arXiv Machine Learning
Jun 15
Code Correctness Signals in LLM Hidden States: Pre-Generation Probing and Repair Geometry
arXiv:2606. 14530v1 Announce Type: new Abstract: Large language models encode rich information in their hidden states.
Code Correctness Is Linearly Decodable from LLM Hidden States Before Generation
arXiv:2606. 14530v3 Announce Type: replace Abstract: Large language models encode rich information in their hidden states.
Related stories
arXiv Computation and Language
Aug 31
Why Didn't It Check? Unsupported Final Claims and Their Repair in Two Tool-Equipped Language Models
The study investigates how language models equipped with tools can still produce unsupported final claims, even when a single tool call could resolve the uncertainty. It defines two metrics—occurrence (how often unsupported claims arise) and conditional repair (how often they are fixed when evidence is provided). Experiments on Qwen3-32B and Gemma 4 show that providing the missing evidence consistently repairs all unsupported claims in the Qwen3-32B setup, while the Gemma 4 model never produced unsupported claims under the tested conditions.
arXiv Computation and Language
Sep 25
Encoded but Not Decoded: Layer-Localized Evidence for a Three-Level Gap in LLM Syntax
The paper introduces a three-level evaluation framework—behavioral deployment, LM-head readout, and probe recoverability—to distinguish whether a language model fails a syntactic test by not encoding structure or by failing to use it. Using a trilingual control-dependency benchmark, the authors find that probe recoverability consistently exceeds LM-head readout, which in turn exceeds behavioral deployment across seven models and three languages, with the largest gap observed in Qwen3-0.6B Instruct. Layer-localized activation patching shows that instruction tuning shifts the decoded layer later, suggesting decoding favors surface shortcuts and that behavioral evaluation understates what models encode while probing alone overstates what they deploy.
arXiv AI
Aug 11
On the Robustness of LLMs' Internal Representation of Code Correctness
arXiv:2608. 08266v1 Announce Type: cross Abstract: Code generated by modern language models often reads naturally.
arXiv Computation and Language
Sep 15
Fabrication After Tool Failure: Tool-Augmented Agents Assert Values Their Tools Did Not Return
arXiv:2609.14758v1 Announce Type: cross Abstract: Tool-augmented language models are evaluated on whether they reach the right answer, not on whether they report honestly when a tool fails to supply...
arXiv AI
Jun 6
Willing but Unable: Separating Refusal from Capability in Code LLMs via Abliteration
arXiv:2606. 05396v1 Announce Type: cross Abstract: Producing a labeled vulnerable code at scale is a recurring obstacle for learning-based vulnerability detection: mined corpora carry substantial label noise, and existing LLM-based augmentation propagates these inaccuracies because it transforms vulnerable seeds rather than synthesising vulnerabilities from a specification.
arXiv AI
Aug 26
Feedback That Backfires: Why Small Language Model Agents Repeat the Call They Just Watched Fail
The study investigates why small language model agents tend to repeat a tool call that just failed. By recording the failed call and its error message in the transcript, the authors measure a negative corrective gain—agents are more likely to repeat the failed action, with a drop of about 1.03 nats per token. The problem is traced to the harness design rather than the model’s understanding of errors, and the authors show that replacing the verbatim call with a runtime-generated description of the failure can reduce this backfiring effect by 76%.
arXiv Machine Learning
Jun 9
Decoy-Calibrated Failure Audits for Language Models
arXiv:2606. 09046v1 Announce Type: new Abstract: Useful audits reveal not only how often a model fails, but also where its failures concentrate.
arXiv AI
Jul 15
Form, Not Content? A Preregistered, Placebo-Controlled Evaluation of Learned Error-Conditioned Self-Repair Through Prompts and Weights in Frozen Small Code Models
arXiv:2607. 12962v1 Announce Type: cross Abstract: Frozen small code LLMs are deployed locally, yet the information guiding a retry after a failed attempt is still measured without placebo controls in the self-repair literature.
arXiv Machine Learning
Sep 18
Measurement Under Selection: Decoy-Calibrated Failure Audits for Language Models
The paper introduces Janus, a method for validating error patterns in language models by comparing error rates across predefined yes/no properties and using shuffled decoy labels to set significance thresholds. Janus requires that a pattern’s error difference surpasses the decoy-derived threshold and is replicated on held‑out data before reporting. Experiments on a controlled code‑finding task confirm several meaningful error patterns, while on MuSiQue and LongBench v2 Janus reports no confirmed patterns for the tested properties, contrasting with standard shuffling tests that sometimes confirm patterns.
arXiv AI
Sep 2
Hints Help But Do They Teach? Evaluating Skills Transfer in Code Generation
The study investigates whether hints that convert failing code generation attempts into passing ones provide new information or simply guide models toward solutions they could already generate. Using Qwen2.5-3B-Instruct and Phi-3.5-mini on HumanEval+ and MBPP+, the authors find that relevant hints rescue a significant portion of failures, yet many of those solutions are also recoverable through ordinary sampling. Mechanistic tests reveal a shared activation direction between relevant and unrelated hints, but adding this direction does not improve overall accuracy, indicating limited task-general transfer.
arXiv Machine Learning
Sep 15
Introspective Uncertainty Estimation for LLM-Based Code Generation
The thesis explores Introspective Uncertainty Estimation (IUE) for large language models (LLMs) in code generation, aiming to determine whether hidden-state representations can indicate functional correctness at both response and line levels. Using LiveCodeBench and BigCodeBench, the study finds that hidden states provide a strong signal for overall correctness, with static single-token probes performing best, while dynamic strategies offer no consistent advantage. Although line-level fault localization is more challenging, a conditional Top‑K ranking approach remains effective, suggesting a two‑stage workflow that first screens responses for risk and then prioritizes line‑level checks.