5arXiv cs.LG (Machine Learning)·
PoPE: Placebo-controlled evaluation of error-conditioned self-repair in frozen small code LLMs
A preregistered arXiv preprint introduces PoPE (Popperian Placebo-controlled Evaluation), a methodology for rigorously measuring whether execution error feedback actually improves code generation in frozen small LLMs (0.5–1.5B parameters). The study uses channel-specific placebos—ablating error content while preserving scaffold form—to test both prompt-based and adapter-based self-repair. Results across both channels failed to confirm content-attributable superiority over placebos or baselines, suggesting that apparent self-repair gains in prior work may reflect form rather than semantic error content. The paper argues that writing oracle-derived representations back into generation state replaces testing with conditioning.
Related guides (2)
Related events (8)
Pre-registered study finds Popperian code-generation prompt skills add no benefit beyond structural scaffolding
A pre-registered two-tier ablation study tests whether 'Popperian falsificationist' prompt skills improve LLM code generation through their procedural content or merely through structural scaffolding. Using Claude Sonnet 4.6 and Qwen2.5-Coder-0.5B with execution-based evaluation (HumanEval+ unit tests) rather than LLM-as-judge, the authors find that on the small model, structured prompts lift correctness by 20-22 points but the full Popperian skill shows no separable benefit over a labels-only scaffold. The paper contributes a calibrated negative result and a reusable disambiguation protocol for evaluating prompt-skill families, while also documenting that LLM self-judges at 0.5B scale perform no better than random selection.
Post-hoc falsification operators for frozen small code models fail to beat Best-of-N in leakage-free evaluation
A measurement study evaluates 26 post-hoc operators (selection, verification, repair, elimination, portfolios) applied to frozen small code models (≤1.5B parameters) against a Best-of-N baseline under a strict leakage-free, matched-compute protocol. None of the semantic operators improves held-out accuracy over BoN, with the failure traced to three structural mechanisms: a coverage wall, a capability scissors, and a near-empty consensus trap. Two non-semantic operators do provide value: an expression-layer recovery method (M1) lifts DeepSeek-Coder-1.3B by +12 tasks on HumanEval+ (p=2.4e-4), and an adaptive consensus early-stop saves ~19% compute with no accuracy harm. The paper's core lesson is that harness quality and coverage measurement should precede investment in semantic post-hoc reasoning.
LLMs judge worse than they generate: empirical challenge to self-evaluation pipeline assumptions
A new arXiv preprint tests the implicit assumption that LLM evaluation is easier than generation, using a controlled in-context QA setup across four benchmarks (SQuAD 2.0, DROP, HotpotQA, MuSiQue) and two models. Results show generation accuracy exceeds self-evaluation accuracy on three of four benchmarks, with attention analysis revealing that evaluation attends to context 3–5x less than generation does. LoRA fine-tuning experiments confirm the asymmetry is not a training artifact, with cross-task interference observed in both directions. The findings directly challenge assumptions underlying LLM-as-a-Judge and self-evaluation pipelines widely used in RLHF and agentic systems.
Study: LLM-Derived Error Highlights and APE Suggestions in MT Post-Editing
Researchers conducted a controlled study with professional En-Nl translators comparing post-editing (PE) workflows augmented with LLM-derived error highlights and automatic post-editing (APE) correction suggestions against regular PE and QE-derived highlights. No condition produced measurable productivity or quality gains over standard PE. However, APE-derived highlights were preferred over QE-derived highlights, and correction suggestions improved subjective user experience.
Generative Compilation: Real-time compiler feedback during LLM code generation via sealors
Researchers introduce generative compilation, a technique that provides compiler feedback on partial programs during autoregressive LLM decoding rather than only after generation completes. The core mechanism is a 'sealor'—a syntax-guided transformation that converts partial programs into complete ones that standard compilers can analyze, formally verified in Lean. Evaluated on repository-level Rust coding tasks across frontier black-box and open-weight models, the approach reduces non-compiling outputs and improves functional correctness by catching errors early and preventing error cascades. The method works without white-box model access, distinguishing it from constrained decoding approaches.
Self-improving VLMs can silently regress when verifier quality is task-mismatched
A new arXiv paper demonstrates that verifier-driven self-DPO, a common recipe for self-improving visual-language models, can silently degrade student model performance when the verifier's task-rubric accuracy is insufficient for the target task. Experiments on Qwen-3-VL-2B and Qwen-2.5-VL-3B across MathVista, MMMU, and BLINK show regressions of 3.4–10.9 percentage points below frozen baselines, with the counterintuitive finding that more accurate-but-still-wrong verifiers cause larger regressions than near-random ones. The authors provide a mechanistic explanation via a variance theorem for progress-gated replay and offer operational guidance: measure target-task rubric accuracy before running any verifier-driven loop and rank verifiers by task-specific quality rather than parameter count.
PowerCodeBench: Knowledge Boundary Probing and Intervention for LLM-Based Power System Code Generation
This paper introduces PowerCodeBench, an execution-validated benchmark for evaluating LLMs on power-system simulation code generation using the pandapower library. The authors identify that failures are dominated by API-knowledge boundary errors (hallucinated function names, misused parameters) rather than reasoning failures, and propose a boundary-aware intervention combining API demand estimation with targeted documentation injection. Evaluated across ten open-weight models (1.5B–480B) and four commercial APIs on 2,000 tasks, the intervention yields 32–56 accuracy point improvements while using only 41% of baseline prompt-token cost. Open-weight models in the 70B–120B range match commercial mid-tier accuracy, with Llama-3.1-405B and Qwen3-Coder-480B leading.
Study finds no detectable self-preference bias when LLMs revise their own instruction-following drafts
A new arXiv preprint tests whether LLMs resist valid corrections to their own writing by using IFEval's deterministic verifier to establish ground-truth correctness, bypassing model-as-judge subjectivity. Across four mid-tier model families and 85 author-versus-fresh comparisons, no statistically significant self-preference bias was detected (gap -5.1 pp, 95% CI [-12.9, +2.7]). A qualitative finding shows that when authors do reject verified-good fixes, 97% of stated reasons are substantive flaw-catching rather than preference. The result challenges the assumption that documented self-preference in judging tasks extends to self-revision contexts.