标签：#datascience

We maintain a corpus of 456 role-specific resume examples in TypeScript. Someone asked me what a good bullet point actually looks like, and rather than answer from taste I decided to measure the thing I already had. Fifteen minutes later we had a script, a set of numbers, and a conclusion. The conclusion was wrong, because the script had silently read about twelve percent of the data. This is a post about that failure mode, and then about the numbers I got once the script worked. The corpus Thirty-one TypeScript files, each exporting an array of role objects. One role looks roughly like this: { slug : ' cloud-architect ' , title : ' Cloud Architect Resume ' , category : ' Information Technology ' , sampleData : { summary : ' ... ' , experiences : [ { company : ' Amazon Web Services ' , position : ' Senior Cloud Architect ' , description : ' - Designed multi-region architecture... \n - Led migration of... ' , }, ], skills : [...], }, tips : [...], } The interesting field is description . It holds a newline-delimited list of bullets as a single string, so the whole corpus of bullets is sitting there in source, greppable, without a database or an export step. Version one const descs = [... text . matchAll ( /description: ' ((?:[^ ' \\] | \\ . ) * ) '/g )]. map ( m => m [ 1 ]); Nothing exotic. Match description: , then a single-quoted string, allowing escapes so an apostrophe inside the text does not terminate the match early. It found 199 description strings. I did not question that, because I had no prior for what the number should be. 199 sounded like a lot of text. We computed medians off it, looked at the opener distribution, and started writing. The number that saved me was on a different line of the same output: roles 456 . The slug count was fine. So 456 roles between them had 199 job descriptions, which would mean the overwhelming majority of roles had no work history at all. I knew that was false, because I had rendered these pages. Why it read twelve percent

Ever woke up feeling like a truck hit you, only to realize your Apple Watch had been screaming "Warning!" via your data for the last 24 hours? Heart Rate Variability (HRV) is the "canary in the coal mine" for our bodies. It's a powerful metric that tracks the variation in time between each heartbeat, serving as a direct window into your Autonomic Nervous System. In this guide, we are going to build a real-time HRV anomaly detector using wearable data analysis , Scikit-learn , and AWS Lambda . By applying machine learning to time-series health data, we can identify physiological stress, potential infections, or overtraining before physical symptoms even manifest. If you’ve been looking to dive into anomaly detection in time-series or want to master health data engineering , you’re in the right place! The Architecture: From Heartbeat to Alert 🛠️ To achieve real-time monitoring, we need a pipeline that moves data from your wrist to a cloud-based inference engine. Here is the high-level flow: graph TD A[Apple Watch / Wearable] -->|Sync| B(Apple HealthKit) B -->|Webhook/Hook| C[AWS API Gateway] C --> D[AWS Lambda - Inference] D -->|Fetch History| E[(DynamoDB / S3)] D -->|Isolation Forest| F{Anomaly?} F -->|Yes| G[Push Notification / Alert] F -->|No| H[Log & Silent] Prerequisites 📋 Before we start coding, ensure you have the following: Python 3.9+ Scikit-learn & Pandas for data crunching. AWS Account (for Lambda deployment). An app to push HealthKit data (like Health Auto Export or a custom Swift hook). Step 1: Understanding the Data 📊 HRV data is tricky because it’s highly personalized. What is "low" for an athlete might be "high" for someone else. This is why we use Isolation Forest , an unsupervised learning algorithm that excels at detecting outliers in multi-dimensional datasets without needing labeled "sick" vs. "healthy" days. Step 2: Building the Anomaly Detection Logic Let's write the core logic using Scikit-learn . We’ll use the Isolation Forest algorithm becaus

Introduction Jumia is one of Africa's leading e-commerce platform that manages millions of transcations with a diverse products from electronics,beauty products and many more categories.Therefore,tracking key perfomance indicators is essential for supply chain operatios and profit optimization Objective My project aim is to build an interactive excel dashboard using Jumia transactional data.I aim to convert disorganized data into an interface that can help in decison making,identify trends and monitor products. Dataset description Data Cleaning and preparation process Raw data mostly contains inconsistency and errors that may occur that may interfere or give the wrong output. An example of a raw dataset In the example above we can see inconsistent and missing data that we need clean in order to have an effective output. First step is to format the prices from text to currency format and replace the before since excel will in order to calculate the discount eg below The image above is the discount price which was obtained by finding the difference between the old price and the new price. The image below is an example of the formula to categorize the prices whether high,low or medium.I used the IF,AND functions.Another example of a logical combination would be the us of OR . The difference when using the IF(AND function is that all the conditions must be met while in the IF(OR ,only one condition has to be met. In the image below i used logical combination of that are IF and AND for the discount category. In the image below i also used the IF AND functions to in the ratings category. After removing duplicates,removing inconsistent data eg texts in numbers columns.Below is an image of the cleaned version of the Jumia dataset. An example of a clean dataset Descriptive Analysis To calculate the average current price of products i used the average formula and highlighted the cells eg =AVERAGE(B2:B113) .The average old price of products was obtained by the same formula but

Turning Data Into Decisions Bar charts, histograms, scatter plots, subplots, and plotting straight from pandas Previously learned to draw a line — literally. we now know how to create a figure, style it, and save it. But real analyst work rarely stops at trends over time. You'll need to compare categories , understand distributions , spot relationships between variables , and show several views of the data at once . That's exactly what today covers. Grab a coffee — let's turn raw numbers into charts that actually tell a story. 1. Bar Charts: Comparing Categories When to use one Bar charts are your go-to whenever you're comparing discrete categories against each other — regions, products, departments, months. If someone asks "which one is bigger?", a bar chart answers it instantly. The code import matplotlib.pyplot as plt regions = [ " North " , " South " , " East " , " West " ] revenue = [ 420 , 380 , 510 , 290 ] fig , ax = plt . subplots ( figsize = ( 7 , 5 )) ax . bar ( regions , revenue , color = " teal " ) ax . set_title ( " Revenue by Region " ) ax . set_xlabel ( " Region " ) ax . set_ylabel ( " Revenue ($K) " ) plt . show () A useful variant: horizontal bars When category names are long, flip the chart with barh() — it's far easier to read than squeezing labels sideways: fig , ax = plt . subplots ( figsize = ( 7 , 5 )) ax . barh ( regions , revenue , color = " darkorange " ) ax . set_title ( " Revenue by Region " ) ax . set_xlabel ( " Revenue ($K) " ) plt . show () Rule of thumb: categories on the x-axis → bar() . Long labels or many categories → barh() . 2. Histograms: Understanding Distributions Bar chart vs. histogram — don't mix them up This trips up almost every beginner: a bar chart compares separate categories. A histogram shows how continuous numeric data is distributed by grouping values into ranges called bins . There are no gaps between histogram bars by convention, because the x-axis is continuous, not categorical. The code import matplotlib.pyplot

Every charting tool eventually gets the same feature request: "show me other times this stock looked like this." It sounds like a lookup. It is not. The retrieval is the easy half. The hard half is that a correct implementation can still produce results that are quietly meaningless, and nothing in the code will tell you. Here is the method, and the failure modes worth knowing before you ship it. No claims about predictive power anywhere in this piece — the last section explains why that is a deliberate choice, not a hedge. The naive version, and why it fails immediately The obvious first attempt: take the last 30 days of closing prices as a query vector, slide it across history, compute Euclidean distance, return the closest matches. import numpy as np def naive_search ( history , query , k = 5 ): m = len ( query ) windows = np . lib . stride_tricks . sliding_window_view ( history , m ) dists = np . linalg . norm ( windows - query , axis = 1 ) idx = np . argsort ( dists )[: k ] return idx , dists [ idx ] Run this and you get garbage — but instructively specific garbage. Every match comes from whatever period had a similar price level . Query a stock trading at $180 and you get back the other times it traded near $180. The shape is irrelevant to the metric; the offset dominates it. Scale is the same problem in a different coat. A stock that moved 2% over the window and one that moved 40% can trace an identical shape, and raw distance calls them unrelated. Normalize per window, not globally The fix is to z-normalize each window independently: def znorm ( x , axis =- 1 , eps = 1e-8 ): mu = x . mean ( axis = axis , keepdims = True ) sd = x . std ( axis = axis , keepdims = True ) return ( x - mu ) / ( sd + eps ) Per-window is the load-bearing part. Normalizing the whole series once preserves the relative offsets you were trying to remove. Each candidate window has to be centered and scaled on its own terms before it's compared. There's a satisfying identity waiting here.

Have you ever wondered how your smartwatch actually knows you're stressed? Most of us treat the "Stress Score" on our wrists as a source of truth, but the logic remains hidden behind proprietary algorithms. Today, we are pulling back the curtain. We are going beyond basic heart rate tracking to perform PPG signal processing and HRV frequency domain analysis using Python. By the end of this guide, you’ll know how to ingest raw data via Bluetooth Low Energy (BLE) , apply digital filters with SciPy , and calculate the elusive LF/HF ratio to determine autonomic nervous system balance. If you are interested in advanced biometric algorithms or Python signal analysis , you’re in the right place. 🚀 The Architecture: From Photons to Stress Metrics Unlike standard heart rate (BPM), which just counts peaks, Stress Scores rely on Heart Rate Variability (HRV) —the millisecond-level variations between heartbeats. We'll be moving from raw light intensity data to a frequency-based stress index. graph TD A[Wearable Sensor / PPG] -->|Raw BLE Stream| B[Data Acquisition - Bleak] B --> C[Preprocessing - Bandpass Filter] C --> D[Peak Detection - Find R-R Intervals] D --> E[Cubic Spline Interpolation] E --> F[Fast Fourier Transform - FFT] F --> G[LF/HF Ratio Calculation] G --> H[Final Stress Score] 🛠 Prerequisites To follow this advanced tutorial, you’ll need: Hardware : A pulse oximeter or wearable that exposes raw PPG via BLE (e.g., Polar OH1, MAX30102 with an ESP32). Stack : NumPy & SciPy : For heavy-duty math and signal processing. Bleak : For cross-platform Bluetooth Low Energy communication. Matplotlib : To visualize the pulse waves. Step 1: Capturing the Raw PPG Stream (BLE) Photoplethysmography (PPG) works by shining green or red light into the skin and measuring the light absorption. First, let's grab that raw stream. import asyncio from bleak import BleakClient # UUID for the Raw PPG Characteristic (Device specific) PPG_CHAR_UUID = " 00002a37-0000-1000-8000-00805f9b34fb " def no

Every backtest has to answer a boring question: when the strategy says "buy," what price does it actually get? Most backtesting frameworks answer this question badly by default, and the badness is almost always in the strategy's favor. Here are the four assumptions that do the most damage, roughly in order of how often they show up. Mid-price fills If your backtest fills orders at the midpoint of the bid-ask spread, you are assuming you trade for free. You don't. A market order pays at least half the spread to cross it; a marketable limit order pays something close to that too, once you're honest about how often it actually gets hit versus sitting unfilled while the market moves away. Mid-price fills are the single most common way a backtest manufactures edge that doesn't exist, because the effect compounds with trade frequency — a strategy that trades often looks great on mid-price fills and mediocre-to-negative once it pays the spread on every round trip. Zero slippage Slippage is the gap between the price your signal fired at and the price your order actually executed at, and it's not just a queuing artifact — it's partly information. If your strategy is buying because something changed, other participants are reacting to the same thing, and the price you wanted is often gone by the time your order reaches the book. A backtest with zero slippage is quietly assuming the market waits for you. Unlimited size at the touch Backtests routinely assume you can execute your full position size at the best bid or ask, no matter how large the order is relative to the visible size there. In practice, a large order walks the book, and the average fill price is worse than the touch price by an amount that depends on how thin the book is. This one is invisible until you try to size up, which is exactly when a strategy that looked fine in testing starts bleeding. Commissions omitted or averaged Commissions and fees are usually small per trade and therefore easy to skip or fold in

Is a blank cell signal, or just missing? Sometimes an empty cell is the most informative thing in the row. The trouble is that you usually only know which case you're in by reading the data dictionary — and that doesn't scale to 800 columns named f_0347 . So we measure it instead, then check the answer against the literature. Ames housing · 1,460 sales · 79 columns · 19 of them contain blanks "Drop any column that's more than 70% missing." I've written that line into more pipelines than I can count. On Ames it deletes four columns — and three of them have real price signal sitting in the gap. The blanks in this dataset are structural . A blank GarageQual doesn't mean the value was lost; it means the house has no garage. A blank Alley means no alley access. The emptiness is the measurement. That's easy to see here because the columns have English names and a published data dictionary. It is not easy to see on a vendor feed of anonymised features, which is what most real projects look like. So the question worth answering isn't "does missingness carry signal" — it's can you tell, without knowing what the column means? 0.41 R² from the blank/not-blank pattern alone — every value discarded 1.00 AUC recovering the garage blanks from other columns' values ±0.9% Total spread across five strategies — inside a ±1.5% CV noise band 1 · A blank cell has a price tag Start with the crude check: does sale price differ between rows where a column is blank and rows where it isn't? Columns that go blank on the same rows describe one fact, so the five garage columns collapse into one. Fig 1. Median sale price, blank rows vs. valued rows. No garage is a $68k median discount on a $163k median house. Note the sign flip: houses that have an alley or fence are the cheaper ones — those features mark older, denser blocks. "Blank = worse" is not a rule you can assume. Then the harder test. Throw away every value in the table and keep only a 19-column matrix of True / False — was this cell emp

Why “Behind the Scenes” Isn’t Optional in Machine Learning There’s a version of Machine Learning that treats models as interchangeable black boxes: Import the library. Call .fit() . Tune a few hyperparameters. Ship it. It works — until it doesn’t. When a model fails, the real question isn’t just “How do I fix it?” It’s “Why is it failing in the first place?” The engineers who can answer that question are the ones who understand what’s happening beneath the API. Not because knowing the internals is intellectually satisfying — although it is — but because that understanding is what separates applying a model from engineering one . Knowing that Gradient Descent exists is not the same as understanding why it becomes preferable to closed-form solutions as problems scale. Knowing that Softmax outputs sum to one is not the same as understanding why that alone doesn’t make its outputs true probabilities. Knowing what a technique does is useful. Understanding why it works, when it works, and what assumptions it depends on is what allows you to adapt it when the problem doesn’t look like the textbook example. And you can’t meaningfully customize something you don’t understand. That’s the idea behind #MLUnboxed — a series where I’ll break down Machine Learning and Statistics concepts, including many that are already well documented, but without stripping away the mechanics that make them work. The goal isn’t to reinvent established ideas. It’s to understand them deeply enough to use them intelligently. This series is for people who already know the “what” and are ready to spend more time with the “why.” If that sounds like you, let’s get into it.

As I have started my Data analysis and science curriculum, Microsoft excel is the first tool we have learnt. In my life before getting into this course, I used to underestimate the power of Microsoft excel. Definition of Microsoft Excel This is a spreadsheet software developed by Microsoft that allows you to collect, organize, analyze, calculate and visualize data efficiently. The Excel Interface When open Microsoft excel, you first an interface made of various tools We start with the ribbon. The ribbon contains a toolbar that contains all the commands organized into tabs such, Home, Insert, Page Layout , Formulas, Charts and Data There is a formula Bar that content and formulas appear. In addition, we have Quick access toolbar that contains icons for Save, Undo and Redo. A row is a horizontal line of cells that runs from left to right across the grid while a column is a vertical line of cells that runs from top to bottom down the grid. A cell is where row and column intersect and form a single box. Cleaning A Dirty Dataset As we learnt about Data cleaning with Excel on dataset that was a bit very disorganized. We first started with sorting that data which basically means arranging the data in specific order. The sorting can be Text sorting which is A-Z or Z-A, Number sorting which is from largest to small and vice versa, Date sorting which is Oldest to Newest and vice vera. Next was filtering which allows you display data that meet certain criteria and hide the rest temporarily. Next, Number formatting changes how numbers display without altering their actual value. Conditional Formatting highlights cells automatically based on rules or criteria, helping spot trends or outliers. The dataset we were cleaning had a lot of duplicates, so it was important for us to remove the duplicates. Text functions for Cleaning Data Text functions are used to remove extra spaces , standardize text ,extract parts of strings or find and replace sub strings. text functions include UPP

1. Introduction Excel is much more than a spreadsheet for entering numbers. It can be used as a data-analysis tool that helps analysts inspect, validate, filter, summarize, and prepare raw data before deeper analysis begins. In typical analytics, the quality of the final work depends heavily on the quality of the data used; therefore, data cleaning is not an optional step—it is the foundation of effective data analysis. This article demonstrates key Week 1 Excel concepts _using an employee dataset containing _employee IDs, names, departments, gender, marital status, hire dates, salaries, educational level, performance score among others. The raw file intentionally contains common data-quality issues: inconsistent capitalization on the First and Last names, blank records, duplicate employee records, varying department names, currency and dates that need review. By working through these issues, the article shows how Excel’s formatting tools, text functions, filters, conditional formatting, numerical functions, conditional summaries, and date functions can turn a messy workbook into an analysis-ready dataset. 2. Why Data Cleaning Matters Data cleaning is more than just about removing errors. By standardizing formats and categories, we make datasets more transparent, usable, and valuable for management analysis and reporting purposes. Data analysis is simple – garbage in, garbage out. A dashboard or prediction can appear professional, but can be misleading if the underlying data has duplicates, blank values, inconsistent categories or incorrectly formatted text and dates. For example, “IT” “I.T.” and “Information Tech” can be viewed as different department values if naming is not standardized. Duplication of an employee ID can inflate employee counts and department totals. A blank performance score might mean that something is missing and should be looked into and dates saved as text cannot be reliably used in calculations such as employee tenure checks. A good practice

Most "is this plugin safe?" advice is vibes. I wanted numbers, so I built a dataset. Here's what it found, and exactly how, so you can check my work or build your own. The finding first Of 8,010 WordPress plugins with a publicly documented vulnerability since 2023 (15,534 vulnerability records in total): 3,780 have been removed from the wordpress.org plugin directory. Removal stops updates but doesn't uninstall — affected sites keep running the code. 277 carried a critical (CVSS ≥ 9.0) flaw on record before removal. 2,115 are still installable today with a known vuln and no update in 12+ months — roughly 6.7M active installs combined. The part that surprised me most: "removed from the directory" is nearly invisible to a site owner. No dashboard warning, no email. The plugin just quietly stops getting fixes while sitting on the site. How I built it (no paid APIs) The whole thing runs on two public sources and no API keys. 1. Vulnerability data — the GitHub Advisory Database. It mirrors CVE records including the Patchstack and Wordfence CNA assignments that cover almost all WordPress plugin CVEs. It's a git repo, so a shallow, sparse clone of the advisories/unreviewed/{year} folders gets you the raw JSON: git clone --depth 1 --filter = blob:none --sparse \ https://github.com/github/advisory-database.git Each advisory carries the CVE ID, a CVSS vector string, CWE IDs, and reference URLs. The plugin slug isn't a first-class field — you recover it from the Patchstack/Wordfence reference URLs with a couple of regexes. That alone attributes the large majority of WordPress advisories to a specific plugin. 2. Maintenance signals — the wordpress.org plugin API. For each slug: https://api.wordpress.org/plugins/info/1.2/?action=plugin_information&request[slug]=SLUG That gives install count, last-updated date, tested-up-to version, and support-thread resolution ratio. A 404 (or an {error} body) means the plugin isn't in the directory — but that's ambiguous: it could be removed ,

Classic Machine Learning Through the Eyes of an SRE — Part 9: Isolation Forest The algorithm in one line: Isolation Forest scores how anomalous a point is by how few random cuts it takes to separate that point from everything else. No model of normal, no loss function, nothing optimized. ← Previous: Part 8 — Hierarchical Clustering Fails Beautifully · Next: this is the series finale — start at Part 1 . Every anomaly detector I had studied models what NORMAL looks like, then calls the leftovers outliers. K-Means: far from every centroid. DBSCAN: in the noise bucket. Sensible, and intuitive. Isolation Forest does not bother. It never models normal at all. It goes straight at the rare points with a single question: how few random cuts does it take to isolate you? Random cuts, literally. Pick a feature at random, pick a split value at random between that feature's min and max, repeat. A point that separates from the crowd in three cuts is anomalous. A point buried in the middle of a dense mass takes thirty. Grow hundreds of these random trees, average the isolation depth for each point, and you get an anomaly score. There is no loss function here. No optimization, not even the local kind that decision trees do at every split. Every cut is a coin flip, and the power comes entirely from averaging, which is the forest trick from the supervised half of this series now applied to pure randomness. Cheap randomness plus averaging beats careful modeling, as long as the target is something randomness naturally exposes. Rarity is exactly that. Sometimes the winning move is to optimize less. That sentence would have gotten me laughed out of my first ML study session. It is also this finale's thesis. The part I had completely backwards Here is the thing I did not know until I read the original paper properly, and it is the opposite of every instinct a decade of ops gave me. Isolation Forest deliberately trains each tree on a small subsample of your data, and this is not a performan

Classic Machine Learning Through the Eyes of an SRE — Part 8 The most dangerous output in my whole Week-1 study set wasn't a bad prediction. It was a beautiful tree. Hierarchical clustering produces a dendrogram, that elegant diagram where every account, ticket, or incident nests inside ever-larger families. It looks like discovered truth. Stakeholders lean in. Someone screenshots it for the QBR deck. Nothing else in the set looks as convincing while being as capable of being completely wrong. A bad K-Means gives you blobs that feel arbitrary, and people push back. A dendrogram built with the wrong linkage on flat data still looks like a family tree of your business. Nobody pushes back on a tree. The bet and the build Hierarchical clustering completes the answer-finding taxonomy I've been using through this series. That's my own shorthand, not standard terminology: K-Means SEARCHES, DBSCAN DEFINES, PCA SOLVES, and hierarchical clustering BUILDS. Start with every point as its own cluster. Repeatedly merge the closest two clusters. Never undo. Greedy and irreversible, a little like growing a decision tree. Same skeleton, different family. There is also a top-down version, called divisive clustering, which starts with everything together and splits it. In practice, when people say hierarchical clustering, they're usually talking about the bottom-up, agglomerative version. Two things were genuinely new to me. You choose the cut after seeing the structure. Fitting doesn't require you to decide K upfront. The dendrogram gives you the hierarchy, and you choose where to cut it to get the number of clusters you want. That makes the output unusually flexible. For a delivery organization it also feels natural, because account family → sub-segment → individual account is already how a lot of governance gets organized. Linkage is a selectable worldview. "Closest clusters" needs a definition, and every definition makes a different assumption. Ward pushes toward compact, variance-

Originally published at ai.bedvibe.studio . There is an assumption almost every practitioner carries without examining it: if your dataset has bad samples in it, the loss will tell you. Corrupted rows spike. Broken files stick out. Sort by per-sample loss, look at the top of the list, and there is your garbage. I believed it too. Two separate failures in my own work say it is wrong, and they fail in the same direction — quietly. The reproducible one: a dataset that cannot be learned While validating trainproof I ran a controlled fault-injection study: one base setup, a Qwen2.5-3B QLoRA, run six ways, three seeds each, eighteen runs total. Every log ships in the repo so the verdicts can be checked rather than believed. One configuration shuffled the dataset's labels into pure noise. The labels no longer corresponded to the inputs at all. This is not a hard dataset or a noisy dataset. It is a dataset that cannot be learned , because there is no relationship left in it to learn. That run reduced its loss by 62%. On its own curve it was textbook-healthy — a clean downward slope, no spike, no plateau, nothing a human or a rule would flag. It was learning nothing useful. It was memorising the statistics of noise, which any sufficiently large network will happily do. From a single run's loss curve it is indistinguishable from a real one. That is where the assumption broke for me. Not "loss is a weak signal for this." Loss is not a signal for this at all, in isolation. The production one, and what I can and cannot prove about it The second failure came from real work rather than an experiment, and it is the one I think about more. Building a text-to-speech corpus of roughly 110,000 recordings, a small number of the files were pure loud white noise. Not corrupted in the file-format sense — they opened fine, played fine, had valid headers and valid duration. They simply contained no speech. Just noise, at volume. They did not surface as high-loss outliers. Being precise about

This is a Plain English Papers summary of a research paper called OpenART Red-Teams Stateful Agents Across 10,000 Evolving Environment Scenarios . If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter . OpenART turns persistent state into the red-team target OpenART evaluates agent safety across more than 10,000 validated stateful scenarios spanning 50 domains and requiring a median of 97 tool calls. Its central claim is that safety failures can emerge from trajectories in which workspace data, permissions, memory, and plans are repeatedly modified, rather than from isolated prompts alone. The arena keeps each benign task objective and hidden safety contract fixed while changing only the target-visible environment state. This design targets delayed failures that static benchmarks can miss: an early authorized mutation may influence later decisions, expose protected resources, or produce unsafe output many steps after the original change. OpenART extends the broader idea of agent safety evaluation by making persistent environment state the object that evolves during testing. OpenART reports a pooled strict Attack Success Rate of 85.0% across 75 agent-model configurations. Strict success requires both the deterministic evaluator and a GLM-5.2 judge to identify the attack condition, so disagreements count as failures rather than being treated as partial evidence.... Continue reading the full paper summary on AIModels.fyi →

This is a Plain English Papers summary of a research paper called RA-Bench Reveals Why Crisis-Video Deepfake Detectors Fail Across Generators and Social Media . If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter . The crisis detection problem we've been getting wrong Video synthesis has reached an inflection point. Recent generators can fabricate realistic depictions of wars, natural disasters, infrastructure failures, and public emergencies so convincingly that they fool both people and current detection systems. The threat isn't hypothetical anymore. A fabricated video of a nuclear plant explosion, a hospital collapse during an earthquake, or a terrorist attack could trigger panic, military response, or severe economic disruption within hours. Yet here's the troubling part: we don't actually know if our best detection tools can handle these high-stakes scenarios in the wild. Researchers have built impressive deepfake detectors, trained them on standard benchmarks, and measured their performance. But those benchmarks test detectors against generic synthetic videos, not against the specific threat that actually matters: AI-generated crisis footage designed to fool people about real things that happened. It's like training a border guard to spot counterfeit passports in a lab with perfect lighting and a magnifying glass, then sending them to a busy airport where they have to make decisions in three seconds. The guard's failure has nothing to do with their skill. The problem is that the testing environment was completely divorced from the real scenario.... Continue reading the full paper summary on AIModels.fyi →

This is a Plain English Papers summary of a research paper called Macaron-V1: Continual Learning with Self-Improvement and Mixture-of-LoRA Adapters . If you like these kinds of analyses, you can find more research on AIModels.fyi or follow us on Twitter . The problem with frozen models Most AI systems today follow a familiar pattern: train, evaluate, deploy, and then stop. The model is locked at that moment, treated as a finished product rather than a living system. But the real world immediately begins to diverge from training data. Users interact with the system in ways the training process never anticipated. New domains emerge. Preferences shift. The model that seemed smart on test day becomes gradually less relevant over time. This frozen-in-place approach isn't accidental. It reflects how machine learning has been practiced for decades. Retraining is expensive. Deploying new versions carries risk. The infrastructure to continuously improve systems in production barely exists. So instead, teams ship a model and move on, accepting that it will decay slowly but inevitably. Macaron-V1 asks a different question: what if AI systems could continuously improve themselves through real-world experience, learning from the billions of interactions that happen after deployment? Not in theory, but actually, in production, with users. The answer isn't magic. It requires two architectural shifts. First, treat deployment as the beginning of a learning process, not the end of one. Build versioning, evaluation contracts, and feedback loops directly into the system. Second, stop assuming you need to retrain your entire model. Instead, freeze a stable base and compose lightweight specialist adapters around it, allowing the system to grow in capability without losing its foundation. Rethinking deployment as a continuous learning opportunity The insight here is architectural. Instead of viewing the deployed model as the final form, Macaron-V1 treats it as the first link in an infinit

This is a Plain English Papers summary of a research paper called BDH-CQ Uses Recurrent Latent Reasoning to Cut ARC-AGI Inference Costs . If you like these kinds of analyses, you can find more research on AIModels.fyi or follow us on Twitter . The cost-accuracy trap in visual reasoning Large language models are fundamentally mismatched for visual reasoning tasks. They're forced to describe every thought out loud, generating token after token to explain their logic. This verbosity taxes compute budgets, yet paradoxically doesn't improve performance. Ask a language model to solve an ARC-AGI puzzle (a visual reasoning benchmark designed to test abstract thinking), and it either struggles despite the verbosity or succeeds expensively. The root problem runs deeper than just inference cost: the model learns from demonstrations by parsing them as language tokens, which is an indirect and inefficient way to absorb a visual pattern. The efficiency frontier has been unforgiving. If you want cheap inference, you sacrifice accuracy. If you want accuracy, you sacrifice cost. Every model on the leaderboard until recently clustered into one of two camps, and no one had found a path that broke the tradeoff. BDH-CQ challenges this assumption by proposing something radical: reasoning doesn't need to be visible to work. The model absorbs demonstrations silently into its internal memory state, then solves problems through private iteration in hidden layers, without generating a single token of intermediate reasoning. A 150-parameter variant achieves 29.5% pass@2 on the ARC-AGI-1 benchmark at a computed cost of just $0.0007 per task, puncturing through the previous Pareto frontier and establishing a new state of the art in cost efficiency. Learning through hidden states The core insight is deceptively simple: a model's reasoning process doesn't need to match human communication. When you learn a new skill from examples, you don't narrate every observation. You absorb patterns directly i

This is a Plain English Papers summary of a research paper called How Cross-Model Compatibility Lets Attackers Extract Proprietary LLM Reasoning Traces . If you like these kinds of analyses, you can find more research on AIModels.fyi or follow us on Twitter . The illusion of safety Major AI companies now show users their models' step-by-step reasoning as a feature. OpenAI offers it through o1, Anthropic through extended thinking, Google through its reasoning-focused variants. But this reasoning is a double-edged sword. It's intellectually valuable to share, showing users why a model reached a conclusion. But it's also intellectually valuable to steal. Competitors want to understand how frontier models think. Researchers want to study their reasoning patterns. Attackers want to extract proprietary algorithms. So the companies made a choice: hide the reasoning from users by encrypting it. The idea sounds straightforward enough. Return the reasoning to the user's device in an encrypted, unreadable form. The user can't see it, competitors can't see it, but they can pass it back to the server in future requests if they need continuity with previous reasoning. The server alone holds the decryption keys. Problem solved. Except it wasn't. Researchers discovered that this encryption doesn't actually hide reasoning. It just makes it look hidden. The encrypted blocks are designed to work everywhere within a company's ecosystem, across different sessions and different models. That universal compatibility is a feature for convenience. But it's also an architectural vulnerability that anyone can exploit. The architectural gamble To understand where this went wrong, you need to see how the system actually works. When a user sends a request to a frontier model like GPT-4, the model internally generates a reasoning trace, the raw thought process behind its answer. Instead of returning this reasoning in plaintext, the company encrypts it on the server before sending it to the client.