← Artificial intelligence this week · 16 November 2026 · 5 min read
Artificial intelligence this week: What impressive scores leave out
New preprints examine the distance between headline performance and useful knowledge, while showing how simulations and physical models can help when data are scarce.
How to check the findings and keep reading? Illustrated guide
How these papers were chosen among the week's 2471 papers
All 2471 papers of the week whose main category is in this field were read, title and start of the abstract, and 100 were kept. They were then ordered by what can be counted: how many research profiles of our readers they come close to, whether a rising topic of the field includes them, and how many categories they were posted in. The abstracts of the first 40 were read in full to choose these, no more than two from one category and one per team.
2 of them belong to a rising topic. The 17 numbers of the text were checked against the abstracts. 27 statements on what the papers found were looked for word for word in their abstracts.
Sources: arXiv.
A virtual cell can look right while missing the response #
Predicting how a cell reacts when a gene is disrupted could help researchers explore experiments before performing them. But the usual score may reward models for predicting what does not change. In a study of cellular perturbation prediction, GEARS achieved a global Pearson correlation of 0.977 between predicted and observed gene expression. When the researchers instead measured changes relative to an unperturbed reference, its correlation fell to 0.076. A simple baseline using the average change in the training data scored 0.267. Thousands of largely unresponsive genes had made the overall prediction look much better than its account of the biological response.
The authors propose evaluating those changes directly and train a model to emphasize them. This improves performance without beating the simple baseline on every measure, while transfer to an unseen cell line offers evidence that the learned representation contains useful information. The broader lesson is familiar well beyond biology: a prediction can closely resemble reality while missing precisely the departure from normal that matters.
-
The Delta Gap: Cellular World Models for Perturbation Response Prediction Are Far Harder Than Global Correlation Implies
Junhui Gao · arXiv cs.AI · Benchmark · Preprint, not yet peer reviewed
Rests on: 509 unseen perturbations, 5000 genes
Main limit: The delta-aligned model still trails the trivial baseline on delta Pearson in the Replogle evaluation.
Read with it: A systematic comparison of single-cell perturbation response prediction models (Science Advances, 2026) · SubcellPert: Benchmarking Subcellular Perturbation Prediction (bioRxiv (Cold Spring Harbor Laboratory) preprint, 2026)
Worth knowing?
Prepare a press kit Premium+ →
Simulated outbreaks help most when local records are thin #
Disease forecasting faces an awkward shortage: the outbreaks a health agency most needs to anticipate may have few local precedents. The Mantis researchers investigated how to balance simulated epidemics with observed surveillance data, holding forecasting tasks fixed while varying the available history. Simulation-heavy training was especially helpful when local observations were limited. Its relative advantage declined as more observations became available, and it also reduced exceptionally large forecast errors.
Their resulting model, Mantis 2, combines simulated and observational training to forecast up to six variables sixteen weeks ahead. Across 30 forecasting tasks, it achieved the strongest aggregate probabilistic accuracy among the evaluated models and missed roughly 30% fewer outbreak peaks than the strongest comparison. The important finding is not that artificial data can replace surveillance, but that mechanistic simulations can supply useful prior experience where surveillance is sparse.
-
Universal Infectious Disease Forecasting with Mantis 2
Carson Dudley, …, Marisa Eisenberg (6 authors) · arXiv cs.AI · Experiment · Preprint, not yet peer reviewed
Rests on: 30 forecasting tasks
Main limit: The advantage of simulation-rich training diminishes as more observational history becomes available.
Read with it: Leveraging synthetic and genetic data to improve epidemic forecasting (PLoS Computational Biology, 2026) · GENIE: Generative Neural Inference for Epidemics (arXiv (Cornell University) preprint, 2026)
Worth knowing?
Prepare a press kit Premium+ →
The hidden language surcharge #
Language-model services charge for tokens, the fragments into which text is divided. Those fragments are not equally economical across languages. Researchers compared nine public tokenizers using professionally translated, English-aligned sentences in the languages of India, alongside human-translated Hinglish-English social media pairs. With GPT-4o's tokenizer, equivalent content required 1.57 times as many tokens in Hindi as in English, 5.01 times in Odia, and 13.7 times in Santali written in Ol Chiki script.
This affects more than a bill: extra tokens consume context space and can increase processing time. The study also found weak alignment between token boundaries and meaningful word parts in Hindi and Marathi. Its cost calculations are explicitly illustrative rather than measurements of production services. Nevertheless, the comparison makes an infrastructure choice visible: the way a model divides writing into pieces can distribute its practical benefits unevenly across language communities.
-
The Indic Token Tax in 2026: Measuring Tokenizer. Disparity and Its Cost for the Languages of India
Johin Johny Arimbur, …, Shervin Sam (3 authors) · arXiv cs.CL · Benchmark · Preprint, not yet peer reviewed
Rests on: 2,009 professionally translated sentences aligned with English
Main limit: The monetary estimates depend on stated assumptions and list prices rather than measured production costs.
Read with it: The Tokenizer Tax: Quantifying and Explaining the Cross-Lingual Cost of Subword Tokenization for Indian Languages (arXiv (Cornell University) preprint, 2026) · Is Sanskrit the Most Token-Efficient Language? A Quantitative Study using GPT, Gemini, and SentencePiece (arXiv (Cornell University) preprint, 2025)
arXiv PDF Code or data Summary Ask this paper Cite (RIS) Journal club Premium+ Propose for the journal clubWorth knowing?
Prepare a press kit Premium+ →
Stricter instructions are not necessarily better judgment #
When a system prompt makes a language model refuse more requests, has the model become safer—or simply more reluctant? A controlled evaluation of 13 models used signal detection theory to separate the ability to distinguish safe from unsafe requests from the threshold for refusing them. Keeping benchmark questions fixed, the researchers varied the prompt family and its clarity. System prompts mainly moved the refusal threshold; changes in discrimination were smaller and less consistent.
Clearer instructions likewise increased willingness to refuse without reliably improving the underlying distinction. Prompts also shifted responses toward behaviors such as clarification and safe assistance, showing why a simple comply-or-refuse score is incomplete. For people deploying these systems, the implication is practical: a more restrictive assistant should not automatically be interpreted as a more discerning one.
-
System Prompts Affect Response Bias More Than Discriminability in LLM Safety
Mohan W. Gupta and Sandy Tanwisuth · arXiv cs.LG · Experiment · Preprint, not yet peer reviewed
Rests on: HarmBench, XSTest, and IHEval
Main limit: The evidence concerns responses to fixed benchmark items rather than demonstrated safety outcomes in deployment.
Read with it: From Detection to Refusal: Safer LLMs via Circuit-Guided Weight Scaling (Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP), 2026) · When Refusal Looks Safe: The Refusal-Cue Shortcut in Safety Guard Models (arXiv (Cornell University) preprint, 2026)
Worth knowing?
Prepare a press kit Premium+ →
Cheap arithmetic can conceal expensive forgetting #
Lower-precision arithmetic promises cheaper model training, but an almost unchanged loss curve may conceal a specific learning deficit. Researchers trained transformers on synthetic biographies with controlled exposure counts, both alone and mixed into web text. NVFP4 needed 1.32 times and 1.52 times as many exposures as BF16 to reach comparable factual recall, despite validation-loss differences of only 0.5% and 2%. The gap persisted when evaluation used BF16 arithmetic, pointing to a training effect rather than just a readout problem.
Controlled noise experiments implicated degraded gradient information: making updates noisier could impede rare-fact learning while barely disturbing the aggregate loss. Increasing the learning rate did not recover the cost. These experiments do not establish how much factual knowledge every large production model loses, but they identify a blind spot in efficiency comparisons. Matching average predictive performance need not mean learning infrequently encountered information equally well.
-
What the Loss Curve Misses: FP4 Pretraining Slows the Acquisition of Rare Facts
Chang Lu and Mengjiao Zhang · arXiv cs.LG · Experiment · Preprint, not yet peer reviewed
Rests on: synthetic biographies with controlled exposure counts
Main limit: The rare-fact results come from controlled synthetic biographies and emulated low-precision arithmetic, not a direct measurement of factual learning in production models.
Read with it: FP4 All the Way: Fully Quantized Training of LLMs (2025)
Worth knowing?
Prepare a press kit Premium+ →
Recovering tissue properties by running the physics backward #
An optical coherence tomography image mixes tissue properties with the physics of the instrument that recorded it. A new method tries to disentangle them by adjusting a physical model until its simulated image agrees with the observed one. Using gradients through that model, the researchers recover a per-pixel optical attenuation coefficient, a quantity associated with tissue composition and pathology, from a single image without training data or segmentations.
The team checked the method against analytical solutions, validated it on five in vitro phantoms, and applied it without modification to clinical intravascular images from 128 patients. Application to clinical images is not itself proof of improved diagnosis, but the approach offers a way to extract quantitative information without training a separate predictor for each imaging setup. Alongside the outbreak work, it illustrates a productive role for scientific knowledge: not merely interpreting a model afterward, but structuring the computation itself.
-
Differentiable Inverse Rendering Recovers Intrinsic Properties from Intravascular Optical Coherence Tomography
Ajay Manicka, …, Elazer R. Edelman (10 authors) · arXiv cs.CV · New method · Preprint, not yet peer reviewed
Rests on: five in vitro phantoms
Main limit: Application to clinical images does not establish that recovered attenuation improves diagnosis or patient outcomes.
Read with it: Development of an Intravascular Doppler Optical Coherence Tomography Imaging Technique for Neurovascular Applications (2021) · Development and applications of optical imaging techniques for microcirculation and cardiovascular imaging (Arrow@dit (Dublin Institute of Technology), 2016)
Worth knowing?
Prepare a press kit Premium+ →
These preprints share a concern with what counts as evidence of progress. Overall correlation can hide missed biological changes; average loss can hide weaker factual learning; more refusals can masquerade as better judgment. Token counts expose another overlooked dimension: who pays for apparently universal capabilities. The forecasting and imaging studies offer a complementary direction, using mechanisms and physics to make limited observations more informative. Across these fields, the useful question is not simply whether a model scores well, but whether the score captures the scientific or practical job it is meant to do.
Rising in this field
- Can AI inspect and edit another AI’s internal settings? Artificial Intelligence (cs.AI) · ×15.3 compared with usual · 10 papers from 10 teams
- Learning patterns that respect rotations Machine Learning (cs.LG) · ×9.2 compared with usual · 8 papers from 8 teams
- Spotting unusual defects in images Computer Vision and Pattern Recognition (cs.CV) · ×12.1 compared with usual · 6 papers from 6 teams
Was this letter useful to you?
Thank you: your answers help choose the papers of the next letters.
Written by AI from the 3045 papers of the week in this field. Most are preprints, not yet peer reviewed: the text says what their authors report.
Get one letter by email every week with a free account, all of them with Premium. Create a free account