Open-sourced an RL model to give LLM the sales strategies
这条记录涉及编程工具或代码能力更新，适合开发者评估工作流变化和可复用价值。
针对大型语言模型（LLM）在销售场景中过于顺从的问题，研究人员开源了一个强化学习（RL）模型，旨在增强LLM的销售策略。该RL模型通过训练，根据信任、兴趣、预算匹配等客户特征的数值，预测销售行动，如推销或促成交易。模型训练无需大量销售数据集，而是通过模拟数百万个不同环境，以收入作为奖励来优化策略。…
The main problem faced when using these LLMs for sales usage is that they are perfect, smooth, polite, always accepting, and agree with whatever I say, even with strict prompting; things always go for acceptance in the long run. The same for Fable 5 ( handicapped now), Opus 4.8, Gemini 3.1 pro, and the GPT 5 series. Always thought about augmenting these responses with a trained RL policy that understands sales nuances. We dont need large sales dataset to train these model, we dont need a dataset at all!, we just need numbers that represent the sales features or the customer values, like trust, interest, budget fit etc. you give them numbers, then train a PPO model with revenue as reward on mllions of environments with different numbers for each, the idea is to predict action, the actions are like close, pitch, rapport etc. say if trust value increases above a threshold, the interest value should also be increased, if many of these conditions are above a certain threshold, revenue aka reward become larger numbers, else zero. So, without words, we train an RL with just numbers and sales rules. Now this RL has to be bridged with the residual streams of the LLMs, so we can add the hidden features and the action states from the RL to the LLM to augment its final response, so we train a bridge MLP layer using Gemma 4 E4Bs layers, frozen and frozen RL layers, the whole idea is to perfecly bridge the hidden features from the trained LLM to the LLM. During inference, one LLM generates a JSON for the features like trust, interest, and the RL model uses this to create the hidden features, and the action states are injected into the LLM’s residual flow, both use two instances of the same LLM, btw. But we can juse use the json directly from first LLM response and use it in second with second LLM, but it doesnt know the future, it doesnt playes 40 million sales games, the policy makes it more interesting, that is 1024 hidden layers from the RL during inference gives the reason why it made the decision, and 8 action head gives which is the best move to be taken.
TLDR: A trained RL model on 22 customer states like trust, interest, etc., to predict which action to take, like pitch or close, injecting on an open-source LLM residual flow to augment the final response. For the LLM APIs, we don’t need to inject, just a system prompt after the RL output, and augment the final response
Pypi package at: https://pypi.org/project/rl-sales-augment/
GitHub repo at: https://github.com/NandhaKishorM/rl-sales-augment
Build on top of my 1 year back arxiv paper: https://arxiv.org/abs/2510.01237
Now new arxiv submission just submitted. Will share the paper once it's accepted