EdiShorts

UPSC Prelims 2026 GS Paper 1 Q67 · Science & Technology

Year2026
PaperGS Paper 1
QuestionQ67
SubjectScience & Technology

LLMs in machine learning - which are correct?
1. Assign probabilities to next words & pick highest
2. Process data through mathematical optimization to minimise prediction errors
3. Produce unbiased outputs
Select the correct answer using the code given below:

  1. a1 only
  2. b1 and 2 only
  3. c2 and 3 only
  4. d1, 2 and 3
Show answer & explanationHide answer
Correct answer(b) 1 and 2 only

LLMs predict text by generating probability distributions over words and are optimized via mathematical algorithms like gradient descent, but they inevitably absorb and reproduce biases present in their training datasets.

Statement-by-statement analysis

✓ Statement 1
  • LLMs operate auto-regressively by calculating probability distributions across a vocabulary to predict the most likely next word or token given a context sequence.
  • Sampling techniques such as greedy decoding or top-k sampling pick the highest or most appropriate probability tokens to generate coherent text.
✓ Statement 2
  • During the training phase, models use mathematical optimization algorithms such as stochastic gradient descent and Adam.
  • These algorithms continuously update internal model weights to minimize a loss function representing prediction error.
✗ Statement 3
  • LLMs are trained on massive, uncurated web datasets that inherently contain societal, cultural, and demographic biases.
  • Consequently, LLMs are known to reflect and often amplify these training data biases rather than producing completely unbiased outputs.

Concept to know

  • Large Language Models (LLMs) are deep learning architectures primarily based on the Transformer model structure.
  • The core task of a basic LLM is conditional probability estimation of sequential text tokens.
  • Training involves iterative forward and backward passes (backpropagation) to adjust parameters and reduce prediction loss.
  • Data bias is a structural issue in machine learning where models inherit real-world historical inequalities present in training data.
  • Alignment techniques like Reinforcement Learning from Human Feedback (RLHF) aim to mitigate harmful outputs, but complete un-biasing remains unsolved.

Related facts for revision

  • Transformer Architecture: Neural network design introduced in 2017 utilizing self-attention mechanisms to process sequential data in parallel.
  • AI Hallucination: A phenomenon where an LLM generates authoritative and confident statements that are factually incorrect or unsupported.
  • Parameters vs Tokens: Parameters refer to the internal weights learned by the model, whereas tokens are the basic chunks of text processed as inputs/outputs.
  • Generative vs Discriminative AI: Generative AI creates new data instances based on learned patterns, whereas Discriminative AI classifies existing data into labels.
How UPSC tests this

Understanding of fundamental operational principles of artificial intelligence models alongside their recognized ethical and technical constraints.

Trap: Assuming that computer models produce perfectly neutral and unbiased outputs simply because they rely on mathematical operations.

More Science & Technology PYQs

All Science & Technology questions →

Answer from the official UPSC answer key. Explanation by the EdiShorts editorial team.