alibaba-cloud-6baacc3b·2 events·first seen Aliases: Alibaba Cloud
A new arXiv preprint investigates whether independently trained code language models converge on the same internal representations, using a 2x2 factorial design crossing Python/Rust with Qwen2.5-Coder-7B and DeepSeek-Coder-V1-6.7B. The study measures 58 Python and 57 Rust grammatical concepts via concept-circuit extraction across all four conditions. Key findings: which concepts earn dedicated circuits is consistent across models (task-determined), but where and how those circuits are organized differs substantially by model — Qwen processes concepts in late layers (~L17-19), DeepSeek in early layers (L6-7). The work reframes the 'circuit universality' question, concluding universality holds for representational content but not computational organization.
Researchers use learned soft prefixes (opaque continuous vectors) to probe the logical stability of LLMs on syllogistic reasoning benchmarks without modifying model weights. Across Qwen3.6-35B-A3B MoE, Qwen3-8B, and Gemma 4 31B, successful prefixes redirect correct answers at flip rates of 72–90% for Qwen3.6 MoE and 54–56% for Gemma, far exceeding random controls by 37–99 percentage points. The dominant effect is a broad answer-preference bias rather than symbol-level forcing, and the bias generalizes across unseen logical forms and prompt interfaces. Model-specific differences in how this bias manifests suggest substantial variation in logical robustness across architectures.