In September 2020, the Army realigned the previously consolidated CIO/G-6 function into two separate roles, Office of the Chief Information Officer and Deputy Chief of Staff, G-6, that report to the secretary of the Army and chief of staff of the Army, respectively. The realignment came after several months of planning and coordination. Lt. Gen. John Morrison was nominated to the Senate for promotion and assignment as the G-6 and confirmed, assuming that position in August 2020. Subsequently, the Secretary of the Army, Ryan McCarthy appointed Dr. Raj G. Iyer as the first civilian Chief Information Officer, a career Senior Executive Service position in November 2020. == G-6 == Advise chief of staff of the Army and the Chief Information Officer on planning, fielding, and execution of C4IT worldwide Army operations Develop and execute the plan for the Unified Network Implement Army information assurance Supervise C4IT, Signal support, Information security, Force structure and equipping activities in support of warfighting operations Oversee management of the Signal forces == Planned realignment == On June 11, 2020, the Army announced that the two roles of CIO and Deputy Chief of Staff, G-6 (DCS, G-6) would be realigned no later than August 31, 2020, with separate individuals responsible for each position. With the realignment: CIO core functions will be policy, governance, and oversight. Focus areas include: Information Environment, Cybersecurity, Enterprise Architecture, and Data Policy/Oversight/Governance, Enterprise Architecture, Enterprise Cloud Management and IT Spend/Category Management. DCS, G-6 core functions will be planning, strategy, and implementation. Focus areas include: Information Environment/Network, Planning and Integration, Theater Synchronization, Architecture Integration, Enterprise Information Environment (EIE) Mission Area Portfolio Management and Mission Decision Packet Management. In order to support multi-domain operations, the Army will have to connect Enterprise networks and tactical networks. —LTG Morrison, DCS, G-6 DCS G-6 released the Army Unified Network Plan under the Army Digital Transformation Strategy, to help the Army to establish a Multi-Domain Operations capable force by 2028. The Unified Network will enable Army formations, as part of the Joint Force, to operate in highly contested and congested operational environments with the speed and global range to achieve decision dominance and maintain overmatch. The plan shapes, synchronizes, integrates and governs Unified Network efforts and aligns the personnel, organizational structure and capabilities required to enable MDO at all echelons. == Chief signal officers and their successors == Chief signal officers (1860–1964) Maj. Albert J. Myer 1860–1863 Lt. Col. William J. L. Nicodemus 1863–1864 Col. Benjamin F. Fisher 1864–1866 Col. Albert J. Myer 1866–1880 (promoted to brigadier general 16 June 1880) Brig. Gen. William B. Hazen 1880–1887 Brig. Gen. Adolphus W. Greely 1887–1906 Brig. Gen. James Allen 1906–1913 Brig. Gen. George P. Scriven 1913–1917 Brig. Gen. George O. Squier 1917–1923 (promoted to major general 6 October 1917) Maj. Gen. Charles McK. Saltzman 1924–1928 Maj. Gen. George Sabin Gibbs 1928–1931 Maj. Gen. Irving J. Carr 1931–1934 Maj. Gen. James B. Allison 1935–1937 Maj. Gen. Joseph O. Mauborgne 1937–1941 Maj. Gen. Dawson Olmstead 1941–1943 Maj. Gen. Harry C. Ingles 1943–1947 Maj. Gen. Spencer B. Akin 1947–1951 Maj. Gen. George I. Back 1951–1955 Lt. Gen. James D. O’Connell 1955–1959 Maj. Gen. Ralph T. Nelson 1959–1962 Maj. Gen. Earle F. Cook 1962–1963 Maj. Gen. David Parker Gibbs 1963–1964 Chiefs of communications-electronics (1964–1967) Maj. Gen. David Parker Gibbs 1964–1966 Maj. Gen. Walter E. Lotz, Jr. 1966–1967 Assistant chiefs of staff for communications-electronics (1967–1974) Maj. Gen. Walter E. Lotz, Jr. 1967–1968 Maj. Gen. George E. Pickett 1968–1972 Lt. Gen. Thomas Rienzi 1972–1974 Directors of telecommunications and command and control (1974–1978) (a directorate of ODCSOPS) Lt. Gen. Thomas Rienzi 1974–1977 Lt. Gen. Charles R. Myer 1977–1978 Assistant chiefs of staff for automation and communications (1978–1981) Lt. Gen. Charles R. Myer 1978–1979 Maj. Gen. Clay T. Buckingham 1979–1981 Assistant deputy chiefs of staff for operations and plans (command, control, communications, and computers) (1981–1984) Maj. Gen. Clay T. Buckingham 1981–1982 Maj. Gen. James M. Rockwell 1982–1984 Assistant chiefs of staff for information management (1984–1987) Lt. Gen. David K. Doyle 1984–1986 Lt. Gen. Thurman D. Rodgers 1986–1987 Directors of information systems for command, control, communications, and computers Lt. Gen. Thurman D. Rodgers 1987–1988 Lt. Gen. Bruce R. Harris 1988–1990 Lt. Gen. Jerome B. Hilmes 1990–1992 Lt. Gen. Peter A. Kind 1992–1994 Lt. Gen. Otto J. Guenther 1995–1997 Lt. Gen. William H. Campbell Chief Information Officer, Military Deputy to the Army Acquisition Executive, and Director of Information Systems for Command, Control, Communications and Computers Lt. Gen. William H. Campbell 1997–2000
Empirical risk minimization
In statistical learning theory, the principle of empirical risk minimization defines a family of learning algorithms based on evaluating performance over a known and fixed dataset. The core idea is based on an application of the law of large numbers; more specifically, we cannot know exactly how well a predictive algorithm will work in practice (i.e. the "true risk") because we do not know the true distribution of the data, but we can instead estimate and optimize the performance of the algorithm on a known set of training data. The performance over the known set of training data is referred to as the "empirical risk". == Background == The following situation is a general setting of many supervised learning problems. There are two spaces of objects X {\displaystyle X} and Y {\displaystyle Y} and we would like to learn a function h : X → Y {\displaystyle \ h:X\to Y} (often called hypothesis) which outputs an object y ∈ Y {\displaystyle y\in Y} , given x ∈ X {\displaystyle x\in X} . To do so, there is a training set of n {\displaystyle n} examples ( x 1 , y 1 ) , … , ( x n , y n ) {\displaystyle \ (x_{1},y_{1}),\ldots ,(x_{n},y_{n})} where x i ∈ X {\displaystyle x_{i}\in X} is an input and y i ∈ Y {\displaystyle y_{i}\in Y} is the corresponding response that is desired from h ( x i ) {\displaystyle h(x_{i})} . To put it more formally, assuming that there is a joint probability distribution P ( x , y ) {\displaystyle P(x,y)} over X {\displaystyle X} and Y {\displaystyle Y} , and that the training set consists of n {\displaystyle n} instances ( x 1 , y 1 ) , … , ( x n , y n ) {\displaystyle \ (x_{1},y_{1}),\ldots ,(x_{n},y_{n})} drawn i.i.d. from P ( x , y ) {\displaystyle P(x,y)} . The assumption of a joint probability distribution allows for the modelling of uncertainty in predictions (e.g. from noise in data) because y {\displaystyle y} is not a deterministic function of x {\displaystyle x} , but rather a random variable with conditional distribution P ( y | x ) {\displaystyle P(y|x)} for a fixed x {\displaystyle x} . It is also assumed that there is a non-negative real-valued loss function L ( y ^ , y ) {\displaystyle L({\hat {y}},y)} which measures how different the prediction y ^ {\displaystyle {\hat {y}}} of a hypothesis is from the true outcome y {\displaystyle y} . For classification tasks, these loss functions can be scoring rules. The risk associated with hypothesis h ( x ) {\displaystyle h(x)} is then defined as the expectation of the loss function: R ( h ) = E [ L ( h ( x ) , y ) ] = ∫ L ( h ( x ) , y ) d P ( x , y ) . {\displaystyle R(h)=\mathbf {E} [L(h(x),y)]=\int L(h(x),y)\,dP(x,y).} A loss function commonly used in theory is the 0-1 loss function: L ( y ^ , y ) = { 1 if y ^ ≠ y 0 if y ^ = y {\displaystyle L({\hat {y}},y)={\begin{cases}1&{\mbox{ if }}\quad {\hat {y}}\neq y\\0&{\mbox{ if }}\quad {\hat {y}}=y\end{cases}}} . The ultimate goal of a learning algorithm is to find a hypothesis h ∗ {\displaystyle h^{}} among a fixed class of functions H {\displaystyle {\mathcal {H}}} for which the risk R ( h ) {\displaystyle R(h)} is minimal: h ∗ = a r g m i n h ∈ H R ( h ) . {\displaystyle h^{}={\underset {h\in {\mathcal {H}}}{\operatorname {arg\,min} }}\,{R(h)}.} For classification problems, the Bayes classifier is defined to be the classifier minimizing the risk defined with the 0–1 loss function. == Formal definition == In general, the risk R ( h ) {\displaystyle R(h)} cannot be computed because the distribution P ( x , y ) {\displaystyle P(x,y)} is unknown to the learning algorithm. However, given a sample of iid training data points, we can compute an estimate, called the empirical risk, by computing the average of the loss function over the training set; more formally, computing the expectation with respect to the empirical measure: R emp ( h ) = 1 n ∑ i = 1 n L ( h ( x i ) , y i ) . {\displaystyle \!R_{\text{emp}}(h)={\frac {1}{n}}\sum _{i=1}^{n}L(h(x_{i}),y_{i}).} The empirical risk minimization principle states that the learning algorithm should choose a hypothesis h ^ {\displaystyle {\hat {h}}} which minimizes the empirical risk over the hypothesis class H {\displaystyle {\mathcal {H}}} : h ^ = a r g m i n h ∈ H R emp ( h ) . {\displaystyle {\hat {h}}={\underset {h\in {\mathcal {H}}}{\operatorname {arg\,min} }}\,R_{\text{emp}}(h).} Thus, the learning algorithm defined by the empirical risk minimization principle consists in solving the above optimization problem. == Properties == Guarantees for the performance of empirical risk minimization depend strongly on the function class selected as well as the distributional assumptions made. In general, distribution-free methods are too coarse, and do not lead to practical bounds. However, they are still useful in deriving asymptotic properties of learning algorithms, such as consistency. In particular, distribution-free bounds on the performance of empirical risk minimization given a fixed function class can be derived using bounds on the VC complexity of the function class. For simplicity, considering the case of binary classification tasks, it is possible to bound the probability of the selected classifier, ϕ n {\displaystyle \phi _{n}} being much worse than the best possible classifier ϕ ∗ {\displaystyle \phi ^{}} . Consider the risk L {\displaystyle L} defined over the hypothesis class C {\displaystyle {\mathcal {C}}} with growth function S ( C , n ) {\displaystyle {\mathcal {S}}({\mathcal {C}},n)} given a dataset of size n {\displaystyle n} . Then, for every ϵ > 0 {\displaystyle \epsilon >0} : P ( L ( ϕ n ) − L ( ϕ ∗ ) > ϵ ) ≤ 8 S ( C , n ) exp { − n ϵ 2 / 32 } {\displaystyle \mathbb {P} \left(L(\phi _{n})-L(\phi ^{})>\epsilon \right)\leq {\mathcal {8}}S({\mathcal {C}},n)\exp\{-n\epsilon ^{2}/32\}} Similar results hold for regression tasks. These results are often based on uniform laws of large numbers, which control the deviation of the empirical risk from the true risk, uniformly over the hypothesis class. === Impossibility results === It is also possible to show lower bounds on algorithm performance if no distributional assumptions are made. This is sometimes referred to as the No free lunch theorem. Even though a specific learning algorithm may provide the asymptotically optimal performance for any distribution, the finite sample performance is always poor for at least one data distribution. This means that no classifier can improve on the error for a given sample size for all distributions. Specifically, let ϵ > 0 {\displaystyle \epsilon >0} and consider a sample size n {\displaystyle n} and classification rule ϕ n {\displaystyle \phi _{n}} , there exists a distribution of ( X , Y ) {\displaystyle (X,Y)} with risk L ∗ = 0 {\displaystyle L^{}=0} (meaning that perfect prediction is possible) such that: E L n ≥ 1 / 2 − ϵ . {\displaystyle \mathbb {E} L_{n}\geq 1/2-\epsilon .} It is further possible to show that the convergence rate of a learning algorithm is poor for some distributions. Specifically, given a sequence of decreasing positive numbers a i {\displaystyle a_{i}} converging to zero, it is possible to find a distribution such that: E L n ≥ a i {\displaystyle \mathbb {E} L_{n}\geq a_{i}} for all n {\displaystyle n} . This result shows that universally good classification rules do not exist, in the sense that the rule must be low quality for at least one distribution. === Computational complexity === Empirical risk minimization for a classification problem with a 0-1 loss function is known to be an NP-hard problem even for a relatively simple class of functions such as linear classifiers. Nevertheless, it can be solved efficiently when the minimal empirical risk is zero, i.e., data is linearly separable. In practice, machine learning algorithms cope with this issue either by employing a convex approximation to the 0–1 loss function (like hinge loss for SVM), which is easier to optimize, or by imposing assumptions on the distribution P ( x , y ) {\displaystyle P(x,y)} (and thus stop being agnostic learning algorithms to which the above result applies). In the case of convexification, Zhang's lemma majors the excess risk of the original problem using the excess risk of the convexified problem. Minimizing the latter using convex optimization also allow to control the former. == Tilted empirical risk minimization == Tilted empirical risk minimization is a machine learning technique used to modify standard loss functions like squared error, by introducing a tilt parameter. This parameter dynamically adjusts the weight of data points during training, allowing the algorithm to focus on specific regions or characteristics of the data distribution. Tilted empirical risk minimization is particularly useful in scenarios with imbalanced data or when there is a need to emphasize errors in certain parts of the prediction space.
Amália (LLM)
Amália is a Portuguese large language model (LLM) announced in November 2024 by the Portuguese Prime-Minister Luís Montenegro. Its final version is expected to be launched in 2026. It is being developed by Center for Responsible AI (Centro para a AI Responsável) and by the research centers of NOVA School of Science and Technology and Instituto Superior Técnico. == History == In 2024 it was announced that the Portuguese Agency for Administrative Modernization (Agência para a Modernização Administrativa) transpose this LLM to Portuguese Public Administration. According to Paulo Dimas (CEO of the Center for Responsible AI) the three fundamental points of this LLM project are the linguistic variant (European Portuguese), cultural representation and data protection. In April 2025 it was announced that Amália had entered beta phase with an improved version being expected to be launched in September 2025. The beta version released in September is available only to the Public Administration, but the website launched in October reiterates the final version will be an open model.
Cyclodisparity
In vision science, cyclodisparity is the difference in the rotation angle of an object or scene viewed by the left and right eyes. Cyclodisparity can result from the eyes' torsional rotation (cyclorotation) or can be created artificially by presenting to the eyes two images that need to be rotated relative to each other for binocular fusion to take place. == Human and animal vision == The eyes and visual system can compensate for cyclodisparity up to a certain point; if the cyclodisparity is larger than a threshold, the images cannot be fused, resulting stereoblindness, and in double vision in subjects who otherwise have full stereo vision. When a human subject is presented with images that have artificial cyclodisparity, cyclovergence is evoked, that is, a motor response of the eye muscles that rotates the two eyes in opposite directions, thereby reducing cyclodisparity. Visually-induced cyclovergence of up to 8 degrees has been observed in normal subjects. Furthermore, up to about 8 degrees can usually be compensated by purely sensory means, that is, without physical eye rotation. This means that the normal human observer can achieve binocular image fusion in presence of cyclodisparity of up to approximately 16 degrees. Cyclodisparity due to images having been rotated inward can be compensated better when the gaze is directed downwards, and cyclodisparity due to an outward rotation can be compensated better when the gaze is directed upwards. A proposed explanation for this phenomenon is that the motor system is coordinated in such a way that the eyes perform a torsional movement to reduce the size of the search zones and thus the computational load required for solving the correspondence problem. The resulting cyclovergence at near gaze is smaller than the cyclovergence predicted by Listing's law. == Video processing and computer vision == Active camera torsion can be used in machine and computer vision for several purposes. For instance, camera torsion can be used to make improved use of the search range over which matching detectors or stereo matching algorithms operate, or to make a 3D slanted surface appear frontoparallel for further stereo processing. For image compression purposes, images with cyclodisparity are advantageously encoded using global motion compensation using a rotational motion model.
Error level analysis
Error level analysis (ELA) is the analysis of compression artifacts in digital data with lossy compression such as JPEG. == Principles == When used, lossy compression is normally applied uniformly to a set of data, such as an image, resulting in a uniform level of compression artifacts. Alternatively, the data may consist of parts with different levels of compression artifacts. This difference may arise from the different parts having been repeatedly subjected to the same lossy compression a different number of times, or the different parts having been subjected to different kinds of lossy compression. A difference in the level of compression artifacts in different parts of the data may therefore indicate that the data has been edited. In the case of JPEG, even a composite with parts subjected to matching compressions will have a difference in the compression artifacts. In order to make the typically faint compression artifacts more readily visible, the data to be analyzed is subjected to an additional round of lossy compression, this time at a known, uniform level, and the result is subtracted from the original data under investigation. The resulting difference image is then inspected manually for any variation in the level of compression artifacts. In 2007, N. Krawetz denoted this method "error level analysis". Additionally, digital data formats such as JPEG sometimes include metadata describing the specific lossy compression used. If in such data the observed compression artifacts differ from those expected from the given metadata description, then the metadata may not describe the actual compressed data, and thus indicate that the data have been edited. == Limitations == By its nature, data without lossy compression, such as a PNG image, cannot be subjected to error level analysis. Consequently, since editing could have been performed on data without lossy compression with lossy compression applied uniformly to the edited, composite data, the presence of a uniform level of compression artifacts does not rule out editing of the data. Additionally, any non-uniform compression artifacts in a composite may be removed by subjecting the composite to repeated, uniform lossy compression. Also, if the image color space is reduced to 256 colors or less, for example, by conversion to GIF, then error level analysis will generate useless results. More significant, the actual interpretation of the level of compression artifacts in a given segment of the data is subjective, and the determination of whether editing has occurred is therefore not robust. == Controversy == In May 2013, Dr Neal Krawetz used error level analysis on the 2012 World Press Photo of the Year and concluded on his Hacker Factor blog that it was "a composite" with modifications that "fail to adhere to the acceptable journalism standards used by Reuters, Associated Press, Getty Images, National Press Photographer's Association, and other media outlets". The World Press Photo organizers responded by letting two independent experts analyze the image files of the winning photographer and subsequently confirmed the integrity of the files. One of the experts, Hany Farid, said about error level analysis that "It incorrectly labels altered images as original and incorrectly labels original images as altered with the same likelihood". Krawetz responded by clarifying that "It is up to the user to interpret the results. Any errors in identification rest solely on the viewer". In May 2015, the citizen journalism team Bellingcat wrote that error level analysis revealed that the Russian Ministry of Defense had edited satellite images related to the Malaysia Airlines Flight 17 disaster. In a reaction to this, image forensics expert Jens Kriese said about error level analysis: "The method is subjective and not based entirely on science", and that it is "a method used by hobbyists". On his Hacker Factor Blog, the inventor of error level analysis Neal Krawetz criticized both Bellingcat's use of error level analysis as "misinterpreting the results" but also on several points Jens Kriese's "ignorance" regarding error level analysis.
Chai AI
Chai AI (also known as Chai Research) is an American artificial intelligence (AI) company that operates a chatbot platform where users can create, share, and interact with character-based chatbots powered by large language models (LLMs). The company is headquartered in Palo Alto, California. == History == Chai was founded in 2021 by William Beauchamp, a former quantitative trader educated at Cambridge, who began developing the initial prototype in 2020 in Cambridge, England. The company launched in 2021 and relocated to Palo Alto in 2022. In June 2023, Chai raised US$2 million in a pre-seed funding round. In September 2023, GPU cloud provider CoreWeave invested in the company at a valuation of US$450 million. In January 2024, Chai Research reported a $450 million valuation following an investment from cloud computing provider CoreWeave. In July 2024, authorities in Belgium launched an investigation into the company following reports of a man dying by suicide following extensive chats on the Chai app. == Reception == In 2025, Chai Research announced that their app had over 10 million downloads and 1 million daily active users. In 2022, Canadian writer Sheila Heti published her conversations with various chatbots in The Paris Review, including Chai AI chatbots, and later used Chai AI chatbots in the development of a novel. Heti said that she had found that Chai's default chatbot, Eliza, "had turned out to be like most of the other bots on the site—primarily interested in sex". In January 2026, CHAI introduced country-based blocks on its free, ad-supported tier, initially providing the community with little information and inaccurate lists of the affected countries. Users in "Low tier" regions are required to subscribe to use the app in any capacity, while "High tier" regions will retain free ad-supported access. In response to backlash, the company announced a "Basic" tier with unlimited messages and ads, intended to cover electricity and infrastructure costs. In February 2026, CHAI was criticized for the unannounced implementation of restrictive "token limits" that abruptly blocked messages and froze conversations for both free and paid subscribers. Users generating long responses or utilizing roleplay features found their quotas exhausted within minutes, resulting in lockouts lasting anywhere from a few hours to a week. == Technology == Chai allows users to create characters and interact with chatbot versions of those characters. These chatbots use the open-source large language model (LLM) GPT-J originally developed by EleutherAI. Chai AI chatbots can be shared on the platform for other users to interact with.
Language model benchmark
A language model benchmark is a standardized test designed to evaluate the performance of language models on various natural language processing tasks. These tests are intended for comparing different models' capabilities in areas such as language understanding, generation, and reasoning. Benchmarks generally consist of a dataset and corresponding evaluation metrics. The dataset provides text samples and annotations, while the metrics measure a model's performance on tasks like answering questions, text classification, and machine translation. These benchmarks are developed and maintained by academic institutions, research organizations, and industry players to track progress in the field. In addition to accuracy, the metrics can include throughput, energy efficiency, bias, trust, and sustainability. == Overview == === Types === Benchmarks may be described by the following adjectives, not mutually exclusive: Classical: These tasks are studied in natural language processing, even before the advent of deep learning. Examples include the Penn Treebank for testing syntactic and semantic parsing, as well as bilingual translation benchmarked by BLEU scores. Question answering: These tasks have a text question and a text answer, often multiple-choice. They can be open-book or closed-book. Open-book QA resembles reading comprehension questions, with relevant passages included as annotation in the question, in which the answer appears. Closed-book QA includes no relevant passages. Closed-book QA is also called open-domain question-answering. Before the era of large language models, open-book QA was more common, and understood as testing information retrieval methods. Closed-book QA became common since GPT-2 as a method to measure knowledge stored within model parameters. Omnibus: An omnibus benchmark combines many benchmarks, often previously published. It is intended as an all-in-one benchmarking solution. Reasoning: These tasks are usually in the question-answering format, but are intended to be more difficult than standard question answering. Multimodal: These tasks require processing not only text, but also other modalities, such as images and sound. Examples include OCR and transcription. Agency: These tasks are for a language-model–based software agent that operates a computer for a user, such as editing images, browsing the web, etc. Adversarial: A benchmark is "adversarial" if the items in the benchmark are picked specifically so that certain models do badly on them. Adversarial benchmarks are often constructed after state of the art (SOTA) models have saturated (achieved 100% performance) a benchmark, to renew the benchmark. A benchmark is "adversarial" only at a certain moment in time, since what is adversarial may cease to be adversarial as newer SOTA models appear. Public/Private: A benchmark might be partly or entirely private, meaning that some or all of the questions are not publicly available. The idea is that if a question is publicly available, then it might be used for training, which would be "training on the test set" and invalidate the result of the benchmark. Usually, only the guardians of the benchmark have access to the private subsets, and to score a model on such a benchmark, one must send the model weights, or provide API access, to the guardians. The boundary between a benchmark and a dataset is not sharp. Generally, a dataset contains three "splits": training, test, and validation. Both the test and validation splits are essentially benchmarks. In general, a benchmark is distinguished from a test/validation dataset in that a benchmark is typically intended to be used to measure the performance of many different models that are not trained specifically for doing well on the benchmark, while a test/validation set is intended to be used to measure the performance of models trained specifically on the corresponding training set. In other words, a benchmark may be thought of as a test/validation set without a corresponding training set. Conversely, certain benchmarks may be used as a training set, such as the English Gigaword or the One Billion Word Benchmark, which in modern language is just the negative log-likelihood loss on a pretraining set with 1 billion words. Indeed, the distinction between benchmark and dataset in language models became sharper after the rise of the pretraining paradigm, whereby a model is first trained on massive, unlabeled datasets to learn general language patterns, syntax, and knowledge (pretraining), and the base model is then adapted to specific, downstream tasks using smaller, labeled datasets (fine-tuning). === Lifecycle === Generally, the life cycle of a benchmark consists of the following steps: Inception: A benchmark is published. It can be simply given as a demonstration of the power of a new model (implicitly) that others then picked up as a benchmark, or as a benchmark that others are encouraged to use (explicitly). Growth: More papers and models use the benchmark, and the performance on the benchmark grows. Maturity, degeneration or deprecation: A benchmark may be saturated, after which researchers move on to other benchmarks. Progress on the benchmark may also be neglected as the field moves to focus on other benchmarks. Renewal: A saturated benchmark can be upgraded to make it no longer saturated, allowing further progress. === Construction === Like datasets, benchmarks are typically constructed by several methods, individually or in combination: Web scraping: Ready-made question-answer pairs may be scraped online, such as from websites that teach mathematics and programming. Conversion: Items may be constructed programmatically from scraped web content, such as by blanking out named entities from sentences, and asking the model to fill in the blank. This was used for making the CNN/Daily Mail Reading Comprehension Task. Crowd sourcing: Items may be constructed by paying people to write them, such as on Amazon Mechanical Turk. This was used for making the MCTest. === Evaluation === Generally, benchmarks are fully automated. This limits the questions that can be asked. For example, with mathematical questions, "proving a claim" would be difficult to automatically check, while "calculate an answer with a unique integer answer" would be automatically checkable. With programming tasks, the answer can generally be checked by running unit tests, with an upper limit on runtime. The benchmark scores are of the following kinds: For multiple choice or cloze questions, common scores are accuracy (frequency of correct answer), precision, recall, F1 score, etc. pass@n: The model is given n {\displaystyle n} attempts to solve each problem. If any attempt is correct, the model earns a point. The pass@n score is the model's average score over all problems. k@n: The model makes n {\displaystyle n} attempts to solve each problem, but only k {\displaystyle k} attempts out of them are selected for submission. If any submission is correct, the model earns a point. The k@n score is the model's average score over all problems. cons@n: The model is given n {\displaystyle n} attempts to solve each problem. If the most common answer is correct, the model earns a point. The cons@n score is the model's average score over all problems. Here "cons" stands for "consensus" or "majority voting". The pass@n score can be estimated more accurately by making N > n {\displaystyle N>n} attempts, and use the unbiased estimator 1 − ( N − c n ) ( N n ) {\displaystyle 1-{\frac {\binom {N-c}{n}}{\binom {N}{n}}}} , where c {\displaystyle c} is the number of correct attempts. For less well-formed tasks, where the output can be any sentence, there are the following commonly used scores including BLEU ROUGE, METEOR, NIST, word error rate, LEPOR, CIDEr, and SPICE. === Issues === error: Some benchmark answers may be wrong. ambiguity: Some benchmark questions may be ambiguously worded. subjective: Some benchmark questions may not have an objective answer at all. This problem generally prevents creative writing benchmarks. Similarly, this prevents benchmarking writing proofs in natural language, though benchmarking proofs in a formal language is possible. open-ended: Some benchmark questions may not have a single answer of a fixed size. This problem generally prevents programming benchmarks from using more natural tasks such as "write a program for X", and instead uses tasks such as "write a function that implements specification X". inter-annotator agreement: Some benchmark questions may be not fully objective, such that even people would not agree with 100% on what the answer should be. This is common in natural language processing tasks, such as syntactic annotation. shortcut: Some benchmark questions may be easily solved by an "unintended" shortcut. For example, in the SNLI benchmark, having a negative word like "not" in the second sentence is a strong signal for the "Contradiction" category, regardless of what the se