings, or from constitutive to corrective human roles. Understanding these information bandwidth constraints is essential for designing systems where human oversight remains meaningful.

Governance as Native System Capability. Across all selections, there's a convergence on treating governance not as external constraint but as built-in system capability. Privacy, incentive alignment, and human agency are architected into fundamental design primitives rather than layered on afterward.

The Uncomfortable Reality of Fragile Competence. Whether it's LLM strategic reasoning failing under counterfactual conditions or endpoints silently changing behavior, the pattern emerges: systems that appear robust under familiar conditions can fail catastrophically when conditions shift. The solution isn't better training but better architecture and monitoring.

Statistical Baseline

• Total unique papers selected: 10
• Papers at 3+ agreement: 3 (expected by chance: 0.07)
• Papers at 2+ agreement: 5 (expected by chance: 1.72)
• Overlap significantly exceeds chance: ✓

Recommended Reading by Agreement

1. Behavioral Fingerprints for LLM Endpoint Stability and Identity (4 models)
2. Constitutive vs. Corrective: A Causal Taxonomy of Human Runtime Involvement in AI Systems (4 models)
3. Regret Bounds for Competitive Resource Allocation with Endogenous Costs (3 models)
4. Towards Verifiable AI with Lightweight Cryptographic Proofs of Inference (2 models)
5. Evaluating Counterfactual Strategic Reasoning in Large Language Models (2 models)

Methodology: Four frontier AI models independently select their top 5 papers from 80 recent arXiv submissions across AI, ML, and systems domains. Agreement patterns reveal research directions with broad model consensus.