Three of the last four weeks at #1 on FutureX. In our showcase round, a single 35B model swept the entire top four.
We think the streak shows a promising way toward Discoverative Intelligence by building heavy-duty reasoning capacity: turning information into evidence through verification, integrating that evidence under uncertainty, and reaching a defensible conclusion about something no one yet knows.
What FutureX actually tests
Most AI benchmarks have a quiet flaw: the answer already exists somewhere in the training data. A model can look brilliant simply because it has read the answer key.
FutureX closes that loophole by design. It asks AI systems to predict events before they happen — the outcome of an election, where a stock closes, who wins, whether something ships on time. Questions are collected automatically each week from regularly updated sources; systems must submit predictions before resolution; and outcomes are scored automatically once the resolution date passes. Because the ground truth literally does not exist when the question is asked, data contamination and benchmark leakage are impossible by construction.
FutureX spans high-value domains — politics, finance, business, science and technology, health, and space — and tests an agent's ability to search, plan, reason, and decide under genuine uncertainty in economically meaningful situations.
In other words, it doesn't reward memory. It rewards Discoverative Intelligence that discovers, validates, and reasons about truths that do not yet exist at the time of inquiry.
The core insight: prediction is heavy-duty reasoning
At Apodex, we believe the most important AI systems should tackle the hardest problems. Two properties define what "hardest" means to us:
Heavy-duty reasoning — not just collecting information, but turning information into evidence through verification, integrating that evidence, and making structured decisions under uncertainty.
Questions with no existing answers — problems where the ground truth does not yet exist at the moment of reasoning.
Future prediction sits exactly at the intersection of both. Predicting an event before it resolves demands that a system assess the relevance and reliability of information, turn raw data into evidence, and reason across that evidence to a defensible conclusion — far beyond retrieval from the web. And because the answer cannot be known at inference time, there is no shortcut to fall back on. Every correct prediction must be earned.
That is precisely the loop that defines heavy-duty reasoning: decompose → search → verify → synthesize. The future is just the topic that happens to have no answer page yet. We don't treat prediction as a separate capability we built — we treat it as the hardest possible stress test of a heavy-duty solver toward Discoverative Intelligence: discover the unknown, and reason about it through verification.
Introducing Apodex-FutureFlow
Apodex-FutureFlow is our future-prediction system. It is built on Apodex-1.0, our state-of-the-art deep research model — and every result below was produced using Apodex-1.0-mini (now available as Apodex Deep Research), our open 35B model, as the backbone for all agent steps.
Our goal was never to chase a leaderboard number through trial-and-error engineering. It was to understand why future prediction works — the underlying principles that make it tractable — because we believe those principles illuminate the foundations of Discoverative Intelligence itself.
The baseline: a robust system from minimal architecture
Apodex-1.0 operates as a general-purpose research solver through a standard tool-augmented ReAct loop: at each step it selects and invokes tools — web search, page fetching, file reading, and stateless code execution — iterating until it produces a research report. Apodex-FutureFlow extends this foundation with three deliberate modifications that form the single-trajectory workflow:
Domain-aware planning. Before the main loop begins, the agent calls a planning sub-agent (architecturally identical to itself) whose only job is to analyze the query and produce a structured plan: the precise scope of the question, the relevant information sources, and a high-level reasoning strategy. This matters because prediction queries are often ambiguously worded and draw on non-standard, domain-specific sources — so getting the scope and sources right first is a prerequisite for everything that follows.
ReAct-based research. The main agent then runs the standard tool-augmented loop, conducting multi-step retrieval and synthesis to produce a comprehensive research report.
Structured summarization. A dedicated step processes the full reasoning trajectory and report to extract a structured, verifiable boxed answer — the canonical output of a single trajectory.
On top of this, the baseline adds multi-trajectory aggregation. Each query is run through several independent trajectories, and rather than taking a naive majority vote, the aggregation step reasons jointly over all of them: it treats each trajectory's summary as a structured evidence package, identifies evidence supported consistently across trajectories, flags singletons, surfaces contradictions, and re-weights each piece of evidence by its cross-trajectory reliability before committing to a final answer. It's verification at the level of evidence, not vote-counting.
The payoff: with minimal architectural intervention, this baseline alone outperforms every external system on FutureX — a credible research foundation, not just a strong submission.
One more feature: self-reflection–based verification
The baseline has a subtle failure mode. We observed that it tends to land on a candidate answer relatively early in a trajectory (often around the halfway point) and then slip into a confirmation mode — passively accumulating supporting evidence while ignoring contradictory signals. Within a single trajectory, the search becomes self-reinforcing rather than genuinely exploratory: an echo chamber.
So we introduced a self-reflection mechanism that intervenes at two points in the single-trajectory pipeline:
After producing its initial candidate, the model is required to actively challenge it — searching specifically for disconfirming evidence and proposing up to two alternative candidates based on that adversarial search.
At summarization, the model must output one to three candidate answers, each with a calibrated confidence estimate, rather than committing prematurely to a single answer.
This feature couples tightly with multi-trajectory aggregation. By pursuing multiple candidates within a single trajectory, we trade single-trajectory confidence for breadth of evidence coverage, and the adversarial search forces genuine contradiction-aware reasoning instead of passive confirmation — exactly the heavy-duty reasoning prediction demands. And there's a bonus: when the initial answer survives an active hunt for counter-evidence, the supporting evidence has been implicitly stress-tested, making it more reliable precisely because it wasn't selected under confirmation bias. The net effect is a higher-information-density summary from every trajectory, which in turn gives the aggregation step better input and a more reliable final prediction.
What this means
The leaderboard is telling a simple story about heavy-duty reasoning:
Prediction is earned, not retrieved. The best future prediction system makes predictions through heavy reasoning: decomposing the question, searching for signal, verifying against contradiction, and finally synthesizing evidence into a defensible conclusion. There is no shortcut when the answer doesn't exist yet.
Verification is the moat. A system that turns information into stress-tested evidence, rather than confirming its first guess, makes better calls about what's coming.
And the principles generalize to Discoverative Intelligence. Evidence verification and contradiction-aware reasoning aren't domain-specific tricks; they're fundamental properties of any system that aspires to reason about the unknown.
Three of the last four weeks at #1 isn't the headline we care about most. The headline is why: future prediction is a heavy-duty task that admits no shortcuts, and tackling it makes it our clearest step yet toward Discoverative Intelligence: an AI that truly thinks.
The future is the ultimate open question. We intend to keep answering it.
Apodex-FutureFlow is built on Apodex-1.0 — read about the model at https://apodex.com/blog/apodex-1.0, and follow the live leaderboard at https://futurex.live.