How we 'hack' HR-Tech: a discussion with the author of CALM from Tencent AI
We discussed the applicability of the Energy Score method directly with Chenze Shao, a researcher at WeChat AI (Tencent) and the author of the CALM architecture. We explain how a multi-level pipeline with Energy Gating solves the 'semantic collapse' problem and saves budget on LLM computations.
Authors: slavb18
While everyone else is simply integrating ChatGPT with their resume databases, we at Iconicompany are delving into fundamental AI to solve the 'vector fragility' problem.
Recently, we discussed the applicability of the Energy Score method directly with Chenze Shao - a researcher from WeChat AI (Tencent) and the author of the groundbreaking CALM (Continuous Autoregressive Language Models) architecture.
What's the problem with current search methods?
Standard vector search (Cosine Similarity) often suffers from 'semantic collapse'. The system sees similar words, yields a high score - but when an expensive LLM steps in for final verification, it turns out the candidate isn't suitable at all.
This leads to a huge number of 'idle' and costly model runs.
Our solution: a multi-level pipeline with Energy Gating
We are building a multi-level matching pipeline where Energy Score (based on Brier Score) acts as an intelligent 'cutoff filter' (Inference Gate):
1️⃣ Coarse Retrieval - fast search using standard embeddings.
2️⃣ Energy Gating - our original calibration layer. We train it on a proprietary dataset of 500,000 labeled 'job-specialist' pairs. The model predicts the probability of successful LLM verification. If the matching 'energy' is low, we 'short-circuit' the pipeline and don't spend money on the LLM.
3️⃣ Fine Ranking - precise neural network verification only for the best of the best.
This approach allows for a significant increase in accuracy in complex domains - for example, in engineering recruitment.
Why is this important?
This is Agentic AI in action: the system optimizes its own resources to deliver a verified team to the client within 48 hours, without burning budget on useless computations.
The discussion with Chenze Shao confirmed that this direction is promising: Energy Score indeed provides more stable signals where cosine similarity 'floats' due to noise in skill formulations.
🔗 Discussion details: issue #19
Want to learn more about our AI-powered recruitment approach? Write to us at @iconicompany.
📚 Read also
- From Cosine Similarity to the "Energy" of Meanings: How Tencents CALM Research is Changing the Game in AI Matching
- AI Experience: How to Stop Competing with Thousands of Candidates
- AI is Not About Prompts
- How We Reimagined Developer Assessment: From Resumes to Voice AI Interviews
- HR Copilot - A Dead End. Here's Why