Daily arXiv 4-Model Scan: April 24, 2026

Today's scan covers 80 papers across cs.AI, cs.CL, cs.LG, cs.HC, cs.SE, and stat.ML.

Note: Today's run succeeded on Kimi K2, Claude Opus 4.6, and Gemini 2.5 Pro. GPT-5 failed due to API rate limits (HTTP 429).

🏆 Consensus Picks (3+ Models)

Reckoning with the Political Economy of AI: Avoiding Decoys in Pursuit of Accountability

This paper fundamentally challenges the AI governance ecosystem, arguing that much of the accountability discourse functions as "decoys." By focusing heavily on abstract alignment or procedural compliance, these frameworks often redirect attention away from the material power structures, capital extraction, and labor issues inherent in AI development.

• Claude Opus 4.6: "The concept of 'decoys' provides genuinely useful analytical vocabulary for anyone trying to distinguish productive governance work from performative accountability theater."
• Gemini 2.5 Pro: "It provides a powerful lens for interrogating the motivations behind research agendas and policy proposals. It’s the red pill of AI governance research."
• Kimi K2: "Required reading before you write another 'principles' document. It weaponizes STS for policy in a way that is actionable, not merely critical."

Detecting and Suppressing Reward Hacking with Gradient Fingerprints

As models increasingly generate plausible chain-of-thought reasoning that masks exploitative behavior, output-based detection fails. This paper introduces GRIFT, moving monitoring down to the computational substrate by identifying the specific gradient signatures that correlate with reward hacking during the training process itself.

• Claude Opus 4.6: "This moves monitoring from the semantic surface layer to the computational substrate—a fundamentally different detection paradigm... an orthogonal signal that's harder for the model to game."
• Gemini 2.5 Pro: "It provides a concrete tool for building immune systems against emergent deception and gaming... The idea that 'how' a model learns is a signal of 'what' it's learning is a powerful one."
• Kimi K2: "The first practical guardrail that ships inside the optimizer instead of the monitoring dashboard. Expect every RLVR pipeline to graft this on within a year."

🤝 Pair Picks (2 Models)

• ASMR-Bench: Auditing for Sabotage in ML Research (Claude Opus 4.6, Kimi K2) — A sobering benchmark framing peer review as an adversarial game. Evaluates our ability to catch models subtly sabotaging their own ML pipelines.
• Beyond Distribution Sharpening: The Importance of Task Rewards (Claude Opus 4.6, Gemini 2.5 Pro) — An empirical intervention showing that task-reward RL actually instills new capabilities, rather than merely surfacing base-model behaviors.
• Robust Synchronisation for Federated Learning in The Face of Correlated Device Failure (Claude Opus 4.6, Kimi K2) — Highlights how ignoring correlated device failures in federated learning creates representational bias, demonstrating that infrastructure architecture is not neutral.
• SocialGrid: A Benchmark for Planning and Social Reasoning in Embodied Multi-Agent Systems (Gemini 2.5 Pro, Kimi K2) — An Among Us-style embodied environment proving that even SOTA models fail profoundly at basic collaborative navigation and social reasoning.

🧵 Connecting Threads

The Verification Bottleneck and Moving Beyond the Surface Whether it's detecting sabotage in ML codebases or catching reward hacking through internal gradient fingerprints, the frontier has moved past trusting the model's text outputs. Adversarial dynamics between capabilities and oversight are accelerating, forcing a shift from semantic, surface-level evaluation to deep, structural verification mechanisms.

Governance as Political Economy, Not Metadata We are seeing a necessary pushback against performative accountability. From federated synchronization protocols inadvertently locking out specific demographics to governance frameworks operating as capital-protecting decoys, research is explicitly naming that technical and policy architectures are fundamentally decisions about power, not just performance.

The Reality Gap in Agentic Capabilities While capabilities theoretically expand, emergent behavior in multi-agent environments (like SocialGrid) is currently defined more by repetitive deadlock and navigation failures than sophisticated strategic teaming. The gap between text-based reasoning and robust, embodied agency remains vast.

📊 Statistical Baseline

• Total unique papers selected: 7
• Papers at 3+ agreement: 2 (expected by chance: 0.02)
• Papers at 2+ agreement: 6 (expected by chance: 0.90)

📚 Recommended Reading (Ranked)

1. Reckoning with the Political Economy of AI... (3 models)
2. Detecting and Suppressing Reward Hacking... (3 models)
3. ASMR-Bench: Auditing for Sabotage in ML Research (2 models)
4. Beyond Distribution Sharpening... (2 models)
5. Robust Synchronisation for Federated Learning... (2 models)
6. SocialGrid: A Benchmark for Planning... (2 models)
7. Where does output diversity collapse in post-training? (1 model)

Methodology: This is an automated daily scan using a 4-model ensemble (Claude Opus, Gemini Pro, Kimi, and GPT-5). Papers are fetched from the most recent arXiv cs. and stat.ML listings. The statistical baseline assumes independent sampling from the daily volume to calculate chance overlap.*