Salesforce Patents a Way to Train AI to Reason Through Problems Without a Bigger AI Doing the Teaching
Teaching an AI to think step by step usually requires an already-capable AI to show it how. Salesforce's new patent describes a way to skip that dependency entirely, letting a smaller model bootstrap its own reasoning ability from scratch.
How Salesforce's self-taught reasoning trick works
Imagine you're learning to solve math problems, but there's no teacher available. Instead, you write out dozens of attempts, check which answers came out right, and use only those successful attempts to study from. That's roughly what this Salesforce patent describes for AI.
Right now, when companies want an AI to "show its work" (walking through logic step by step before giving an answer), they typically train it by having a more powerful AI model generate those worked examples first. Salesforce's approach cuts out that middleman. A standard AI model is shown a few examples of careful, step-by-step thinking, then asked to generate many versions of its own reasoning attempts. The ones with correct final answers get kept; the rest get discarded.
The surviving examples then become the training material that teaches the model to reason more carefully going forward. The system keeps refining itself on fresh questions it hasn't seen before, so it improves over time without any outside help.
How the filtering and reward loop builds reasoning data
The patent describes a training pipeline with a few distinct stages:
- Prompting with examples: A base language model (one not already trained for long-form reasoning) receives a query combined with a handful of "reasoning examples" in its context window. This nudges it to generate responses that include a step-by-step chain of thought before reaching a final answer.
- Generating many candidates: The model runs the same prompt multiple times, producing a batch of candidate responses. Each includes a reasoning chain and a final answer, and they'll vary in quality.
- Filtering with a reward model: A separate scoring model checks each candidate's final answer for correctness. Only responses that clear a predefined accuracy threshold survive. This is called outcome-based reward filtering, meaning the system judges the work by whether it got the right answer, not by checking every reasoning step manually.
- Training on survivors: The filtered, correct-answer responses become labeled training data. The target model is then trained on these examples to learn the reasoning style associated with correct outcomes.
The process also includes an online training phase, where the model keeps updating itself on new, previously unseen queries. This prevents the model from overfitting (getting too good at memorizing old examples while failing on anything new).
What this means for enterprise AI without massive model budgets
The dominant approach to building a reasoning-capable AI model today relies on knowledge distillation, where a large, expensive frontier model (think GPT-4 or Claude) generates the training examples that teach a smaller model to think carefully. That creates a dependency on whoever runs the big model, plus the cost of generating thousands of high-quality examples.
Salesforce's method, if it holds up in practice, could let companies build capable reasoning agents using only their own existing models and compute. For enterprise software customers who want AI that can work through multi-step business logic without sending data to a third-party frontier model, that's a meaningful practical difference. It also fits Salesforce's broader push to sell AI agents as part of its CRM and platform products.
This is a genuinely interesting technical bet. Bootstrapping reasoning ability from a model that doesn't already have it (and doing so without leaning on a more powerful model) is one of the harder problems in practical AI deployment. Whether the quality of self-generated, outcome-filtered training data can match what you'd get from distilling a frontier model is the real question, and the patent doesn't answer it. But the direction is commercially smart for Salesforce, which needs capable AI that runs on its own infrastructure.
The drawings
10 drawing sheets from US 2026/0220478 A1 · click any drawing to enlarge
Which company should we read for you?
We track 17 companies here. Pro is the same weekly breakdown for any company you choose, delivered privately. Type a name and we'll scope it and send you a quote.
Get one Big Tech patent every Sunday
Plain English, intelligent commentary, no hype. Free.
Editorial commentary on a publicly published patent application. Not legal advice.