Reasoning Is Becoming Search Again

The interesting AI signal this week was not another bigger chat model. It was a pair of reminders that useful reasoning systems increasingly look like search procedures wrapped in disciplined harnesses: one paper makes latent reasoning stochastic and parallel; one security evaluation shows frontier cyber models need decomposition, adversarial review, and proof-producing workflows to matter.

1. Recursive reasoning gets a width dimension

Sources: alphaXiv discussion, arXiv abstract, PDF, project page

What changed: Generative Recursive Reasoning introduces GRAM, a probabilistic extension of Recursive Reasoning Models. The basic move is simple but important: instead of a deterministic recursive model refining one latent state toward one answer, GRAM treats reasoning as a stochastic latent trajectory. At inference time, it can scale by both depth — more recursive refinement — and width — multiple sampled trajectories in parallel.

The paper reports improvements over deterministic recursive baselines on structured reasoning and constraint tasks including Sudoku-Extreme, ARC-AGI, N-Queens, graph coloring, and unconditional Sudoku generation. The headline is not that this is a frontier LLM replacement. It is that the authors are explicitly decoupling “more thinking” from “more visible tokens.”

Why it matters: A lot of current reasoning progress is still framed as longer chain-of-thought, tree-of-thought, or agent traces. GRAM points at a different abstraction: reasoning as latent exploration over possible solution paths. That matters because many hard tasks are not single-path derivations. They require hypothesis diversity, backtracking, uncertainty, and coverage of multiple valid solutions. In that world, deterministic recurrence is brittle; stochastic recurrence starts to look more like a learned search process.

There is also a practical inference-economics angle. If latent trajectories can be sampled in parallel, the scaling knob may become closer to “how many candidate internal computations can I afford?” rather than “how many tokens can I afford to emit?” That is a different latency/cost tradeoff, and it may pair naturally with specialized accelerators or batched inference.

Contrarian read: This is still a structured-task result, not evidence that stochastic latent recursion will help messy software agents or open-ended research systems. Sudoku and ARC-style tasks reward exactly the kind of constraint propagation GRAM is designed for. The transfer question is everything. The paper is best read as a conceptual probe: a useful direction for post-CoT reasoning architectures, not a proof that language-model agents should be rebuilt around GRAM tomorrow.

2. Cyber agents are moving from “find a bug” to “prove the exploit chain”

Sources: Cloudflare: Project Glasswing — what Mythos showed us, bookmark context

What changed: Cloudflare published a unusually practical writeup from testing Anthropic’s Mythos Preview, a security-focused frontier model, against more than 50 internal repositories. Their strongest claim is not merely that the model finds vulnerabilities. It is that Mythos can reason through exploitability: connecting low-severity primitives into chains, generating proof-of-concept code, and iterating until a finding becomes actionable.

Cloudflare’s phrasing is worth taking literally: “A suspected flaw without a working proof is speculation, and Mythos Preview closes that gap on its own.” That is a meaningful shift in the security workflow. The value is no longer just triage acceleration; it is ambiguity reduction.

Why it matters: The more general lesson is about harnesses. Cloudflare says pointing a generic coding agent at a large repo is the wrong tool for high coverage. Their effective pattern was narrow parallel tasks, structured decomposition, and deliberate disagreement between agents. In other words: capability emerges from the model plus the operating system around it.

That rhymes with the GRAM paper. Both are anti-single-trajectory stories. GRAM samples multiple latent reasoning paths; Cloudflare decomposes cyber research into many narrow adversarial workstreams. The shared trend is that frontier AI reasoning is becoming less like “ask a smart assistant” and more like “run a search process with verification pressure.”

For builders, this is the strategic point: the moat may shift from prompting skill to harness design. The best systems will know how to split tasks, preserve context, force disagreement, validate outputs, and decide when a proof is strong enough to act on.

Contrarian read: Security demos can overstate attacker readiness. This was controlled access, on Cloudflare’s own infrastructure, under a responsible disclosure process, with limited public detail on false positives and failure cases. Also, “faster patching” is not automatically enough if models also accelerate exploit development — but the operational economics of real adversaries still matter. Treat Mythos as a warning shot, not a finished map of the threat landscape.

What to watch next

• Whether stochastic latent-reasoning papers move beyond structured puzzles into code, tool use, theorem proving, or long-horizon agent benchmarks.
• Replications of GRAM-style width scaling: does sampling many latent trajectories beat sampling many natural-language chains at equal compute?
• Cyber-agent evaluations that report false positives, exploit reproducibility, time-to-proof, and coverage across large real repositories — not just compelling anecdotes.
• Harness patterns that generalize across domains: adversarial agent pairs, narrow task decomposition, persistent project memory, and proof-first workflows.

Review note

Local Obsidian source for review: /Users/hiroyoshisuzuki/Documents/Obsidian Vault/AI news/AI analysis cron/2026-05-24 Reasoning Is Becoming Search Again.md

Raw source packet: /Users/hiroyoshisuzuki/Documents/Obsidian Vault/AI news/AI analysis cron/_materials/2026-05-24/source-packet.md