Large Language Models are increasingly used as zero-shot classifiers in complex reasoning tasks. However, standard constrained decoding suffers from a phenomenon we define as Renormalization Bias. When a model is restricted to a small set of target labels, the standard softmax operation discards the probability mass assigned to semantic synonyms in the original distribution. This loss of information, which we call the Silent Vote, results in artificial overconfidence and poor calibration. We propose Semantic Softmax, an inference-time layer that recovers this lost information by aggregating the scores of the semantic neighborhood surrounding each target label. We evaluate this approach on Qwen-3 and Phi-4-mini models using GoEmotions and Civil Comments datasets. Our results demonstrate consistent improvements across all evaluation metrics: Semantic Softmax substantially reduces Expected Calibration Error (ECE) and Brier Score, while simultaneously enhancing discriminative performance in terms of AUROC and Macro-F1. By accounting for linguistic nuances, our method provides a more calibrated and accurate alternative for zero-shot classification.

Multimodal Large Language Models (MLLMs) have achieved remarkable success in semantic visual reasoning, yet their capacity for fine-grained, low-level perception remains critically under-evaluated. This perceptual fragility limits their reliability in noisy, real-world environments where visual signals are degraded. Furthermore, existing benchmarks often entangle visual perception with language priors, masking these underlying deficits. To address this, we introduce the **FAint numeric Detection Evaluation (FADE)** dataset, a novel evaluation suite designed to probe the limits of zero-shot Optical Character Recognition (OCR) in frontier MLLMs. By embedding synthetic, strictly numerical sequences over cluttered natural backgrounds at varying levels of transparency (𝛼), FADE explicitly disentangles pure visual perception from semantic predictability. We evaluate state-of-the-art models including Gemini 3.0, Claude 4.5 Sonnet, and Gemma 3 against a specialized UNet segmentation baseline. Our results reveal a striking limitation in frontier architectures: while they achieve near-perfect transcription at high visibility, their performance collapses under high transparency. Conversely, the UNet pipeline maintains robust spatial grounding, significantly outperforming generalist models at the lowest visibility thresholds. FADE provides a reproducible dataset to expose and diagnose the perceptual breakage points of modern multimodal systems.

Advanced reasoning typically requires Chain-of-Thought prompting, which is accurate but incurs prohibitive latency and substantial test-time inference costs. The standard alternative, fine-tuning smaller models, often sacrifices interpretability while introducing significant resource and operational overhead. To address these limitations, we introduce Prompt-Level Distillation (PLD). We extract explicit reasoning patterns from a Teacher model and organize them into a structured list of expressive instructions for the Student model’s System Prompt. Evaluated using Gemma-3 4B, PLD improved Macro F1 scores on StereoSet (57% to 90.0%) and Contract-NLI (67% to 83%), while increasing LogiQA accuracy to 70%. Similar results on Mistral Small 3.1 demonstrate cross-architecture generalizability, enabling these compact models to match frontier performance with negligible latency overhead. These expressive instructions render the decision-making process transparent, allowing for full human verification of logic, making this approach ideal for regulated industries such as law, finance, and content moderation, as well as high-volume use cases and edge devices.

We present LegalSim, a modular multi-agent simulation of adversarial legal proceedings that explores how AI systems can exploit procedural weaknesses in codified rules. Plaintiff and defendant agents choose from a constrained action space (for example, discovery requests, motions, meet-and-confer, sanctions) governed by a JSON rules engine, while a stochastic judge model with calibrated grant rates, cost allocations, and sanction tendencies resolves outcomes. We compare four policies: PPO, a contextual bandit with an LLM, a direct LLM policy, and a hand-crafted heuristic; Instead of optimizing binary case outcomes, agents are trained and evaluated using effective win rate and a composite exploit score that combines opponent-cost inflation, calendar pressure, settlement pressure at low merit, and a rule-compliance margin. Across configurable regimes (e.g., bankruptcy stays, inter partes review, tax procedures) and heterogeneous judges, we observe emergent “exploit chains”, such as cost-inflating discovery sequences and calendar-pressure tactics that remain procedurally valid yet systemically harmful. Evaluation via cross-play and Bradley-Terry ratings shows, PPO wins more often, the bandit is the most consistently competitive across opponents, the LLM trails them, and the heuristic is weakest. The results are stable in judge settings, and the simulation reveals emergent exploit chains, motivating red-teaming of legal rule systems in addition to model-level testing.