Entity · paper

Logical Judgments Under Pressure: Diagnosing Syllogistic Stability with Learned Soft Prefixes

paperactiveprovisionallogical-judgments-under-pressure-diagnosing-syllogistic-stability-with-learned-soft-prefixes-480de439·1 events·first seen 46h ago

Aliases: Logical Judgments Under Pressure: Diagnosing Syllogistic Stability with Learned Soft Prefixes

Co-occurring entities

Qwen3.6-35B-A3B soft prefix Google Qwen3-4B Gemma 4 Alibaba Cloud

More like this (12)

From Correctness to Utility: Gain-Based Prefix Evaluation for LLM Reasoning Explicit Fuzzy Logic in the Feed-Forward Layer: Self-Forgetting Quantifiers Discover Legible Grammatical-Licensing Detectors Scaling LLM Reasoning from Minimal Labels: A Semi-Supervised Framework with a Lightweight Verifier Operadic consistency: a label-free signal for compositional reasoning failures in LLMs Judgement-of-Learning latent reasoning Knowledge Knows, Verbalization Tells: Disentangling Latent Directions for Mathematical Solvability in LLMs Long-context Reasoning Benchmarks Causally Evaluating the Learnability of Formal Language Tasks DAIS: Dependency-Aware Intermediate QA Supervision for Complex Reasoning When the Chain of Thought Knows Better: Failure Modes in Multi-Turn Reasoning Models Operads for compositional reasoning in LLMs

Recent events (1)

5arXiv · cs.CL·46h ago·source ↗

Soft prefixes can override correct logical judgments in LLMs, revealing model-specific stability limits

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.

Evaluation and Benchmarking AI Safety Research Qwen3.6-35B-A3B soft prefix Google +4 more