Yuvraj Verma | Search arXiv

arXiv2026

Try Again, Don't Look Back: Blind Resampling Outperforms Self-Repair in Small Code Models

Self-repair - returning a failed program to the model together with its test output and asking for a correction - is a standard component of code agents, and is almost always evaluated against a baseline that does not retry at all. We argue that this comparison confounds the value of the feedback with the value of the extra attempt. Using a placebo-controlled design on MBPP+ at three model scales (1.5B, 3B, 7B), we compare four matched-budget retry conditions: blind resampling, a content-free failure notice, genuine execution feedback, and feedback augmented with verbal self-reflection. Blind resampling is the strongest condition below 7B, and remains statistically tied with the best condition at 7B, while consuming 2.5-5.5x fewer tokens; conditioning on the model's own failed attempt costs 6.1 points at 1.5B (p=0.006), and the informational content of execution feedback adds nothing measurable over the placebo. We attribute this to anchoring: when shown its previous attempt, a model reproduces a near-identical program in 33-68% of retries, against 2-14% under blind resampling. Two further experiments delimit the effect. Retrieved solutions to other tasks change nothing (bounded to +/-3.5 points), which localizes the harm to self-conditioning rather than context length; and reflection, the only condition that measurably weakens the anchor, remains dominated on cost. Replication rules out two competing explanations: the penalty is unchanged at full precision, and it reproduces on an independent model family. Across six configurations spanning two families and two precisions, its magnitude is predicted by baseline quality alone (r=0.96) - the cost of anchoring is the cost of committing to a bad first attempt.

cs.SE↗

arXiv2026

What Does 99% Accuracy Measure? A Reproducible Audit of Shortcut Learning in a Widely Used Fake News Corpus

Text classifiers trained on the ISOT/Kaggle "Fake and Real News" corpus routinely report accuracy and F1 above 0.98, a level of performance that sits uneasily beside the difficulty of assessing veracity. Using a transparent TF-IDF and linear-classifier pipeline as a measurement instrument, we audit the corpus along three leakage channels and two distribution-shift protocols, releasing all code and derived numbers. First, the benchmark is partly degenerate: a classifier given only the subject metadata field, with the article text discarded, attains F1 = 1.000, since the two classes have disjoint subjects. Second, removing all three leakage channels, metadata, a newswire source tag present in 99.2% of real articles, and 6,251 duplicate documents contaminating 19.4% of a naive test split, lowers F1 by only 1.21 points (0.9935 to 0.9814); the residual signal is diffuse editorial style rather than a few giveaway tokens, since deleting the 1,000 highest-weight unigrams still leaves F1 = 0.926. Third, this style signal does not transfer: under a topic-disjoint protocol, average precision falls from 0.9995 to 0.9475 and deployed F1 from 0.9905 to 0.8067, with a prior-matched analysis confirming a genuine 5.2-point loss of discrimination, while temporal transfer is nearly lossless. A fine-tuned DistilBERT is stronger in-distribution (F1 = 0.9993) but degrades far more under topic shift, losing 12.9 average-precision points against the linear model's 5.2. Transferred to the independent LIAR benchmark, all three models fall to near-chance ranking (ROC-AUC 0.54-0.57), none beating a majority-class baseline. We conclude that within-corpus scores here quantify source and topic separability rather than veracity, that added capacity exploits the shortcut rather than avoiding it, and we recommend metadata-only, small-sample, and topic-disjoint baselines as inexpensive diagnostics for future work.

cs.CL↗