.png){kind=logo}
In 2019 I ran across Counterfactual Regret Minimization while searching for something completely different — three citations away. There were four researchers at the University of Alberta (Martin Zinkevich, Michael Johanson, Michael Bowling, and Carmelo Piccione), authors of the CFR paper back in 2007. Nobody beyond a tiny academic community noticed.
Nowadays, CFR derivatives manage real capital at quant trading shops, help develop adversarial attacks for Pentagon-funded cybersecurity tools, and run negotiation AI platforms that make millions of dollars worth of procurement decisions. Over the past couple of years I have followed this journey and honestly it is one of the most impressive examples of technology transfer I've ever witnessed.
Almost all optimization techniques try to find the single "optimal" action. Try all the options, pick the winner. Reasonable sounding until you're in a situation where the "optimal" action changes depending on what another player chooses — and they will intentionally identify any exploitable patterns in your actions.
CFR does not try to optimize any single action. Instead, it computes regret. After each decision, it simply asks one question: how good would each option have been if it was chosen instead? Do that many times over many simulated scenarios and update the strategy to minimize the total regret; and you'll end up with some mathematical properties that are very interesting indeed — including a Nash equilibrium where there isn't anything that any opponent can do to improve their outcome by adjusting their strategy.
The "counterfactual" portion is really where Zinkevich's group found the breakthrough. They realized how to break down the huge multi-agent problem into many independent decision points, calculate the regret at each point independently and combine these to create an efficient solution. Prior to this paper, those types of problems were never computationally tractable.
Libratus was developed by Carnegie Mellon using Monte Carlo CFR and defeated four top expert humans in a 120,000-decision competition in January 2017. The four human experts lost $1.77 million in simulated stakes. None of them won.
Here is the part of the story that captured the attention of non-academics: Libratus did not rely on a fixed strategy. Rather it adapted and changed its approach in real time during live competition to refine its strategy based on patterns of behavior that it identified in its opponents' plays. This adaptability — strategy that develops against adversary activity in real time — was novel. And that's precisely what numerous industries have been seeking.
Pluribus — another CFR derivative — was developed two years after Libratus by Meta AI and solved six-player settings by computing subproblems in real time rather than attempting to pre-compute solutions for an impossibly large decision tree. The architectural design pattern of solving subproblems locally and stitching together a cohesive global strategy has proven to be the key to enabling CFR to be applied outside of university research labs.
A quant researcher I am familiar with transitioned directly from the imperfect-information games community to a mid-sized hedge fund. He needed someone who understood competitive bidding in illiquid markets — submitting orders without having access to other participants' positions, prices, or intent. His experience with CFR was clearly far more valuable than an MBA in finance.
That makes perfect sense once you consider that financial markets are textbook examples of imperfect information environments. Many agents simultaneously make decisions with private knowledge. An agent's optimal course of action is dependent upon all other agents' behavior — but since only indirect evidence of others' behavior is available via price movements and order flow — it is difficult to infer.
By 2024, at least three major quant firms have published research that builds off of CFR ideas. They referred to the use of CFR as "regret-based portfolio optimization," "equilibrium-seeking algorithms," etc. However, the intellectual lineage was evident. One paper even referenced Zinkevich et al. 2007 in its bibliography.
What CFR provides quant firms that standard quant models do not: robustness against adversarial exploitation. Traditional quant models assume market behavior follows statistical patterns — prices follow distributions, correlations exist. CFR assumes that there exists an adversary that seeks to exploit any predictability exhibited by your behavior. Given that algorithmic trading dominates volumes in 2026 — that assumption is simply more honest.
Another significant migration route is through red-team testing. Red teams attempt to emulate the mindset of attackers: probe defenses, identify vulnerabilities and exploit them prior to actual attackers doing so. Decisions made sequentially under uncertainty against an adaptive defender. Sound familiar?
Modern red-team tools utilize CFR-derivative algorithms to plan attack sequences based on observed defensive reactions. It does not employ a static playbook — but rather adjusts its approach based on what the defense reveals through its reactions. Just as Libratus did when it adjusted daily.
Since 2021 DARPA has funded multiple programs investigating game-theoretic AI for adaptive cyber defense. Their "adaptive cyber defense" program explicitly references the imperfect information game theory literature. The pipeline from academia to defense is relatively direct and well-documented.
I spoke with a security researcher last year who stated his tool was similar to Libratus "for networks." I questioned him further regarding how accurate that analogy was — and he replied "around 70% — we solve smaller subgames but the regret-minimization loop remains the same."
Between 2022 and 2025 several startups founded negotiation AI that negotiates procurement deals, contract terms and pricing agreements on behalf of companies. The challenge faced by these startups is similar to that for which CFR was originally created: private information shared among parties involved in sequential commitments constrained by subsequent options available to each party — optimal strategies depend on their respective beliefs regarding their counterpart's position.
Initial negotiation AI employed rule-based methods. "When they offer X — respond with Y." Such methodologies proved brittle. Experienced human negotiators quickly recognized any patterns presented — typically within three iterations of offers exchanged.
Current generation negotiation AI employs belief-space planning — probability distributions representing possible positions held by counterparts, updated with each offer — guiding multi-step strategies accounting for informational revelations throughout the negotiation process. I have viewed demonstrations. The AI adapts its strategy in ways that appear to be random — however, are strategically calculated to prevent the counterpart from developing an accurate understanding of its position.
One founder stated his system exceeds performance achieved by experienced human negotiators on standardized procurement cases by approximately 8%. Although this may not seem like a lot until you recognize they process thousands of negotiations per month. On a large scale, 8% translates to real dollars.
Zinkevich's group solved an abstract game-theoretical problem when they wrote their 2007 paper. They sought to determine equilibria in large extensive-form games. Period.
The notion that their method would impact how hedge funds execute trades, how the Pentagon plans cyber warfare operations, and how enterprise companies negotiate contracts — likely wasn't anticipated by any member of Zinkevich's group. However, this is how foundational research operates. You solve the abstract problem. Then applications find you — often decades later.
CFR did more than simply advance game theory. It provided a practical methodology for rational decision-making under conditions of adversarial uncertainty. That methodology is now being utilized in systems managing billions of dollars — and defending critical infrastructure.
Not too shabby for a paper that barely anyone noticed when it was released.
