EPFO EO/AO 2023 General Ability Test · Question 116 of 120

Correct — D, (d) 1, 2, 3 and 4. All four listed uses are real applications of large language models, so the option that accepts every item is the answer. A large language model is a neural network, almost always built on the transformer architecture, trained on very large quantities of text to predict what comes next in a sequence. Everything such a model does follows from that one capability: given text, produce more text. It can therefore be pointed at any task that can be expressed as text in and text out, and it is adapted to particular tasks either by prompting, which is simply describing the task in the input, or by further training on task-specific data. Item 1, reimagined search engines. Conversational interfaces that answer a question in prose, rather than returning a list of links, are built by putting a language model in front of a search index. This has been the most publicly visible use of the technology and the one that made it a news story. Item 2, health care. Language models are used to draft and summarise clinical documentation, to condense long patient records, to search and synthesise medical literature, and to answer patient-facing questions in plain language. The applications are real, and so is the standing requirement that a clinician remains responsible for any output that affects care. Item 3, software development. Code is text with a strict grammar, which makes it particularly well suited to these models. They are used to complete and generate code, to explain unfamiliar code, to write tests and to help diagnose errors. Item 4, translating a language. Machine translation is one of the oldest tasks in the field, and the transformer architecture that underlies today's language models was introduced in the context of translation research. In India the National Language Translation Mission and its Bhashini platform apply this class of technology to translation across Indian languages. Since every item names a text task, and the model's single capability is to transform text, none of the four can be excluded. The structure of the option set confirms the reading: with '1, 2, 3 and 4' on offer, the only question worth asking is whether any listed use is impossible, and none is. A card on this subject should also say what these models do not do, because the same topic is easily set as a negative question. They generate fluent text that can be confidently wrong, a failure usually called hallucination; they reflect biases present in their training data; they have no inherent access to information published after their training unless a retrieval system supplies it; and they do not reason from a database of verified facts.

• (a)1 and 3 only — This accepts the two most technology-facing uses, search and software development, and rejects health care and translation. Neither rejection stands. Translation is among the tasks the underlying architecture was designed for, and it is the use with the clearest policy footprint in India through the National Language Translation Mission. Health care applications are equally real, covering clinical documentation, summarisation of records and literature synthesis, even though they are used under professional supervision. A candidate lands here by treating the model as a tool for computing tasks only, when its actual domain is language, and language is present in every profession.
• (b)1, 2 and 4 only — This excludes software development, which is one of the best-established uses of the technology and among the earliest to be deployed commercially. The exclusion probably comes from thinking of a language model as something that works on natural language alone, and of programming as a separate, more formal activity. In fact source code is exactly the kind of material these models handle well, because it is text with a strict and consistent grammar and because very large quantities of it are publicly available to train on. Code completion, code generation, code explanation and test writing are all standard applications.
• (c)2, 3 and 4 only — This excludes the reimagined search engine, which is the most publicly visible application of all and the one that put the technology into general news coverage. A candidate might exclude it on the reasoning that search is a matter of indexing and ranking documents rather than of generating text. That describes traditional search; the reimagined version puts a language model in front of the index so that the system answers in prose and can be questioned conversationally, which is precisely the change the item refers to. Rejecting the item requires denying a use that is in daily view.

A large language model is a machine learning system trained on very large text corpora to predict the next unit of text in a sequence. The architecture almost universally used is the transformer, whose key mechanism, attention, allows the model to weigh the relevance of every part of the input when producing each part of the output, and which made it practical to train on far larger datasets than earlier sequence models allowed. Training proceeds in stages: a pre-training stage on broad text, which produces general language competence, followed by fine-tuning and alignment stages on narrower data, which shape the model to follow instructions and to behave acceptably. Because the trained model's single operation is to transform text into text, its range of applications is set by how many tasks can be cast in that form, and the answer turns out to be very many: summarisation, drafting, question answering, translation, classification, extraction of structured data from documents, and code generation. The limitations follow from the same design. The model produces the most plausible continuation rather than a verified one, so it can generate fluent falsehoods, a failure conventionally called hallucination; it carries the biases of its training data; it knows nothing published after its training data was collected unless a retrieval system supplies the material; and its outputs are probabilistic, so the same prompt need not give the same answer twice. In India the policy conversation around this technology runs through the IndiaAI Mission and, for language specifically, through the National Language Translation Mission and its Bhashini platform.

The current-affairs block of the EO/AO General Ability Test tracks technology stories that have crossed into general news, and large language models were the dominant such story when this paper was set. Questions of this kind are not testing technical depth; they are testing whether the candidate understands what the technology is for well enough to judge a list of proposed uses. The habit rewarded is to learn any new technology by its single core capability and then to reason from that capability to its applications, rather than memorising a list of applications, because the list changes faster than the capability does.

• A large language model is a neural network trained on very large text corpora to predict the next unit of text.
• The transformer architecture, built around the attention mechanism, is what makes these models practical at scale.
• Because the core operation is text in and text out, any task expressible in that form is a candidate application.
• Established uses include conversational search, drafting and summarising clinical documentation, code generation and explanation, and machine translation.
• The models are adapted to tasks by prompting or by further training on task-specific data.
• Known limitations are hallucination, bias inherited from training data, no inherent knowledge of events after training, and non-deterministic outputs.
• In India, the Bhashini platform under the National Language Translation Mission applies this class of technology to Indian-language translation.
• A model's output is a plausible continuation, not a verified fact, which is why professional supervision is required in high-stakes uses.

• Treating a language model as a search engine that retrieves stored facts; it generates text rather than looking up records.
• Assuming these models handle only natural language and not source code.
• Excluding translation because it seems too old a task; it is among the tasks the architecture was built for.
• Confusing hallucination, which is fluent invention, with a simple factual gap in the training data.
• Rejecting an item in a list because the application seems futuristic, when the question asks only whether the use is possible.

Technology questions in EO/AO papers come as a list-and-code item like this one, as a 'best describes the term' item on a piece of vocabulary in the news, or as a two-statement item pairing a capability with a limitation. Because the field moves quickly, the safest preparation is conceptual: learn what the technology does at its root, learn the standard limitations, and learn the Indian policy instrument attached to it, since the examiner frequently connects an international technology story to a domestic mission or statute.

No directly related past PYQ was found.

• practice — not a real PYQ

  Which one of the following best describes a large language model ?

  • (a)A database that stores verified facts and retrieves them on request
  • (b)A neural network trained on very large text corpora to predict the next unit of text
  • (c)A special-purpose supercomputer used only for weather forecasting
  • (d)An encryption standard used to secure electronic communication

  Answer(b) A neural network trained on very large text corpora to predict the next unit of text

• practice — not a real PYQ

  Bhashini, sometimes mentioned in news, is associated with :

  • (a)Translation and language technology for Indian languages
  • (b)Regulation of digital lending applications
  • (c)A national database of land records
  • (d)A unified platform for direct benefit transfers

  Answer(a) Translation and language technology for Indian languages