Synthetic Hallucinations, Real Gains: Hard Negatives from Frontier Models for FIM Hallucination Mitigation

Technical Summary: Synthetic Hallucinations, Real Gains

Problem Statement
Small, open-source code models (3–7B parameters) used for IDE autocomplete frequently suffer from Fill-in-the-Middle (FIM) hallucinations. These are completions that are syntactically natural and locally plausible but violate the project's true state, such as calling non-existent methods, using undefined variables, or importing fictitious packages. Existing mitigation strategies face significant limitations:

1. Execution-based filtering: Requires per-language sandboxes and unit tests, which are unavailable for mid-keystroke edits or files that do not yet parse.
2. Preference optimization (DPO/ORPO): Requires large, human-annotated preference corpora or per-rollout reward signals, which are difficult to generate at the granularity of specific FIM holes.

Methodology
The authors propose an execution-free pipeline that leverages frontier code models to synthesize hard negatives (plausible-but-wrong completions) to train smaller models. The core insight is to invert the standard synthetic data recipe: instead of generating correct code and verifying it, the pipeline generates wrong code and contrasts it against the existing "golden" (ground-truth) developer edit.

The pipeline operates in four stages:

1. Data Sampling: The system scrapes a permissively licensed, multilingual corpus from public GitHub (covering C#, Go, Java, JavaScript, PHP, Python, Ruby, and Rust) to extract ~400k FIM contexts (prefix, golden completion, suffix).
2. Hard Negative Synthesis: A panel of three frontier code generators (GPT-5.2-Codex, GPT-5.4, GPT-5.5) produces one hallucinated completion per context for four specific hallucination types defined by the Delulu taxonomy:
   • Method: Replacing a method name with an invented but plausible one.
   • Parameter: Injecting fake keyword or positional arguments.
   • Undefined Variable: Replacing an identifier with one not defined in scope.
   • Import: Replacing a real import with a fictitious package or symbol.
3. Curation: The system creates a paired dataset (chosen/golden vs. rejected/hallucinated). It enforces quotas per language and type to ensure balance and excludes data overlapping with evaluation benchmarks.
4. Training: The authors apply Supervised Fine-Tuning (SFT) on the "chosen" (golden) side of the data. Crucially, the training targets end exactly at the FIM hole boundary, teaching the model to stop generation correctly. The rejected (hallucinated) side is used for contrastive comparisons (DPO/ORPO) in ablation studies but is not strictly required for the SFT recipe.

A secondary mechanism, the "fool rate," is introduced as a difficulty signal. A panel of LLM judges scores hallucinations based on how often they are mistaken for correct code. This allows for filtering the training set to focus on harder negatives, though the main proof-of-concept does not rely on this filtering.

Key Contributions

• Execution-Free Pipeline: A method to generate paired (chosen, rejected) FIM data without sandboxes, unit tests, or file-parsing requirements, relying solely on the contrast between a ground-truth edit and a frontier-generated hallucination.
• Open-Source Reproducibility: The authors release the full pipeline code (generation prompts, judging harness, curation, and fine-tuning recipes), enabling reproduction on any permissively licensed corpus.
• Base-Model Agnosticism: The recipe is validated across four 7B/3B base families (Qwen2.5-Coder, StarCoder2, CodeLlama), demonstrating that the hallucination-correction signal is encoded in the data rather than specific model architectures.
• Stopping Behavior: Evidence that the recipe teaches models to stop generation at the correct boundary, a critical failure mode in FIM tasks.

Results
Fine-tuning Qwen2.5-Coder-7B-Instruct on a curated 100k-row subset yielded significant improvements:

• Delulu Benchmark: Exact Match (EM) increased by +18.8 points (from ~42.8% to 61.4%) across all seven languages and all four hallucination types. Edit Similarity (ES) improved by +0.22.
• General FIM: The recipe improved performance on HumanEval-Infilling and SAFIM benchmarks across all splits, including control-flow and API completion.
• 3B Model: The same recipe on Qwen2.5-Coder-3B improved Delulu EM by +12.8 points but showed a trade-off, regressing on specific HumanEval-Infilling splits (multi-line) due to capacity constraints.
• SFT vs. Contrastive: On the same paired data, SFT matched or exceeded ORPO performance and significantly outperformed DPO. DPO failed to internalize the stopping behavior, leading to over-generation, whereas SFT successfully learned the correct token sequence and stopping point.
• Ablations:
  • Size: Performance saturates around 50k rows; 100k offers diminishing returns.
  • Language: Cross-lingual transfer requires a minimum of five diverse languages to ignite; beyond that, budget is better spent on scale than diversity.
  • Difficulty: The most valuable supervisory signal comes from the "disputed middle" of the fool-rate distribution (hallucinations that confuse judges but are not unanimous), rather than the easiest or hardest extremes.

Significance and Claims
The paper claims that synthetic hard negatives generated by frontier models can serve as an effective supervised fine-tuning signal to reduce hallucinations in smaller models without the overhead of execution environments. The authors emphasize that:

1. Quality over Quantity: A curated, difficult subset of data (50k rows) is more effective than larger, uncurated sets.
2. SFT Sufficiency: For stopping-sensitive FIM tasks, a high-quality curated SFT corpus is a safer and more effective default than explicit contrastive training (DPO/ORPO), which can struggle with generation length control.
3. Generalization: The approach improves general FIM capabilities (not just hallucination reduction) when applied to 7B models, though 3B models require careful curation to avoid catastrophic forgetting on specific task types.

The authors remain modest about limitations, noting that the training data consists of synthetic hallucinations rather than real-world model errors, and that the 3B model exhibits a capacity-dependent trade-off between hallucination reduction and general FIM accuracy. They do not claim the method solves all hallucination types (e.g., complex logic errors) but demonstrate its efficacy on identifier-level hallucinations.