arXiv subjects

Esther Xin

Publications and source records attributed to Esther Xin.

5 recordsLinked to original sources

arXiv2026

Does Chain-of-Thought Reliance Follow a Scaling Law?Larger Models Follow Their Steps More, Independent of Post-Training

Monitoring a model's chain of thought (CoT) is useful for safety only if the final answer depends on the written steps. Prior measurements use natural tasks where true reasoning is unknown. We measure it against an exact ground truth on a synthetic box-swap task where every intermediate state is known. We replace one step inside a model's own chain with a wrong but valid state and check whether the answer follows the edit. We evaluate 25 open-weight models--Qwen3 (0.6B-14B, base and post-trained), Gemma 3 (1B-12B), Llama 3 (1B-8B), and 8 OLMo 2 checkpoints--on 7,200 items each. We find five results: (1) Reliance rises with size and saturates. In Qwen3, answer reliance rises from 16-18% at 0.6B to 97-100% at 14B; a logistic in log N fits well (R^2 = 0.89), and Gemma 3 and Llama 3 follow this trend. (2) Capability is the main driver. Across all models, reliance and step accuracy correlate at r = 0.99; equally capable models are equally reliant, regardless of family. (3) Size still matters at matched correctness. On problems solved correctly before the edit, reliance increases with size among capable models (rho = 0.80, p = 0.002). (4) Post-training does not change reliance. Base and post-trained Qwen3 differ by at most 0.03 from 4B upward; a large gap at 1.7B matches a large capability gap. (5) No model up to 14B solves the task internally. Without written steps, every capable model is at chance once a problem has four or more swaps. OLMo 2 checkpoints never become capable, making the training-token axis a limitation. On tasks requiring step-by-step computation, scale makes the chain of thought more load-bearing, supporting CoT monitoring in this regime.

cs.CL↗

arXiv2026

What Should We Remove? Evaluating the Reliability and Costs of RLVR Data Decontamination

Reinforcement learning with verifiable rewards (RLVR) math corpora are routinely screened for benchmark overlap, but the reliability and data-loss costs of cleaning policies remain underexamined. We audit sixteen open RLVR math corpora comprising 1,673,514 corpus-specific training items against six benchmarks containing 1,949 questions. We evaluate eight policies based on exact, lexical, and semantic matching using 594 independently and blindly dual-reviewed training-benchmark pairs. Calibration and held-out sets have disjoint benchmark questions and training-example families. A separate review of all 83 candidate matches for 80 sampled training items measures item-level false removal. At cosine threshold 0.80, 52 of 59 sampled removals (88%) were confirmed false removals; this is not a demonstrated population rate. No semantic threshold in our preregistered grid (0.80-0.98) meets the 90% conservative-precision target on calibration. The existing judge-adjudicated policy achieves 98.1% resolved precision and 98.2% candidate-frame recall on held-out corpora. It has zero confirmed false removals among six sampled removals, with one unresolved case; this does not establish zero risk. Within the evaluated candidate frame, this policy offers a more favorable precision-recall balance than the semantic-threshold alternatives tested. Recall does not measure contamination outside the candidate frame. We release reviewed labels, per-policy decisions, and coverage accounting to support evaluation of cleaning reliability and cost rather than overlap counts alone.

cs.CL↗

arXiv2026

When a Data Artifact Isn't a Shortcut: Causal Auditing of Synthetic RLVR Corpora

Several recent pipelines build RLVR training data by masking a span of real corpus text and asking a language model to invent plausible wrong answers around it. The correct option is therefore genuine human prose; every distractor is synthetic. Correctness and provenance become entangled, and a policy could in principle learn the second instead of the first. We audit that possibility in GooseReason-0.7M. First we ask whether the asymmetry is visible at all: a classifier reading only five surface statistics (never the meaning) reaches AUROC 0.562 over 315,499 options, barely above chance. The aggregate hides something, though. Code sits at 0.416, below chance, and manual inspection explains why: code distractors turn out to be single-operator mutations of the gold answer rather than freely written alternatives, so the two classes are nearly identical by construction. Detecting a signal is not the same as showing a model uses it, so we then run an intervention. We build a paraphrase-matched control corpus, hold training-set size identical across arms, and train two policies under one fixed budget. The exploitation gap does not favour the unmodified-data arm: 0.021 against 0.027 for the control. Under our budget, in other words, a detectable artifact went unexploited. We think that dissociation, along with the domain-specific construction finding, is worth knowing for anyone curating corpora of this kind, and we release the audit as a mostly CPU-only protocol.

cs.CL↗

arXiv2026

Are Verifier Errors Independent Within a GRPO Group? Evidence from Qwen2.5 Rollouts

Group-based reinforcement learning with verifiable rewards (RLVR) scoresmultiple completions per prompt using automatic verifiers. Analysesbased on independent verifier errors may overlook dependence associatedwith shared answer formats. We investigate this dependence in24,998 groups of eight completions generated by Qwen2.5-1.5B onMATH, GSM8K, and DeepMath-103K. We estimate a pooled within-groupverifier-error correlation of 0.530 (95% confidence interval:0.500--0.560). Under an exchangeable-error model, this correspondsto a design-effect-adjusted effective sample size of 1.70 for aneight-completion group. Dependence varies substantially across answerforms: fractions, radicals, symbolic expressions, and intervals exhibitstronger clustering than unit annotations and percent signs. Replayinggroup-relative advantages across four rule-based verifier configurationsidentifies at least one advantage-sign disagreement in up to 0.83% ofgroups. Because a group is repeated sampling for one prompt, thiswithin-group clustering may reflect shared prompt difficulty as well asshared answer form, and we do not attempt to separate the two here.Unlike studies of correlated judgments across multiple evaluators, ouranalysis examines dependence across completions scored by the sameverifier. These findings motivate prompt- and answer-form-aware analysesof verifier noise rather than characterizations based solely onaggregate error rates.

cs.LG↗

arXiv2026

Where the Verifier Fails: A Category-Level Audit of Reward Signals in RLVR

Reinforcement learning with verifiable rewards (RLVR) and standard benchmark evaluation both rely on an automatic verifier that turns a free text answer into a binary reward. Prior work reports that one evaluation harness accepts only about 94% of its own ground truth answers, blaming LaTeX parsing. That is an aggregate: it does not say which answer forms consume the error budget. We supply the decomposition. We apply metamorphic testing to the verifier rather than the model, generating certified equivalent answer variants, that is, rewrites that preserve mathematical meaning by construction, so that any rejection is a provable false negative needing no human adjudication. We then measure rejection per answer category across four widely used verifiers over 307,420 verdicts. We find three things. (1) Self validation ranges from 53.8% to 95.2% on identical inputs, a spread of 41.3 points. The published figure describes one implementation, not the task; two configurations of the same library disagree on 49.9% of pairs. (2) The residual is not spread across parsing categories but concentrated in whitespace and punctuation, which account for 93.0% of in contract failures for the default LaTeX configuration. A trailing period or newline dominates the budget. (3) Separating rejection from execution failure shows that verifiers with similar aggregate error fail for opposite reasons, and that a reference numeric cascade accepts off by one wrong answers as a step function of magnitude, from 0% below 10^4 to 100% at or above, because its relative tolerance is scale invariant.

cs.CL↗