@che_shr_cat: 1/ Parameter scale is a brute-force crutch. What if a 35B model could beat a 1,000B model simply by scaling its search …

X AI KOLs Timeline Papers

Summary

Explores whether a 35B parameter model could surpass a 1000B model by scaling its search horizon at test time, using structured process feedback rather than brute-force parameter scaling.

1/ Parameter scale is a brute-force crutch. What if a 35B model could beat a 1,000B model simply by scaling its search horizon at test time? It is not magic. It is structured process feedback. Let us look under the hood. 🧵 https://t.co/wsTyq46MQL
Original Article
View Cached Full Text

Cached at: 07/20/26, 09:29 AM

1/ Parameter scale is a brute-force crutch. What if a 35B model could beat a 1,000B model simply by scaling its search horizon at test time?

It is not magic. It is structured process feedback. Let us look under the hood.

2/ In “Scaling the Horizon, Not the Parameters”, Shanghai AI Lab introduces Agents-A1.

They prove that explicitly training a smaller model on structured execution traces can match or exceed trillion-parameter frontier models on complex tasks.

3/ The framework runs on a Knowledge-Action Graph (KAG).

Instead of treating trajectories as flat text, they formalize steps as verifiable 4-tuples: (Corpus, Action, Observation, Verifier).

Every action is written back to the graph only if it passes verification.

4/ Training uses a 3-stage recipe:

  1. Full-domain SFT
  2. Specialized domain RL teachers (using GRPO)
  3. Multi-Teacher On-Policy Distillation

This solves a massive problem: gradient conflicts when trying to mix tool-use, search, and instruction following.

5/ To merge these teachers, they use Salient Vocabulary Alignment (SVA).

Instead of distilling the entire vocabulary, the student only aligns with the domain-routed teacher on a compact top-k slice of valid tokens.

This keeps the training stable and prevents gradient clash.

6/ The results: Agents-A1 (35B MoE) hits 56.4 on SEAL-0, beating DeepSeek-V4 (55.0) and Kimi-K2.6 (50.5).

On FrontierScience-Olympiad, it scores 79.0, beating DeepSeek-V4-Pro’s 76.0.

We are seeing frontier-class capability from a 35B model.

7/ The catch? Serious latency.

Running an average sequence length of 45K tokens at test time is computationally brutal.

It also lags on MLE-bench (43.9 vs GPT-5.5’s 72.7), proving that long-term goal consistency for complex engineering is still highly challenging.

8/ The takeaway: Stop waiting for bigger models.

We should be focusing on structured process-level RL and test-time search. Unifying specialized teachers through on-policy distillation is a highly viable path for small, local deployments.

9/ Read my full technical breakdown here: https://arxiviq.substack.com/p/scaling-the-horizon-not-the-parameters…

Check out the original paper here: https://arxiv.org/abs/2606.30616

How are you structuring your agent trajectory data? Let’s discuss in the replies.

#MachineLearning

10/ Visual breakdown of the multi-teacher distillation process below. Sometimes a diagram is worth a trillion parameters.

Similar Articles