5arXiv cs.CL (Computation and Language)·
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.
Related guides (4)
Related events (8)
Semantic vs. Surface Noise in LLM Agents: 68-Cell Measurement Study with Held-Out Validation
This paper documents an empirical phenomenon across 10 LLMs from 7 architecture families: meaning-bearing perturbations (paraphrase, synonym substitution) cause final-answer inconsistency ~19.69 percentage points more often than presentation-level perturbations (formatting, reordering) of comparable severity, across GSM8K, MATH, and HotpotQA benchmarks. The effect is validated on a held-out 11th model (qwen2.5-14B-Instruct) with 1,800 trajectories. Trace-level analysis supports a 'stealth-divergence' picture where semantic perturbations preserve the first action but induce divergence in intermediate reasoning steps, while two prior mechanism claims are explicitly retracted. The study is notable for its honest reporting of stress-test failures and pre-registered replication.
Study finds thinking mode in LRMs shifts instruction-following errors by constraint type rather than uniformly degrading performance
A new arXiv paper investigates how enabling built-in chain-of-thought reasoning ('Thinking ON/OFF') in Qwen3 and Hunyuan models affects instruction following on IFEval. Aggregate pass-rate changes are small but 10-20% of prompts switch outcomes, with 'Planning' constraints (global counting, structure) improving under thinking while 'Precision' constraints (exact local form) consistently worsen. Activation patching and trace-relevance analyses reveal an execution gap: thinking traces engage with Planning constraints but fail to translate that engagement into compliance, while Precision failures are more mechanistically recoverable. The findings have practical implications for when to enable reasoning modes in instruction-following applications.
Statistical Re-evaluation of GSM-Symbolic Finds Benchmark Confounds and Overstated Reasoning Conclusions
A re-evaluation of the GSM-Symbolic benchmark (Mirzadeh et al., 2025) challenges its conclusion that LLMs lack genuine reasoning capabilities. Using Generalised Linear Mixed Models on 20 open-weight models, the authors find only half show statistically significant performance drops, and identify a previously unacknowledged distributional shift toward larger integers in GSM-Symbolic relative to GSM8K that accounts for significance in roughly half the remaining cases. After controlling for this confound, model-specific failure profiles emerge—including variable binding fragility, arithmetic limitations, and dual-task interference—suggesting the original blanket claims about LLM reasoning were both statistically premature and mechanistically misleading.
Prefix Utility Model (PUM) trains process reward models on outcome-grounded prefix gain rather than step correctness
A new arXiv preprint proposes replacing local step-correctness signals in process reward models with 'prefix gain' — the improvement in solve-rate induced by conditioning a student model on a given reasoning prefix. The authors train a Prefix Utility Model (PUM) using a pairwise ranking objective and evaluate it across Best-of-N selection, beam search, and RL on mathematical reasoning tasks. PUM shows particular strength when candidate pools are large, search budgets are high, or rule-based rewards are sparse. Code, data, and models are released publicly.
Systematic study reveals effectiveness-fluency trade-offs in LLM conditioning methods
A new arXiv paper systematically evaluates a range of LLM conditioning methods across both concept injection and removal scenarios, finding that efficient steering methods often degrade fluency significantly. A key finding is that activation steering is substantially less effective on instruction-tuned models than on base models, a previously overlooked interaction. Simple prompting and supervised fine-tuning work for concept injection but not removal, and cheap textual metrics are found to correlate well with expensive LLM-as-judge evaluations.
Benchmark study reveals how linguistic framing of user beliefs shifts LLM context-following behavior
A new arXiv paper introduces a typology of 17 linguistically motivated expression-of-belief (EoB) types—spanning form, evidentiality, epistemic stance, and tone—to evaluate how phrasing affects whether LLMs defer to user-stated beliefs or their own prior knowledge. The authors benchmark 16 LLMs across Llama 3, Qwen3, and Gemma3 families at scales from 1B to 30B parameters, finding that larger and instruction-tuned models are systematically less context-following than smaller or base models. Specific linguistic framings (e.g., presuppositions, certainty markers) are identified as statistically more persuasive, with implications for prompt robustness and sycophancy research.
Systematic evaluation of LLM prompt sensitivity in healthcare settings reveals safety risks
Researchers conduct a sensitivity analysis of both general-purpose and medical-specific LLMs using the MedMCQA benchmark, testing robustness to lexical and syntactic prompt perturbations. The study finds that even minor phrasing changes can alter clinical advice, and adversarial prompts can produce dangerous outputs such as incorrect dosages or omitted critical findings. Both general-purpose models (GPT-3.5, Llama 3) and domain-specific models (ClinicalBERT, BioLlama3, BioBERT) exhibit this fragility, with syntactic reordering and misleading contextual cues proving more destabilizing than simple paraphrasing.
LOCOS: Logit-Contribution Scoring identifies non-literal retrieval heads in long-context LLMs
A new arXiv preprint introduces Logit-Contribution Scoring (LOCOS), a method for identifying attention heads responsible for non-literal retrieval in long-context LLMs — cases where models synthesize answers from meaning rather than copying tokens verbatim. Existing detectors fail at this task because they rely on a literal-copy criterion that misses the output-value (OV) circuit mechanism. Evaluated across Qwen3, Gemma-3, and OLMo-3.1, LOCOS outperforms prior attention-based detectors on the NoLiMa benchmark, with ablation of 50 heads on Qwen3-8B collapsing ROUGE-L from 0.401 to 0.000 while the best baseline retains 0.292. The identified heads are retrieval-specific, leaving parametric recall and arithmetic reasoning unaffected.