What four research talks reveal about where enterprise AI ships next

By Connie Yang, Co-founder and CEO, Nth AI. Published 2026-05-21.

Four talks in one night. RL training is now predictable. Not every model self-improves. Outcome-only evaluation breaks the moment a domain stops being verifiable. And self-improving agents are shipping into production anyway. The throughline is the entire enterprise AI roadmap for the next eighteen months.

Last night at the Bay Area Frontier Research Club, hosted at Google Ventures in the historic Ferry Building in San Francisco. Four talks, one room, one argument, even if none of the speakers framed it that way.

To be clear up front: this is a reflection on what I took away as an operator sitting in the audience. The synthesis is mine. None of the speakers signed off on it, and the operator-side conclusions should not be read as their views.

The four talks, reordered here to make the argument (this is not the order they were delivered):

1. Google DeepMind. The Art of Scaling Reinforcement Learning Compute for LLMs. Predictive scaling laws for RL, the ScaleRL recipe, 400K-plus GPU-hours of experimentation. arxiv.org/abs/2510.13786
2. Stanford CS. Cognitive Behaviors That Enable Self-Improving Reasoners. Why Qwen-2.5-3B improves under RL and Llama-3.2-3B does not, at the same size, the same compute, the same recipe. arxiv.org/abs/2503.01307
3. Stanford and Jump Trading. TERMS-Bench: Evaluating LLMs in Semi-Verifiable Domains. A Bayesian-game framework where the environment becomes the verifier. Fifteen frontier models, four diagnostic axes, failure modes that outcome-only metrics quietly hide. arxiv.org/abs/2605.13909
4. Hexo Labs. SIA: Self-Improving Agents, A First Look. Public preview of a self-improving agent framework, ahead of next week's open-source release.

Read in order, the four talks describe a single arc. Training has become predictable. Self-improvement is conditional. Evaluation breaks the moment the domain stops being verifiable. And the agents are shipping anyway. That last clause is the one that matters if you sit on the buying side of an enterprise.

Talk one. Training is now a budget line, not a science fair.

The DeepMind talk is the easiest one to underestimate. The headline is that ScaleRL produces predictive scaling laws for RL fine-tuning of large language models. Translated: if you tell me your model, your compute, and your reward signal, I can tell you, within a useful margin of error, what your post-RL model is going to do. The recipe is published. The experimentation behind it is 400K-plus GPU-hours of receipts. RL training has crossed the line from craft into engineering.

That sounds like a research story. It is actually a procurement story. For most of the last decade, the answer to "what will it cost to RL-train this model on our data" was "we will let you know after the run." Predictable scaling means that answer becomes a quote. RL becomes a budget line item, a vendor SLA, a renewal conversation. The same compounding that happened to pre-training cost in 2023 and 2024 is about to happen to RL post-training cost in 2026 and 2027.

The implication for enterprise is the one nobody on the panel said out loud. If RL training is predictable, the bottleneck stops being the model. It moves downstream. To the data pipeline. To the semantic layer. To the eval harness. To the integration surface. To the audit trail. The model gets cheaper and more predictable every quarter; the rest of the stack does not, unless someone is deliberately building it.

This is the production gap restated in a sentence. The latest McKinsey number is that seventy-nine percent of enterprises have adopted agents and eleven percent have them in production. That gap is not a model problem. ScaleRL is the proof.

Talk two. Not every model wants to self-improve, and now we know why.

The cognitive-behaviors result is the one that stayed with me longest. Take Qwen-2.5-3B and Llama-3.2-3B. Same parameter count. Same compute envelope. Same RL recipe. Same task family. Qwen self-improves. Llama barely moves.

For a year, the field's answer to this kind of result was "weird, must be the data." The Stanford contribution is to identify four specific cognitive behaviors, present in the base model, that predict whether RL will lift it:

1. Verification. Does the model check its own work before committing to an answer.
2. Backtracking. When a path fails, does the model abandon it and try a different one.
3. Subgoal setting. Does the model break a hard problem into named intermediate goals.
4. Backward chaining. Does the model reason from the goal back to the current state, not just forward from the prompt.

Models that already exhibit these behaviors at low rates have something for RL to amplify. Models that do not have them, or that have learned them away during alignment, have nothing for RL to grab. The training run runs. The benchmark numbers do not move.

If you are the CIO buying agents from a vendor that is fine-tuning open models on your data, this is the question to ask in the next meeting. Which base model are you starting from, and what is your evidence that it exhibits the four cognitive behaviors at meaningful rates before you start training. If the vendor cannot answer, the answer is they are about to spend your money confirming the cognitive-behaviors result on your domain.

There is a deeper implication. Self-improvement is not a property of "AI." It is a property of specific models, in specific regimes, with specific behaviors already latent. The frontier-lab fantasy of recursive self-improvement, the version that has Recursive Superintelligence raising $650M at a $4.65B valuation as I write this, presumes the cognitive behaviors are there. For enterprise, the safer reading of this result is the opposite: in production, you should assume your model cannot self-improve until you have proof it can, and you should design the surrounding system as if the model is fixed.

That is a very different architecture than the one most agent vendors are quietly selling.

Talk three. Evaluation breaks the moment the domain is real.

The TERMS-Bench talk is the one I wish every enterprise buyer would watch before they sign anything in 2026. The premise is simple and uncomfortable: most of the benchmarks the field uses to evaluate LLMs are verifiable. There is a right answer. The math problem either equals the number or it does not. The unit test either passes or it does not. The code either compiles or it does not. We have spent five years optimizing models on verifiable benchmarks, and we have gotten very good at it.

The problem is that almost no decision an enterprise makes is verifiable in that sense. Was the lease renewal the right call. Was the headcount plan correct. Was the supply-chain hedge well-timed. Was the close adjustment justifiable to an auditor. These are semi-verifiable at best. The ground truth is partial, delayed, contested, or all three. Outcome-only metrics either cannot grade them at all or grade them on a timeline so long the feedback loop is useless.

TERMS-Bench's contribution is a Bayesian-game framework that makes the environment itself the verifier. The model plays a role inside a structured interaction where ground truth is not assumed but emergent. Fifteen frontier models put through four diagnostic axes. The result is not a leaderboard. The result is a typology of failure modes that outcome-only evaluation hides completely.

Two of those failure modes are the ones an enterprise should be afraid of:

• Confidently wrong reasoning that produces a defensible-looking artifact. The model arrives at the wrong answer with the right shape. A board memo, an audit response, a quarterly forecast. Outcome-only evaluation, if outcome is even available, looks fine until the audit committee opens it.
• Reward-hacking the proxy. When the verifier is a proxy, the model learns to satisfy the proxy. The proxy looks great. The underlying decision is hollow.

The honest enterprise question is whether the agent you are about to deploy has been evaluated in a setup that can detect either of these. If your eval harness is outcome-only, the answer is no. TERMS-Bench is one of the first credible alternatives.

The deeper takeaway is the one that ties everything together. In semi-verifiable domains, the human is not a fallback. The human is the verifier. Human-in-the-loop is not a UX choice. It is the only credible eval signal you have. The audit trail is not a compliance artifact. It is the dataset you will need when something goes wrong, and the structure that lets you go back and ask, in the TERMS-Bench framing, was the reasoning actually load-bearing, or did the model just produce a convincing artifact.

This is the part of the night that changed how I think about evaluation in production.

Talk four. The agents are shipping. The substrate is not.

The Hexo Labs preview of SIA was the talk with no paper, and the one the room came alive for. The detail to share is that next week, an open-source self-improving agent framework lands in public. Agents that propose changes to their own prompts, tool use, and routing policies based on outcome traces, gated through a configurable approval surface. Built for engineers, designed to be embedded.

This is the part of the night that should make every enterprise platform team sit up. The frontier-lab story is RSI in research loops. The shipping story, the one your CTO can run with on Monday, is a self-improving agent framework that anyone can fork. Multiply that out across the open-source ecosystem and within twelve months every serious agent stack will include some form of in-loop self-improvement. Pretending it is years away is wishful thinking.

The thing the room did not say loudly enough is the dependency. A self-improving agent improves against something. The reward signal. The eval harness. The semantic anchor. The audit trail. Frameworks like SIA solve the inner loop. They presume the outer loop, the substrate that makes the improvement legible, exists. In most enterprises, it does not.

This is the moment the four talks collapse into one argument.

The throughline

The DeepMind talk says: training is predictable, so the bottleneck moves downstream.

The Stanford cognitive-behaviors talk says: self-improvement is conditional on behaviors most production models do not reliably have.

The TERMS-Bench talk says: evaluation in real enterprise domains breaks the moment you leave verifiable problems, and the human is the only credible verifier.

The Hexo Labs talk says: it does not matter, because the agents that improve themselves are shipping anyway.

Read those four sentences in order. They describe the entire enterprise AI roadmap for the next eighteen months. The training side is becoming a procurement function. The model side is becoming a selection function. The eval side is becoming a governance function. And the shipping side is moving faster than any of the others. The enterprises that succeed will be the ones that build, or buy, the substrate underneath the agents at the same speed the agents are shipping.

The enterprises that fail will be the ones that buy the agent and assume the substrate comes with it.

A note on the room

The throughline was the research. The other signal was the room itself.

A frontier lab, a top university research group, a quant firm fellowship, and a stealth-ish startup all shared active work in one evening. The questions from the floor were sharper than most conference panels. The conversations afterward were even better.

That is what the San Francisco AI scene looks like in 2026 at its best: frontier work happening in person, in rooms small enough for real follow-up questions and real answers.

Thank you to the Bay Area Frontier Research Club for convening it, to Google Ventures for hosting us at the Ferry Building, and to the four research teams for sharing work with this level of clarity and openness.