From the Archive

From the Archive — 2026-06-16

2026-06-16

A short reading list of papers that landed this week. Each one sits at the edge of a longer conversation already underway in the research literature.

From AGI to ASI

Tim Genewein, Matija Franklin, Alexander Lerchner, et al. · arXiv:2606.12683

As AI researcher timelines continue to compress, the question of what comes after human-level AGI has shifted from theoretical curiosity to urgent practical concern—yet the paths from AGI to superintelligence remain poorly mapped. This report stakes out that terrain by proposing four potential trajectories (scaling, paradigm shifts, recursive improvement, and multi-agent emergence) while acknowledging that progress uncertainties make the distinction between a single transformative step and a cascade of distributed breakthroughs genuinely unclear. What's particularly unresolved is whether recursive improvement loops that accelerate AI research itself represent a fundamental bottleneck or a runaway dynamic—and relatedly, whether human-AI research collaboration can remain competitive as systems approach and potentially exceed organizational-level capability. The report's call for interdisciplinary preparation suggests the real uncertainty isn't whether ASI-level systems are possible, but whether institutions can anticipate and shape change across multiple simultaneous domains rather than bracing for a single discontinuity.

Go deeper into Agentic Systems and Tool Use

Capability Boundaries And Diversity CollapsePersuasion And Epistemic Distortion
Explore these inquiring lines →

q0: Primitives for Hyper-Epoch Pretraining

Bishwas Mandal, Shmuel Berman, Akshay Vegesna, et al. · arXiv:2606.03938

As compute budgets outpace text availability, the pretraining paradigm faces a fundamental constraint: a single model exhausts its learning signal long before exhausting its allocation. This work proposes a conceptual reorientation toward population-based training, where the budget funds diversity rather than refinement—a shift with deep implications for how we think about model aggregation. The approach echoes existing insights about how models trained on diverse imperfect experts develop implicit consensus, but inverts the setup: instead of training on external experts, here the population is grown in parallel and combined at inference time. The use of chain distillation to compound quality across generations, paired with a learned prior to weight ensemble members, also connects to broader questions about how to allocate compute between parallel coverage and sequential refinement. What remains unclear is whether this population strategy generalizes beyond the multi-epoch regime—whether the principles that make diverse trajectories valuable during saturation hold when architectural choices and inference constraints reshape the efficiency frontier.

Go deeper into Model Architecture and Internals

Scaling, Sparsity & Data Trade-offsTest-Time Compute ScalingTraining Dynamics And GeneralizationCapability Boundaries And Diversity Collapse
Explore these inquiring lines →

Consensus is Strategically Insufficient: Reasoning-Trace Disagreement as a Knowledge-Representation Signal

Michał Wawer, Jarosław A. Chudziak · arXiv:2606.04223

The standard wisdom in multi-agent AI design has long favored consensus—through voting, debate, or aggregation protocols—as the path to reliability, yet recent work on agent confidence as a routing signal and consensus failures in LLM teams has begun exposing the limits of that approach. This paper reframes disagreement itself as potentially informative rather than merely an error to be corrected, arguing that in normatively complex domains like content moderation, the *pattern* of disagreement—whether agents diverge in reasoning, conclusions, or both—carries meaning that should shape downstream decisions rather than be suppressed. By mapping disagreement into symbolic states that distinguish convergent from divergent reasoning, the work sits at the intersection of two unresolved tensions: whether LLM debate mechanisms can capture what human expert disagreement actually signals, and whether routing decisions informed by disagreement patterns can gracefully handle the kind of value-laden uncertainty that no amount of agent coordination will fully resolve. The question emerging from this framework is whether we've been asking the wrong question all along—not "how do we eliminate disagreement?" but rather "what decisions are best made *by* disagreement, rather than *despite* it?"

Go deeper into Agentic Systems and Tool Use

Silent Agreement In Multi-Agent ReasoningAgent Reliability ArchitectureTest-Time Compute ScalingHuman-AI Cognitive Boundaries
Explore these inquiring lines →

Position: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!

Subbarao Kambhampati, Karthik Valmeekam, Siddhant Bhambri, et al. · arXiv:2504.09762

The push to improve language model reasoning through intermediate token generation has become orthodoxy, yet whether these traces actually perform reasoning or merely imitate its appearance remains contested—a distinction that matters deeply for how we deploy and evaluate these systems. Recent work has further complicated the picture by showing that deliberately corrupted reasoning traces can match or exceed the performance of correct ones, and that chain-of-thought explanations may optimize for model performance rather than user understanding. These findings suggest a growing tension between what intermediate tokens do for model accuracy and what they appear to reveal about the model's cognition—a gap that hinges on whether we're studying actual reasoning processes or learned surface patterns that merely resemble human problem-solving.

Go deeper into Reasoning, Retrieval, and Evaluation

Reasoning Model Failure ModesChain-of-Thought FaithfulnessReasoning Model Quality & TrainingReasoning Trace Reliability
Explore these inquiring lines →

The Last Human-Written Paper: Agent-Native Research Artifacts

Jiachen Liu, Jiaxin Pei, Jintao Huang, et al. · arXiv:2604.24658

The compression of scientific research into linear papers has long been understood as a necessity of human communication, but this work reframes it as a fundamental liability for AI-native research—one that becomes acute when agents must not only read papers but extend them. Rather than accept the traditional publication format as inevitable, the authors propose restructuring research artifacts around what machines actually need: executable specifications, preserved dead ends, and grounded evidence chains. This connects to broader tensions in AI-accelerated science: autonomous research mechanisms work better together than apart, yet AI-generated research outpaces verification in ways that may require new venue models with closed-loop automated review. The question emerging is whether formalizing research as executable packages will accelerate discovery or whether the loss of narrative flexibility might constrain the kinds of conceptual leaps that have historically driven paradigm shifts.

Go deeper into Agentic Systems and Tool Use

Capability Boundaries And Diversity CollapseSilent Agreement In Multi-Agent ReasoningReasoning Trace ReliabilityPersuasion And Epistemic Distortion
Explore these inquiring lines →

All briefs →