The results show that instruction-tuned 3B models can outperform several evaluated 7B base models, that differences among leading models on MASSIVE are statistically indistinguishable under pairwise McNemar tests, and that widely used benchmarks have become saturated and no longer meaningfully discriminate among current open-weight models.
Abstract
Intent classification is a core component of task-oriented dialogue systems, yet practitioners have limited systematic guidance for selecting deployable open-weight language models under compute, latency, and robustness constraints. We present a systematic zero-shot evaluation of 41 open-weight language models spanning 15 families and the 135M--9B parameter range across eight English single-label intent-classification datasets. A ninth dataset, ATIS, uses five labeled demonstrations and is reported as an auxiliary five-shot result. The evaluation includes standard benchmarks, a large-scale voice-assistant corpus, and production-derived e-commerce datasets. Beyond exact-match accuracy, we analyze confidence calibration, robustness to realistic input perturbations, statistical reliability of model rankings, deployment efficiency, and benchmark saturation. Our results show that instruction-tuned 3B models can outperform several evaluated 7B base models, that differences among leading models on MASSIVE are statistically indistinguishable under pairwise McNemar tests, and that widely used benchmarks such as SNIPS have become saturated and no longer meaningfully discriminate among current open-weight models. Instruction tuning's effect on confidence calibration is inconsistent rather than uniformly harmful. These findings provide practical guidance for selecting and evaluating open-weight language models for intent classification.
This survey provides a comprehensive overview of recent advances in LLM-based evaluation, covering techniques, applications, and challenges across domains, with future directions emphasizing standardized protocols, uncertainty estimation, and human–AI collaboration.
M. Nadăş· Artificial Intelligence Revi...· 0 citations
We benchmark eleven audio classification methods: five task-aware closed-set LLMs (four Gemini models plus open-weight Kimi-Audio-7B-Instruct), four fixed-vocabulary taggers (YAMNet, PANNs, Whisper-AT, and SSLAM), a zero-shot audio-text model (CLAP), and an audio-grounded LLM (BAT). We evaluate them on a closed-set sound-source identification task over 2,242 clips spanning 23 fine-grained classes and 11 categories. Since these methods differ fundamentally in how they receive the task and how outputs are scored, we group them into four evaluation tiers rather than one leaderboard, reporting macro Precision, Recall, F1, and false-negative rate per tier. The best model, Gemini-3.1-Pro-Preview, reaches 85.6 percent category-level F1 and 56.7 percent fine-grained F1. Kimi-Audio is competitive for its size, reaching 67.5 percent category-level F1 and 32.9 percent fine-grained F1, but fails to answer 1.6 percent of samples. SSLAM and CLAP match or exceed the best closed-set model at the category level without seeing the candidate list, but fall behind at the fine-grained level. Analyzing the Gemini models'chain-of-thought across 8,968 responses, we find that response length does not predict accuracy, an apparent"holistic judgment beats detailed analysis"effect is better explained as a difficulty confound, and wrong answers are stated confidently 92 to 100 percent of the time. We report full per-class confusion matrices and metrics for all eleven methods, identify the structural error modes behind most of the accuracy loss between granularities, and give practical guidance for choosing among these method families.
Sajjad Abdoli, Ghassan Al-Sumaidaee, Ahmad El-Shiekh et al.· 1 citation
Competence claims for a language model in a safety-critical domain are credible when measured against a standard the domain already enforces. We evaluate an open-weight 31-billion-parameter multimodal model (Gemma 4 31B-IT) on the U.S. Nuclear Regulatory Commission Reactor Operator Generic Fundamentals Examination (GFE), scoring it paper by paper against the 80% criterion applied to every human candidate, with no rounding up. The evaluation set is a census of every GFE administered at the March sitting from 2015 to 2021, giving seven pressurized water reactor (PWR) and seven boiling water reactor (BWR) papers and 697 scored items. Eight configurations cross three model states, the base model, supervised fine-tuning (SFT) on distilled chain-of-thought rationales and retrieval-augmented fine-tuning (RAFT), with three retrieval conditions, none and BM25 retrieval over the Department of Energy Fundamentals Handbooks under fixed-size and structure-aware chunking. Out of the box it answers 51.94% correctly and passes no paper. SFT with fixed-size chunking retrieval passes 8 of 14, reaching 80.23% on PWR items and 79.77% pooled, with a Wilson interval spanning the threshold. The preferred chunking granularity reverses with training state, structure-aware before fine-tuning and fixed-size after, so chunking optimized against a base model cannot be inherited by its fine-tuned descendant. RAFT trails SFT by 2.2 to 2.3 percentage points overall, and the deficit holds in all four reactor-type and chunking strata. The pipeline runs on one workstation with no network access at run time, and the result approaches operator-level command of engineering fundamentals without reliably achieving it.
LLM-based evaluators of natural language generation (NLG) quality are widely deployed as scoring tools and as automated training signals, yet the internal procedure by which they assign a rating remains poorly understood. We investigate this procedure mechanistically through an eight-attack perturbation taxonomy across the Readability and Adequacy dimensions of NLG quality, a generation pipeline that produces paired clean and corrupt summaries with controlled error intensity and explicit token-level modification maps, and a four-experiment battery of causal tracing, logit-lens vocabulary projection, and attention-head knockout applied to Themis (Llama-3-8B) and Prometheus (Mistral-7B). Both evaluators implement a structured, coherent evaluation pipeline operating in two stages: below layer 15, attention performs local error comparison and routes the result to the final input position; above it, the MLP cascade integrates the signal and writes the rating, with the decision crystallizing in the residual stream at a sharp late layer (L = 26 on Themis, L = 25 on Prometheus). Furthermore, a base-model control at the same scale (Llama-3-8B) reproduces the routing architecture and crystallization but not the stage separation, isolating the two mechanisms that fine-tuning specifically installs, suppression of below-L15 MLP contribution at the last position and a two-layer advance of the crystallization depth, indicating that fine-tuning sculpts an existing substrate rather than building the pipeline from scratch. We release the source code and data at https://github.com/himil-v/judge-mech
This work represents the first application of online control mechanisms to adaptively select prompting strategies in AES, transforming prompt selection from an offline hyperparameter optimization problem into an efficient online learning task.
EYT-Bench is introduced, a human-centered benchmark whose evaluation protocol is built around a decoupled three-party design: a persona-grounded user simulator, a target model evaluated on both intent perception and response generation, and an independent, configurable ensemble of LLM judges.
