Inside Arena's factuality signal: the methodology deep dive

What Arena is adding

• Arena leaderboards currently rank models mainly by human preference.
• Arena is extending this with a new factuality signal so that models that produce incorrect factual claims are penalized, and models that are calibrated about what they claim are rewarded.
• The approach is designed to be complementary: human preference and factuality provide different information about response quality.

---

Why factuality is needed (problem setup)

• Manual fact-checking of model outputs is “very time-consuming” and effectively infeasible at scale.
• Human preference does not reliably reflect factual correctness:
  • Some tasks (e.g., creative writing) may contain no checkable claims, so factuality shouldn’t matter.
  • Other responses may contain many claims; humans may miss subtle inaccuracies, and length/format can bias preference votes.

---

Key product behavior / policy goals

• Don’t penalize abstaining: if a model makes no claims, it should not lose on factuality.
• A model that “doesn’t know” (or avoids uncertain statements) should be treated fairly on factuality—even if the user subjectively dislikes the answer.
• The methodology specifically targets models that shoot off lots of claims where many are likely wrong.

---

How the leaderboard methodology is constructed (overall)

Arena uses pairwise comparisons to learn model “strength” scores, using:

• Human preference battles: sample two models, ask which response wins, producing label (Y_H).
• Factuality battles: sample a subset of battles, fact-check responses, producing label (Y_F).

The final leaderboard score is learned by optimizing a combined objective:

• A convex combination of the human-preference loss and the factuality loss.
• The weights sum to 1, and changing the weight changes rankings (e.g., “if we weight factuality 25%, what happens?”).

---

Human preference part (existing method, recap)

• Training data format: (A, B, Y_H)
  • (A) and (B) are anonymous model outputs
  • (Y_H \in {0, 0.5, 1}) (A wins, tie, B wins)
• Learns scores via Bradley–Terry loss.
• Learned scores are rescaled to an ELO-like scale for display.

---

Factuality part (new method, core details)

Fact-checking labels

For factuality battles, Arena:

• Subsamples some fraction of platform battles
• Fact-checks both responses individually
• Produces a factuality preference label (Y_F \in {0,1}) (no-tie is not mentioned in the excerpt, but the scoring setup can yield comparisons)

Claim-level verifier → probability of truth

For each battle side:

• Extract a set of claims: claims made by model A vs. claims made by model B.
• Run a verifier on each claim to output a 0/1 score, described as a probability that the claim is true.
• Define a side-level factuality metric:
  • Average factuality probability across all claims on that side.
  • If a response has many correct claims, it approaches 1; if half appear wrong, it trends toward 0.5.

Converting side factuality into the battle label (Y_F)

(Y_F) is computed as a sigmoid of the difference between:

• average probability of correctness for side A
• average probability of correctness for the other side

It also includes a temperature term for normalization so the outputs look “clean” on the leaderboard.

---

Abstention handling (how it avoids double-penalizing)

• If a model makes zero claims:
  • It should not lose factuality by default.
• Outcome logic described:
  • both abstain (no claims): effectively a tie/no factuality advantage
  • one abstains and the other makes (true) claims: abstainer can win or tie (as described)
  • abstainer vs. opponent making false claims: abstainer wins (so abstention is not punished as “wrong information”)
• Overall goal: reward calibration (only claim what the model is reasonably confident is true) and penalize overclaiming.

---

Statistical methodology: confidence intervals

• After defining the factuality-based objective, Arena:
  • Treats the learned parameter (\hat{\theta}) as an M-estimator
  • Uses asymptotic normality (central limit theorem / M-estimation theory)
• They can compute:
  • a closed-form covariance (noted to have the “sandwich form”: Hessian-inverse × gradient-covariance × Hessian-inverse)
  • a plug-in estimate (\hat{\sigma})
  • resulting confidence intervals around the estimator
• Discussion notes this is asymptotic, not exactly exact in finite samples.

---

Key analytical takeaways mentioned

• Optimizing only human preference vs. only factuality yields very different rankings, supporting the claim that the signals are complementary.
• The system includes a requirement that factuality vote/label counts don’t become negligible relative to human votes over time (otherwise certain asymptotic results degrade); i.e., balanced sampling is important for stability.

---

Main speakers / sources (as stated in the subtitles)

• Tiana (main explainer of what factuality is and why)
• Evan (mentioned as providing details on how factuality labels are obtained)
• Other speaker(s) (co-presenter/host during the math and statistical discussion; exact identity not named in the subtitles)