More about HKUST
Inference-Time Learning for LLMs: A Survey of In-Context, Retrieval-Augmented, and Test-Time Training
PhD Qualifying Examination
Title: "Inference-Time Learning for LLMs: A Survey of In-Context,
Retrieval-Augmented, and Test-Time Training"
by
Miss Jinghan ZHANG
Abstract:
When deployed in real-world settings, language models often face target data
that differ from their pretraining distribution - whether due to emerging
knowledge, domain-specific styles or noisy inputs. Fine-tuning may be
infeasible when compute is constrained, model weights are inaccessible, or
only unlabeled test streams are available. To address this, we study
inference-time learning: adapting model behavior during prediction to bridge
the distribution gap and enhance reasoning under shift.
We organize inference-time learning along three axes: the learning signal
(e.g., self-supervision, uncertainty-aware objectives, or verifiable
feedback); the form of parameterization (ranging from contextual
conditioning to lightweight adapters or external memory); and the temporal
scale (single-shot adaptation, sliding-window updates, or online continual
learning). Building on this taxonomy, we categorize existing algorithmic
families - including in-context learning, retrieval-augmented reasoning with
tool use, and test-time training and efficient finetuning - and analyze
their trade-offs in stability, plasticity and compute efficiency.
To enable systematic evaluation, we systematize existing protocols and
benchmark suites, clarifying distinctions between online and offline
settings and among contextual input modalities (instructions, retrieved
documents, programmatic scratchpads), while tagging learning content as
knowledge, procedure, or calibration. We further discuss failure modes
under distributional and semantic drift and outline metrics for performance,
calibration, forgetting, regret, and explicit compute cost. Finally, we
outline open challenges for safe long-horizon experience accumulation and
robust deployment.
Date: Wednesday, 12 November 2025
Time: 10:00am - 12:00noon
Venue: Room 5501
Lifts 25/26
Committee Members: Dr. Junxian He (Supervisor)
Dr. Yangqiu Song (Chairperson)
Dr. May Fung