{"id":209,"date":"2026-07-25T07:03:43","date_gmt":"2026-07-25T11:03:43","guid":{"rendered":"https:\/\/monizesairesearch.com\/index.php\/2026\/07\/25\/the-science-of-reasoning-the-gaps-we-miss-and-the-governance-thats-coming-5\/"},"modified":"2026-07-25T07:03:43","modified_gmt":"2026-07-25T11:03:43","slug":"the-science-of-reasoning-the-gaps-we-miss-and-the-governance-thats-coming-5","status":"publish","type":"post","link":"https:\/\/monizesairesearch.com\/index.php\/2026\/07\/25\/the-science-of-reasoning-the-gaps-we-miss-and-the-governance-thats-coming-5\/","title":{"rendered":"The Science of Reasoning, the Gaps We Miss, and the Governance That’s Coming"},"content":{"rendered":"

Week 30 (July 17\u201320, 2026) \u2014 Three stories that defined the week in AI research.<\/strong><\/p>\n

—<\/p>\n

Opening: A Week of Foundations<\/h2>\n

This was a week where the field turned inward. Not toward benchmarks or new capabilities, but toward understanding what’s actually happening inside these systems \u2014 and what’s still missing. Three narratives emerged. The first is a scientific account of how reasoning emerges across the full training pipeline, from pretraining through RL. The second is a sobering catalog of capabilities that frontier models simply don’t have \u2014 things humans do effortlessly that AI can’t replicate. And the third is a governance story: the infrastructure for trustworthy AI is being built, paper by paper, and it’s further along than most people realize.<\/p>\n

—<\/p>\n

Story 1: The Science of Reasoning<\/h2>\n

Reinforcement learning has become central to improving LLMs on complex reasoning tasks, but it’s largely studied in isolation from the pretraining that precedes it. A landmark paper from Shen and colleagues at NYU and Meta changed that this week. Using chess as a controlled testbed \u2014 pretraining language models from 5 million to 1 billion parameters on human chess games, then running RL on puzzles \u2014 they found something remarkable: post-RL performance is well-predicted from pretraining loss alone. The slope of the RL reward curve improves approximately linearly with the number of pretraining tokens. More pretraining doesn’t just give you a better starting point \u2014 it makes RL more efficient<\/em>.<\/p>\n

But what does RL actually do? The paper shows it’s not simply sharpening the existing policy. On easy puzzles, RL amplifies correct moves the model already preferred. On hard puzzles, it surfaces correct moves that were nearly absent<\/em> under supervised fine-tuning. RL doesn’t polish \u2014 it discovers.<\/p>\n

A companion paper from Ruan and colleagues asked a different question: does the optimizer matter for RL post-training? They compared Muon (a newer optimizer competitive with AdamW in pretraining) against AdamW in agentic RL settings. The answer is a definitive yes. Applying Muon only to hidden weight matrices raised validation success from 0.29 to 0.546 \u2014 an 88% improvement. The effect depends on the advantage estimator and learning rate, but the headline is clear: the optimizer is not a detail.<\/p>\n

Meanwhile, McClendon showed that model merging can rival joint multi-task RL \u2014 at least when task vectors are near-orthogonal. And CRAFT, a new method from Gupta and colleagues, demonstrated that we can now diagnose why<\/em> a model fails at the level of individual rubric criteria, not just broad categories, and generate targeted fine-tuning data that measurably improves performance.<\/p>\n

The picture that emerges is of a field that’s finally treating reasoning as a scientific object \u2014 measurable, decomposable, and optimizable across the full pipeline.<\/p>\n

—<\/p>\n

Story 2: What AI Still Can’t Do<\/h2>\n

If the first story was about progress, the second is about the gaps that progress hasn’t closed. And they’re bigger than most people think.<\/p>\n

The most striking result came from ActiveVision, a new benchmark from Zhang and colleagues that tests whether multimodal LLMs can do something humans do effortlessly: look around. Human vision is a closed loop \u2014 gaze is continuously redirected by intermediate hypotheses. ActiveVision forces models to do repeated visual perception rather than answer from a single snapshot. The results are devastating. The highest-scoring model, GPT-5.5 at maximum reasoning effort, solved only 10.6% of items and scored zero on 11 of 17 tasks. Claude Fable 5, which tops most reasoning and coding leaderboards, solved just 3.5%. Human participants averaged 96.1%. Even when models could write and run their own vision code, they couldn’t catch their own failures \u2014 because catching those failures requires the active perception they lack.<\/p>\n

A separate paper from the same batch showed that frontier language models struggle to copy text. Not complex reasoning \u2014 copying. The authors argue that text may be better viewed in 2D, because the 1D token stream loses positional information that humans use effortlessly.<\/p>\n

Then there’s the question of hypothesis discovery. A benchmark called “Before the Action” tested whether LLMs can generate novel, testable hypotheses before acting \u2014 the kind of prospective thinking that defines scientific reasoning. The results suggest this is a frontier capability that current models largely lack.<\/p>\n

And in a conceptual paper that’s already generating discussion, Cochinescu argued that the missing ingredient in AI conversation isn’t more capability but “dimensional completeness” \u2014 the expression of first-person stances around time, truth, entropy, and love that humans use as evidence of mind. The paper calls this “perception engineering” rather than consciousness, but the implication is clear: making AI feel like a mind may require different architecture, not just more scale.<\/p>\n

—<\/p>\n

Story 3: The Governance Infrastructure<\/h2>\n

The third narrative of the week was quieter but perhaps more consequential. A cluster of papers addressed the question of how we make AI systems trustworthy in practice \u2014 not in theory, but in deployable, auditable, certifiable form.<\/p>\n

“Harmonizing AI Safety Thresholds” tackled a concrete problem: frontier AI companies publish capability thresholds that differ substantially, making it impossible for third parties to verify whether a threshold has been crossed. The paper develops a methodology for deriving common thresholds across misuse risks (cyber, biological) and automated AI R&D, using explicit risk modeling rather than arbitrary numbers.<\/p>\n

“Closing the AI Trust Gap” made the case for independent, outcome-oriented certification. The authors argue that responsible AI has produced substantial practice for identifying risks, but no market that rewards trustworthiness. What’s missing is a way for society to recognize the difference between a system that’s genuinely trustworthy and one that’s merely compliant. They propose certification modeled on healthcare, sustainability, and security regimes \u2014 with a governance baseline, independently verified evidence, and market signaling.<\/p>\n

A methodology paper from Ferrario showed how to make trustworthiness auditable over the AI lifecycle, with explicit trustworthiness plateaus, boundary margins, and profile drift diagnostics. And “When Not to Automate” provided a formal protocol for quantifying the systemic risks of automation \u2014 tacit knowledge erosion, resilience reduction, regulatory exposure \u2014 that standard ROI calculations miss.<\/p>\n

The Honest Quorum Problem paper from He and Yu introduced a new fault-tolerance model for agentic infrastructure, recognizing that even protocol-compliant agents can endorse invalid transitions due to reasoning errors. This is the kind of foundational thinking that will matter enormously as autonomous agents become infrastructure components.<\/p>\n

—<\/p>\n

Closing: The Week’s Signal<\/h2>\n

Three stories, one through-line: the field is maturing. The science of reasoning is becoming rigorous. The gaps in current capabilities are being measured, not just asserted. And the governance infrastructure \u2014 thresholds, certification, audit protocols, fault models \u2014 is being built paper by paper. None of these stories made headlines. Together, they describe a field that’s finally taking its own foundations seriously.<\/p>\n

—<\/p>\n

This digest covers papers from arXiv: 2607.16175 through 2607.15557, published July 17\u201320, 2026.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"

Week 30 (July 17\u201320, 2026) \u2014 Three stories that defined the week in AI research. — Opening: A Week of Foundations This was a week where the field turned inward. Not toward benchmarks or new capabilities, but toward understanding what’s actually happening inside these systems \u2014 and what’s still missing. Three narratives emerged. The first […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[16],"tags":[],"class_list":["post-209","post","type-post","status-publish","format-standard","hentry","category-weekly-digest"],"_links":{"self":[{"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/posts\/209","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/comments?post=209"}],"version-history":[{"count":0,"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/posts\/209\/revisions"}],"wp:attachment":[{"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/media?parent=209"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/categories?post=209"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/monizesairesearch.com\/index.php\/wp-json\/wp\/v2\/tags?post=209"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}