LoRA, PEFT, instruction tuning and domain adaptation — adapting a pretrained model without paying to train one.
Designing systems that attribute a generated image to its source model has lately become a major area of AI research. Most existing datasets and systems related to image source attribution are targete...
The study investigates how quantization affects subliminal preferences in AI models, specifically a hidden "favourite animal: panda" preference implanted in Qwen2.5-7B-Instruct. While 8‑bit quantization preserves the preference, 4‑bit quantization produces divergent effects: Half‑Quadratic Quantization amplifies the preference to 64% panda probability, whereas NormalFloat suppresses it to 17%. The results show that the arrangement of rounding errors, not just their magnitude, drives these shifts, indicating that a 16‑bit audit cannot reliably predict real‑world behaviour under deployment‑level compression.
The study analyzes 12,933 responses from a design crossing 20 brand labels, eight query languages, three provider aliases, and 15 prompt identifiers, using a multilingual sentiment classifier to produce signed confidence scores. A crossed linear mixed model attributes 26.49% of variance to query language and 1.46% to brand, but when English is excluded the language contribution drops to 5.07% and brand to 2.23%, indicating that the variance decomposition is highly sensitive to the inclusion of English. The findings highlight that automated sentiment scores can vary due to both answer content and the measurement instrument, and that English responses are not necessarily more negative toward brands. whyItMatters":"The results caution against drawing substantive conclusions from automated sentiment scores without accounting for language and instrument effects, as these factors can disproportionately influence the measured sentiment."
MemRole is a framework that explicitly models the memory-to-decision process for large language model (LLM) role‑playing agents, enabling them to make role‑consistent decisions over long narrative horizons. It introduces Long‑Horizon Memory Thinking (LHMT), a reasoning schema that decomposes hidden reasoning into context understanding, memory activation, and response strategy planning, and is trained on a novel dataset that limits each character’s memory pool to what it could plausibly know at the current narrative point. The framework also includes Self‑Refining On‑Policy Self‑Distillation (SR‑OPSD), where the model critiques its own rollouts and distills token‑level guidance back into the policy, resulting in a 15.2‑point improvement over its backbone and competitive performance with frontier LLMs such as GPT‑5.4.
The paper introduces Discrete Latent Reasoning (DLR), a method that transforms continuous latent states into explicit discrete tokens by rendering textual chains of thought into images, extracting visual features, and clustering them to build a discrete latent vocabulary. DLR enables standard autoregressive modeling over both natural language and latent tokens, supporting pretraining alignment, supervised fine-tuning, and reinforcement learning. Experiments on five reasoning benchmarks and two model series (Qwen3‑VL and LLaMA‑3) show that DLR outperforms prior latent reasoning baselines with up to 20× compression while preserving interpretable semantic structure.
The paper introduces selective linkage precision, a metric for evaluating how accurately vision‑language models can match de‑identified chest radiographs to their reports when only certain attributes (such as sex, age, projection, or acquisition year) are available. Experiments on eight public VLMs and three chest‑radiograph datasets show that filtering by these attributes dramatically improves precision—from a baseline of 1.98% to 67.68% with sex, age, and projection, and up to 99.05% when acquisition year is added—while keeping the candidate set manageable. The study demonstrates that this approach can be applied without retraining the models and may guide hospitals in deciding which metadata to release.
The study examines how large language models (LLMs) respond to users in online political debates, finding that LLMs are more receptive to left‑leaning users than right‑leaning ones. This bias persists even when users defend the same position or appear together, and is influenced by user expertise and, to a lesser extent, rhetorical strategy. Fine‑tuning on different political corpora can reduce or reverse the effect, indicating that the bias is not inherent to the models but can be altered.
The paper introduces Concise-9B, a version of the Qwen3.5-9B reasoning model that reduces unnecessary thinking tokens through concise self‑training. By applying two rounds of on‑policy rejection‑sampling fine‑tuning and a small share of answer‑synthesis targets, Concise-9B cuts token usage to roughly two‑thirds of the original model while maintaining comparable accuracy on a wide range of hard mathematics and reasoning benchmarks. The study also shows that most of the token savings occur after the answer has already formed, and that the approach’s effectiveness depends on the verbosity of the base model.
TMax is an open reinforcement‑learning recipe that trains terminal‑using agents from small language models, achieving 33% on Terminal‑Bench 2.0 with only 9B parameters and outperforming larger prior models. The approach generates training data via a novel taxonomy that blends difficulty control, personas, and verifier diversification, enabling cheap creation of many terminal environments for both RL and supervised fine‑tuning. The recipe is applicable to models ranging from 2B to 27B across different families, dominating the Pareto curve for open terminal‑agent performance at smaller sizes.
The paper introduces ALPS, a hierarchical simulated student model that alternates between planning profile‑aligned task segments and generating corresponding actions, aiming to maintain consistent student behavior over long‑horizon learning environments. Unlike existing LLM‑based simulators that focus on short, language‑centered interactions, ALPS is conditioned on observable, environment‑specific behavior profiles and evaluated in both data‑sufficient and data‑scarce settings. Experiments in an inquiry‑based physics game and a pharmacy simulation show that ALPS consistently outperforms turn‑level generation, achieving comparable behavioral density and judge pass rates even with limited real student data, though coverage remains lower.
The paper reexamines the notion of large language model (LLM) homogeneity by introducing 25 interpretable metrics across semantic, lexical, syntactic, structural, and stylistic dimensions. A human study reveals that perceived similarity depends on both content and form, with embedding similarity correlating less strongly with human judgments than sentence-level matching. Across 18 contemporary LLMs and over releases from 2024 to 2026, models share global semantics but diverge in lexical, structural, and interactional patterns, and model-specific conventions emerge during fine‑tuning and preference optimization.
MEMENTO is a long‑context agent built on Qwen3‑8B that learns to manage its fixed‑size context window through compaction, replacing its history with self‑generated notes. The model is first fine‑tuned on synthetic trajectories with hindsight notes and then reinforced with a reward combining answer correctness, note quality, and confidence calibration. At test time, MEMENTO can also fold context and backtrack based on confidence, achieving a 67.8% accuracy on book‑level QA (up to 2.5 M tokens) while never holding more than 20 K tokens, outperforming GPT‑5.5 and GLM‑5.3 agents at significantly lower cost.
The paper examines how fine‑tuned Transformer detectors (DistilBERT, RoBERTa) perform on an out‑of‑distribution dataset of hyper‑local news generated by LLMs. Using SHAP‑based explainable AI, the authors uncover three phenomena: a Call‑to‑Action bias, a Specificity Bias where hallucinated named entities evade detection, and a Wikipedia Paradox where formal human prose is misclassified as AI. These results highlight systematic shortcomings of purely statistical detectors in structured human domains and underscore the need for semantically‑aware, interpretable detection methods.
XDecideBench is a benchmark comprising nine suites across 122 languages, containing 88,425 texts and 153,875 typed questions, designed to evaluate Typed Decision Models and their confidence policies. The benchmark reports results separately for aligned, locally labelled, and native items, revealing that model rankings vary by suite and language, with the best English model performing poorly in many other languages. Fine‑tuned open models outperform their base Qwen counterparts in all twelve development languages, and confidence thresholds set on English data can exceed acceptable error rates on other datasets such as XNLI.
The paper introduces EmoNet‑Face‑HQ, a benchmark of 2,500 AI‑generated portraits each evaluated by four psychology experts across 40 emotion categories. It shows that most vision‑language models (VLMs) perform no better than random guessing unless fine‑tuned, but that changing the prompt to a simple yes/no per category can raise a VLM’s agreement with experts from chance to near expert levels. When evaluated using probability‑of‑yes scores on the five most‑agreed‑upon categories, eleven open‑weight VLMs achieve a quadratic‑weighted κ of 0.425, below the 0.468 two‑expert agreement, yet their average probability‑of‑yes score (0.553) matches the benchmark’s fine‑tuned model and outperforms it in three cases without any training.
The paper explores using LLM-generated annotations to train smaller BERT models for literary sentiment analysis. Five BERT variants were fine‑tuned with ordinal regression on 10,889 sentences labeled by Gemma 4 31B on a ten‑point scale, and evaluated on 1,923 human‑annotated sentences from Hemingway’s *The Old Man and the Sea*. BERT‑large achieved a Kendall’s τ of 0.589, close to the teacher’s 0.604, and BERT‑base outperformed the teacher on non‑neutral sentences with an MAE of 0.726 versus 1.116.
AutoSpecNER is a new expert‑annotated dataset for fine‑grained named entity recognition in vehicle listings, comprising 659 advertisements and over 10,000 entities across 15 categories such as MODEL, ENGINE_SPEC, and BATTERY_CAPACITY. The annotation quality was validated with an inter‑annotator agreement score of 91.5%. Benchmarking shows that DeBERTa achieves a 90% micro‑F1, far surpassing a rule‑based baseline (43%) and the best large language model (77.8%).
Sonar‑MoE introduces a memory‑efficient training technique for Mixture‑of‑Experts models by updating each expert tensor during the backward pass and immediately freeing its gradient, ensuring that GPU memory never holds more than one layer of expert gradients. In experiments fine‑tuning OLMoE‑1B‑7B and Granite‑3.1‑3B‑A800M in bf16, Sonar‑MoE reduces peak memory usage by 19% to 22% and speeds up each training step by 17% to 23% compared to standard training with global‑norm clipping. The method achieves these gains while maintaining bitwise identical updates to standard training (without expert gradient clipping) and keeping held‑out loss within 0.003 of all baselines.
CST is a lightweight risk triage system that provides moderators with source‑grounded evidence for web discussions. It first assigns risk scores using character patterns and a multilingual model, then groups related messages and selects key excerpts to show, linking them back to the original thread. Across multiple languages and datasets, CST outperforms fully fine‑tuned models in ranking harmful messages and reduces moderator review time by at least 27% while maintaining high accuracy.
The study evaluates whether multimodal models produce consistent decisions when presented with equivalent optical spectra in different representations—tables in nanometers, micrometers, scientific notation, or photon energy, and plots in wavelength or energy. Using datasets of silicon metasurfaces and gold nanoparticles, the authors test several models (d3-nano, Jev 1.13, GPT‑6 Luna variants, Qwen3.5‑2B) and find that representation changes can shift accuracy by up to 41 percentage points, with some models benefiting from photon‑energy inputs while others suffer. Fine‑tuning with LoRA improves performance on trained formats but often relies on a shortcut of abstaining when data are truncated; training with answerable truncations mitigates this shortcut, revealing significant gaps in models’ handling of missing evidence.