H.HARSH SINGH

RESEARCH DIRECTION · ARTIFICIAL INTELLIGENCE

Understanding how
intelligent systems reason.

My central intellectual interest is reasoning-capable AI: systems that can structure difficult problems, deliberate over alternatives, allocate computation, use feedback, and verify their results rather than rely only on immediate answer generation.

AI REASONINGINFERENCE-TIME COMPUTESEARCH & PLANNINGVERIFICATION

From generating an answer
to solving a problem.

I am especially interested in what happens between receiving a difficult problem and producing a reliable answer. A capable reasoning system needs more than fluent output. It may need to identify the structure of a task, decide what information is missing, form intermediate representations, consider more than one candidate path, recognise when it is uncertain, and check whether its proposed solution is valid.

This leads me toward approaches that make computation more deliberate and inspectable. I want to understand how reasoning processes can be organised, when additional inference-time computation is worth its cost, how search and planning can be combined with learned models, and how independent verification can help distinguish a plausible response from a correct result.

I am also interested in the engineering conditions that make such systems useful in practice: bounded resources, reproducible experiments, transparent traces, robust evaluation, and clear separation between model-generated suggestions and independently checked outcomes.

Research positioningThis page describes my research interests and the questions guiding my learning and project work. It does not claim that every direction listed has already been experimentally investigated or that a particular method has been proven effective.

Dimensions of
reasoning AI.

01 / REPRESENTATION

Structured and hierarchical reasoning

How should an intelligent system represent a complex problem before attempting to solve it? I am interested in task decomposition, intermediate states, abstraction, dependency structure, and methods that allow a system to move between high-level plans and lower-level operations. The key challenge is preserving the meaning and constraints of the original task while making the reasoning process manageable.

TASK DECOMPOSITION · HIERARCHIES · INTERMEDIATE REPRESENTATIONS
02 / COMPUTATION

Inference-time compute and adaptive deliberation

Rather than treating every prompt as requiring the same amount of computation, I am interested in systems that adapt their effort to task difficulty and uncertainty. Relevant questions include when to pause and plan, when to generate additional candidates, when to call a tool or verifier, how to allocate a bounded compute budget, and when further reasoning is unlikely to improve the result enough to justify its cost.

TEST-TIME COMPUTE · ADAPTIVE EFFORT · RESOURCE ALLOCATION
03 / SEARCH

Search, planning, and multi-path problem solving

Many difficult tasks admit several possible solution paths. I am interested in how model-guided search can explore alternatives, maintain and compare states, rank candidate continuations, backtrack from weak branches, and decide when to stop. This includes connections between language-model reasoning and established search ideas such as breadth-first search, depth-first search, beam search, and Monte Carlo Tree Search, while recognising that heuristic scores are not guarantees of correctness.

TREE SEARCH · PLANNING · CANDIDATE SELECTION · BACKTRACKING
04 / RELIABILITY

Verification, feedback, and tool-assisted reasoning

A generated explanation can sound convincing while containing a logical, mathematical, factual, or implementation error. I am interested in task-specific verifiers, constraint checking, executable tests, external tools, evidence grounding, and feedback loops that identify and correct invalid intermediate or final results. An important concern is understanding the limits of verifiers themselves and avoiding systems that simply optimise for a weak checker.

VERIFIERS · FEEDBACK LOOPS · TOOL USE · CONSTRAINTS
05 / SCIENCE

Evaluation and experimental methodology

Reasoning systems need evaluation that goes beyond answer presentation or output length. I am interested in carefully specified tasks, meaningful baselines, controlled comparisons, repeated trials where appropriate, ablation studies, robustness checks, uncertainty reporting, and transparent datasets and configurations. Accuracy and task success should be considered alongside consistency, generalisation, latency, token usage, and total cost to an accepted result.

BENCHMARK DESIGN · BASELINES · ABLATIONS · REPRODUCIBILITY
06 / SYSTEMS

Efficient and practical reasoning systems

Reasoning capability must be considered together with the systems that support it. I am interested in the trade-offs among model architecture and size, accelerator memory, inference latency, context length, parallel execution, tool orchestration, and operational cost. This connects AI research to machine learning systems, performance engineering, and scientific computing.

MODEL SYSTEMS · MEMORY · LATENCY · PARALLEL COMPUTATION

Questions that can
be investigated.

Q / 01

Which task characteristics make structured decomposition or multi-path search useful compared with direct generation?

Q / 02

How can a model estimate whether additional inference-time computation is likely to improve a result?

Q / 03

How should intermediate states and candidate paths be represented so that they can be inspected, compared, and verified?

Q / 04

When do external tools and independent verifiers improve reliability, and how can their own failure modes be measured?

Q / 05

How can evaluation distinguish robust problem-solving from prompt sensitivity, memorisation, or longer explanations?

Q / 06

What is the relationship between task success, compute budget, latency, memory use, and total cost?

Rigour from question
to evidence.

01

Ground the question

Understand the relevant literature, define the problem, state assumptions, and identify what existing methods and evidence do—and do not—establish.

02

Operationalise the idea

Translate a research question into a testable system, explicit method, measurable outcomes, and a realistic experimental protocol.

03

Build and instrument

Implement focused prototypes with reproducible configurations, transparent traces, bounded resource use, and clear separation of model output from verification.

04

Evaluate against alternatives

Use appropriate baselines, controls, ablations, task-specific checks, and measurements of quality and cost. Treat negative and inconclusive results as useful evidence.

05

Report with precision

Document data and methods, make limitations visible, preserve provenance, and ensure that each conclusion is no stronger than the supporting evidence.

ENGINEERING THAT SUPPORTS THE INTEREST

Ideas made
inspectable.

My project work explores complementary parts of this direction: research workflows, search and evaluation prototypes, machine-learning visualisation, and tools for reliable AI-assisted engineering.