SearchJev separates search decisions from autoregressive language generation

2026-10-06

SearchJev uses a calibrated System-1 model to handle search agent decisions, improving answer accuracy from 45% to 54% on BrowseComp-Plus while reducing active search time.

On 04 Oct 2026, Huawei researchers released SearchJev to address the latency and confidence issues that happen when generative models handle short, repetitive search decisions. I work in AI and I build agentic AI systems, so the obsession with using a massive LLM for every single "yes/no" relevance check is a point of friction I find tedious. Most agents treat every decision as a prose generation task.

SearchJev changes this by separating the decision from the reasoning. It scores legal options directly based on a decision schema rather than generating tokens autoregressively. To make this work, the authors used Soft-Label Learning for Calibrated Decisions (SLCD). This allows the model to learn probabilities from uncertain supervision and calibrate how confident it actually is about a specific search action.

The results on SearchDecision-Bench show that SearchJev is 5.2-5.3 × faster than same-size Qwen3.5 autoregressive models. It also reduces average expected calibration error by 41-74%. I think the real value here is the dual-system architecture. SearchJev acts as the System-1 for fast decisions, while a System-2 model handles the heavy lifting like planning, query generation, and final answer composition.

When tested on BrowseComp-Plus, these dual-system agents achieved a 3.7-4.7 × speedup in active search time. Accuracy climbed from 45% to up to 54%. This is a concrete win for efficiency, though I don't know how this calibration holds up across domains outside the SearchDecision-Bench. The benchmark itself unifies six types of search decisions, which provides a decent baseline, but real-world search noise is often messier than a benchmark.

I'd argue that the industry is too slow to admit that not every agentic step requires a full generative pass. Using a calibrated scorer to delegate uncertain judgments to a larger model is a more honest way to build a system. It stops the agent from hallucinating confidence in a search decision just because the language model is trained to be helpful and assertive.

I built brendon.bot to surface this kind of architectural shift. Moving away from the "one model does everything" approach is the only way to get latency down to a usable level for active search. We can stop pretending that autoregressive generation is the optimal way to select a search filter or decide if a page is relevant.