Prior work, positioning and novelty

Why it matters

"How is this different from X?" is one of the most predictable questions in a research conversation and one of the most damaging to answer badly. Positioning is also how you avoid spending months rediscovering a known result. For a thesis or internship, placing a proposed idea next to the lab's own papers shows that you can work independently.

Core ideas

Search is a procedure. Start from two or three seed papers. Chase references backwards and citing papers forwards (Semantic Scholar and Google Scholar list citing papers; dblp lists venue proceedings and complete author records). Scan recent proceedings of the venues where the seeds appeared and the reading lists of graduate courses in the area. Keshav's note describes a compact version of this loop. Stop when new candidates mostly cite papers you already have, and record queries and dates so you can repeat the search before submitting anything.

Positioning table. For the five to ten closest works, write one row each: problem, assumptions (failure model, hardware, workload, data, threat model), mechanism, evaluation setup and stated limitations. Columns expose the gap better than prose. Often the opening is an assumption that has stopped holding: a workload became burstier, memory became cheap relative to network bandwidth, a model no longer fits on one accelerator, or an agent's tool calls now have side effects.

Kinds of novelty. A contribution can be a new problem formulation, a new mechanism, a new measurement or characterization, a design point that changes a known tradeoff, a negative result, or an explanation that unifies observations. Each is legitimate, and each needs different evidence.

Novelty versus useful engineering. Re-implementing a known design in a new framework, integrating components or scaling a pipeline can be valuable work, and much research engineering is exactly that. It becomes research when it produces knowledge that transfers: why something works, a tradeoff nobody had quantified, or a measurement that changes what others should build. A useful test is to ask what a reader who never runs your code would learn. If the honest answer is "that it can be done with enough effort", call it engineering, and say so plainly in interviews. Levin and Redell's review criteria (original ideas, reality, lessons, choices, context) remain a sharp checklist.

Contribution statements. Write each as a claim plus its evidence: "we show that ..., measured on ... (Section N)". A contribution without planned evidence is a hope.

Related work, written fairly. Organize by idea, not paper by paper. Describe each line of work in terms its authors would accept, state the assumption or regime where it falls short for your problem, and acknowledge where it is better than your approach.

Common misconceptions

• "Nobody has done this; I searched." One query is not a search. After a documented procedure, the most you can claim is "to our knowledge", and that phrase should mean something.
• "Novelty means a completely new idea." Many strong papers apply known techniques in a new regime or explain known behaviour; the novelty is the precise delta.
• "More engineering effort makes the work more novel." Effort is not a contribution; insight and evidence are.
• "Related work is a list of systems to beat." It shows that you understand the design space; dismissing prior work without engaging its assumptions costs credibility.
• "Concurrent work makes my project worthless." Overlap is common. Position honestly and emphasize differences in assumptions, setting or evidence.

How to approach it

Take one hypothesis from the research-questions topic and spend a bounded session (about three hours) searching from two seed papers. Build the positioning table, write three contribution statements and one sentence on why the closest work does not already answer your question. Before a thesis or internship interview, read the target lab's most recent relevant papers and prepare a two-sentence answer to "how does your idea relate to our work?". Interviewers at every level check that you can name the closest work and its assumptions; senior candidates are also expected to judge which direction is worth pursuing and why now.