arXiv Computation and Language

Where a Model Sends Its Own Repeated Token

arXiv:2609. 31181v1 Announce Type: new Abstract: Black-box model identification works by scoring a model's response to natural-language prompts.

arXiv Computation and Language
Sep 1

Token Counts Are Not Model Lineage: A Frozen-Threshold Holdout Study of Black-Box LLM API Fingerprinting

The study evaluates whether prompt‑token counts can reliably identify the lineage of large language models served via APIs. Using a frozen‑threshold approach on 24 labeled endpoint pairs, the authors find that token‑count consistency perfectly separates development pairs but only half of the holdout pairs meet the strict repeatability criteria, yielding moderate accuracy and perfect specificity. The results confirm token‑count consistency as a fingerprint of shared tokenization stacks but reject it as a standalone test for model‑family attribution.

By Bo Chen
arXiv Computation and Language
Sep 1

Manac\'a-1B: An Open, Reproducible Brazilian-Portuguese Language Model and a Tokenizer-Aware, Paired Evaluation

Manacá-1B is a 1.72‑billion‑parameter, open decoder‑only language model trained from scratch for Brazilian Portuguese, released with a fully containerized, reproducible training pipeline and complete logs. The authors evaluate it against nine open baselines on four Portuguese benchmarks, reporting standard errors and paired significance tests, and find that Manacá-1B outperforms smaller models on LAMBADA‑PT while remaining competitive on commonsense completion. They also uncover a tokenizer‑related evaluation pitfall that can drastically lower accuracy and provide a simple fix, releasing all code, logs, and corrected tokenizer for full reproducibility.

By Bruno Leonardo Santos Menezes, Carlos Leonardo Souza Cardoso, Fabio Andre Machado Porto
arXiv AI
Jul 22

Prompt Design at Scale: How Format, Instruction Count, and Context Length Shape Instruction Adherence and Hallucination in Large Language Models

arXiv:2607. 19257v1 Announce Type: cross Abstract: Practitioners make three prompt-design decisions with almost no controlled evidence behind them: how to format instructions and context (markdown, plain text, prose, or tabular), how many simultaneous instructions a system prompt can carry before compliance degrades, and how much context a model can hold before recall and honesty degrade.

By Netanel Eliav
arXiv AI
Sep 4

CASCADE: A Component Ablation and Corpus Audit of a Layered Local Defense for MCP-Based Systems

The paper evaluates CASCADE, a fully local layered defense for Model Context Protocol (MCP)-based systems, by conducting a component ablation and corpus audit on a fixed 5,000-sample dataset. It demonstrates that the choice of aggregation convention heavily influences reported metrics, that detection performance varies with provenance, and that the released configuration does not fully disclose the operating point. The study also shows that a local review model invoked for a third of requests does not alter classification outcomes, highlighting the importance of reproducibility and transparency in defense evaluations.

By \.Ipek Abas{\i}kele\c{s} Turgut, Edip G\"um\"u\c{s}
arXiv Computation and Language
5d ago

What a Cross-Model Fixed-Point Census Can and Cannot Arbitrate About Repetition

The paper investigates neural text degeneration by measuring the fixed‑point structure of short‑window argmax maps across 17 pretrained models, using 96 random two‑token starts without prompts. It finds a stable four‑way classification that varies across model families and scales, with some models funneling to a single endpoint token while others do not, and shows that this behavior is not solely determined by training data or corpus frequency. The study demonstrates that repetition phenomena are not uniformly explained by either training data or network architecture alone, highlighting the complexity of neural text generation dynamics.

By Nicol\'as Vera Z\'u\~niga
arXiv Machine Learning
Sep 23

Same Quantity, Different Answer: Numerical Representation Invariance in Language Models

The paper introduces a benchmark of 3,600 exact‑rational word problems and 8,600 prompts that test whether language models give the same canonical answer when the same quantity is expressed in different numeric forms (decimal, fraction, percentage, number word, scientific notation, or unit‑converted). After normalizing answer syntax, canonical accuracy is high (0.969–0.996), but correctness across equivalent representations drops to 0.848–0.981, revealing that many errors stem from the evaluator’s number grammar rather than the models’ reasoning. The study also finds that representation consensus does not outperform paraphrase consensus on low‑error subsets and that certain models (e.g., Mistral Small 4) exhibit systematic unit‑conversion errors.

By Ephraim Atta-Duncan
arXiv Machine Learning
Sep 23

What Does 99% Accuracy Measure? A Reproducible Audit of Shortcut Learning in a Widely Used Fake News Corpus

The paper audits the widely used ISOT/Kaggle Fake and Real News corpus and finds that extremely high reported accuracies (≈0.98) are largely due to shortcut signals rather than genuine veracity detection. A simple TF‑IDF linear classifier achieves perfect F1 when using only subject metadata, and even after removing metadata, newswire tags, and duplicate documents, the F1 drops only modestly, indicating that editorial style rather than specific tokens drives performance. Under topic‑disjoint and temporal transfer tests, performance collapses, and models transfer poorly to the independent LIAR benchmark, showing that within‑corpus scores reflect source and topic separability, not truth verification. whyItMatters:"The study demonstrates that current high accuracy metrics on this fake‑news dataset are misleading, highlighting the need for more robust evaluation protocols that guard against shortcut learning."

By Yuvraj Verma
arXiv Computation and Language
Aug 31

Nested Byte-Level Vocabularies Are Cheap to Deploy and Expensive to Share: A Pre-Registered Negative Result

The paper demonstrates that a byte‑level BPE tokenizer can be sliced to create multiple vocabulary sizes from a single trained model, preserving exact logits while reducing deployed weights by 66%. Experiments on 30 models show that while sliced models match the full model numerically, they underperform fixed‑cap specialists by a few percentage points in bits‑per‑byte. Multi‑cap training improves robustness to typographical noise, suggesting benefits from training across multiple granularities rather than from control tokens alone.

By Christos Koutsiaris