Language Model

Large language models are essentially stateless functions whose weights are fixed at inference time and do not learn over time.

The Agent Harness does not replace the reasoning capability of the core model, but instead builds a runtime management shell around the model, responsible for cross-cutting concerns such as tool invocation, state management, and error recovery.

Current mainstream large language models exhibit significant deficiencies in both confidence calibration and selective abstention, tend to be overconfident, and when faced with questions outside the training distribution, tend to produce fluent but incorrect answers.

Large language models cannot structurally distinguish between user instructions and instructions embedded in content, because both are part of the same token sequence from the model's perspective.

Reward model training can use learning-to-rank frameworks such as the Bradley-Terry model, training a neural network to predict the probability of human preference between any two responses, taking a (prompt, response) pair as input and outputting a scalar score.

The tool-calling mechanism allows large language models to dynamically invoke external functions during inference, representing a fundamental leap from language models to action models.

A language model itself can only output text and cannot directly perform operations such as opening files or pressing buttons.

This article is compiled from automatically generated content; the original text was drafted by a language model and checked automatically, with no human author credited.

The larger the parameter count, the stronger a model generally is in language understanding and generation quality, but the memory and compute requirements also increase proportionally.

The same model performs vastly differently across Agent products, and the gap lies mainly not in the model but in the scaffolding built around it (loop control, tool integration, etc.).

Simulated control evaluations (SCEs) use language models to simulate tool responses in place of a real execution environment, allowing control evaluations to be flexibly constructed for specific deployment scenarios.

The leap in language model capabilities is obvious to all, but general-purpose robots that can stably complete household, warehousing and assembly tasks remain at the laboratory and demo-video stage.

The core idea of using language models for protein sequence generation is to treat amino acid sequences as text composed of 20 letters and perform autoregressive or masked pretraining on large-scale protein sequence databases.

The model outputs the most "likely" continuation token by token by probability rather than verifying whether a package actually exists in the npm registry, so package names, function signatures, and API version numbers are high-hallucination areas.

In small-model applications, everything that must be exact (such as deadlines, amounts, reference numbers, issuing institutions) should be handled by deterministic code, and everything that needs to be understood (such as explaining the letter) should be left to the model.

When a language model generates text, it predicts only the next token at each step, and a token may be a whole word, a word fragment, punctuation, or a word with a leading space.

Current language models mainly rely on offline training and maximum likelihood estimation, lacking theoretical guarantees for online continual learning behavior.

In BLACKBOX, the language model is responsible for choosing tools, interpreting graph query results, synthesizing evidence, and generating fault explanations, but the graph always remains the source of operations facts.

Deterministic orchestration demotes the model to a stateless function call that receives only the current sub-problem and returns structured output, holding no cross-round memory.

The three failure modes of dropped steps, narration instead of execution, and empty turns appear in exactly the same way on large and small models, and the common factor is the agentic loop itself rather than model scale.