dechenzhang/valg-ml-theory-agent/research-lit

Research Lit

Use this skill to map the ML-theory literature for one research direction and write LITERATURE_SURVEY.md.

This is the internal literature-survey skill:

• write only LITERATURE_SURVEY.md
• do not invent controller-visible outputs, statuses, or checkpoints
• leave downstream perspective selection to /perspective-select

Contract

Required Input

• $ARGUMENT

Optional Input

• workflow_feedback.md
• current LITERATURE_SURVEY.md draft when rerunning the literature-survey checkpoint
• matching pending rerun feedback from workflow_feedback.md for the literature-survey checkpoint

Required Output

• LITERATURE_SURVEY.md

Execution Default

• default internal execution is non-blocking and completes the survey in one pass

Responsibilities

• map Direct Theory, Foundation Theory, Empirical Practice, and evidence-based gaps
• start with Direct Theory for each active sub-target
• recurse with Component Theory or Testbed Theory on zero-hit scopes (recurse only on zero-hit scopes)
• allow lightweight Foundation Theory when Direct Theory is sparse or technique coverage is thin
• use runtime consultations rather than controller-owned checkpoints
• de-duplicate and classify papers before writing LITERATURE_SURVEY.md

Search Budgets

• LOCAL_PAPER_PATHS = papers/, literature/
• MAX_LOCAL_PAPERS = 20
• LOCAL_PDF_READ_PAGES = 3
• MIN_DIRECT_THEORY_PAPERS = 6
• MAX_DIRECT_THEORY_PAPERS = 12
• MIN_FOUNDATION_PAPERS_PER_OBJECT_OR_TESTBED = 4
• MAX_FOUNDATION_PAPERS_PER_OBJECT_OR_TESTBED = 8
• MIN_EMPIRICAL_PAPERS = 4
• MAX_EMPIRICAL_PAPERS = 8

Use these as targets, not hard failure conditions. Theory budgets apply per active sub-target; foundation budgets apply per recursive component scope or per testbed within one sub-target pass.

Source Priority

Search in this order, filling gaps as needed:

1. Zotero, if configured
2. Obsidian notes, if configured
3. local PDF libraries under papers/ or literature/
4. web search
5. Semantic Scholar, if available
6. DeepXiv or equivalent progressive-reading source, if available

Rules:

• prefer user-local sources first
• skip unavailable optional sources silently
• never fail only because a source is unavailable
• prefer the most informative canonical source when the same paper appears multiple times

Workflow

Step 1: Normalize the research direction

• if LITERATURE_SURVEY.md already exists, read it as the latest draft for this literature-survey-checkpoint rerun
• if workflow_feedback.md contains pending rerun feedback in the Literature Survey section, treat that feedback as binding guidance for rerunning /research-lit
• Apply the Source-Direction Fidelity Contract from ../_shared/checklists/artifact-contracts.md to the effective $ARGUMENT. If $ARGUMENT includes or reflects a RESEARCH_BRIEF.md or equivalent brief, do not assume fixed headings; scan the available prose for explicit or clearly implied constraints on the problem object, required scope, target quantity, theorem type, success or progress standard, acceptable partial progress, forbidden shortcuts, out-of-scope directions, parameter dependence, and evaluation priorities. Treat absent fields as unspecified rather than inventing them.
• read $ARGUMENT and detect:
• whether it has one or multiple primary objects
• whether it already constrains the theorem target
• the initial target families and sub-targets
• whether the current scope is still meaningfully splittable
• treat the initial target families and sub-targets as provisional until the first Direct Theory sweep refines them
• send one runtime consultation update in the format Normalization | Targets: <targets> | Objects: <objects> | Splittable: <yes/no>
• in standalone use, report the detected objects and continue
• when invoked inside the public controller workflow, report the normalization and continue
• if the controller or user later edits the detected objects and reruns the skill, use the edited objects in that rerun

Object-detection rules:

• a primary object is a main analyzed entity named in $ARGUMENT
• coordinated entities such as A and B, A/B, or clearly parallel noun phrases count as multiple objects
• theorem targets such as generalization, optimization, implicit bias, or expressive power do not count as separate objects
• example: theoretical analysis of component A in algorithmic setting B has two primary objects: component A and algorithmic setting B

Step 2: Search theory

Treat each active sub-target as its own search unit.

• start every active sub-target with Direct Theory on the full $ARGUMENT
• after the first Direct Theory sweep, refine the active sub-target set using early retrieval signals:
• repeated theorem types, bound metrics, or claim forms
• recurring canonical settings, algorithm-model pairings, or proof styles
• persistent low-yield or zero-yield patterns after exact-name and alias search
• if early retrieval clusters around a non-listed theorem property within an existing family, activate that family's Other bucket and name the property explicitly; if it clusters around a theorem-target family outside the listed families, activate top-level Other and name it explicitly
• for unconstrained arguments, use the analysis-target taxonomy below as the initial search menu and then drop irrelevant or low-yield passes after refinement
• before dropping a low-yield sub-target after the first sweep, check whether nearby papers use proxy theorem quantities or adjacent terminology that may still qualify under that sub-target
• for constrained arguments, search only the constrained family or sub-target
• search with both object or family terms and sub-target theorem terms; do not rely on object names alone
• for Theoretical Scaling Laws, prefer papers derived from an explicit theoretical setting such as linear regression, kernel regression, or another analyzable model; do not treat purely fitted experimental curves as theory hits

Qualifying theory papers must state the setting, assumptions, theorem goal, and at least one main formal claim that explicitly characterizes the current sub-target. Examples:

• Convergence: rate, guarantee, or stationarity guarantee
• Optimization Landscape: critical points, minima, saddles, or related landscape property
• Excess Risk: explicit excess-risk bound
• Population Risk: population-risk or test-risk theorem that is not better treated as Excess Risk or Theoretical Scaling Laws
• Generalization Gap: generalization-gap, stability, or generalization-error theorem
• Theoretical Scaling Laws: explicit theorem-derived scaling law from an explicit theoretical setting, where the main formal claim is how performance or risk scales with domain-appropriate quantitative categories such as data scale, model or system scale, structural parameters, algorithmic or horizon parameters, regularity or stability parameters, or numerical or approximation scales; fitted experimental scaling curves alone do not qualify
• Optimization / Other: an explicit theorem-level optimization property that is not naturally classified as Convergence or Optimization Landscape
• Generalization / Other: an explicit theorem-level generalization property that is not naturally classified as Excess Risk, Population Risk, Generalization Gap, or Theoretical Scaling Laws
• top-level Other: an explicit theorem-target family that is not naturally classified as Optimization, Generalization, Expressive Power, or Implicit Bias
• Implicit Bias: limiting direction, margin, bias, or selected solution geometry
• Expressive Power: approximation power, expressive complexity, or separation result

A scope is not reasonably exhausted until all of the following have been attempted:

• exact-name search
• alias or family-name search when relevant
• sub-target-theorem search
• one proxy-term or adjacent-terminology check when nearby papers suggest the sub-target may be present under different wording
• seed expansion through references, citations, or closely related papers for 2-3 representative papers when available
• one same-sub-target missed-paper audit before stopping

If a scope finds one or more qualifying theory papers, stay in that scope until budget or reasonable exhaustion. If Direct Theory is nonzero but still sparse, or if technique coverage remains thin after reasonable exhaustion, a lightweight Foundation Theory pass is allowed.

If a scope finds zero qualifying theory papers:

• for a splittable multi-object scope, propose a binary split and recurse under Component Theory (recurse only on zero-hit scopes)
• for a single-object or unsplittable scope, fall back to Testbed Theory

Split proposals are runtime reports, not controller-owned checkpoints:

• send No theoretical papers found | Target Family: <family> | Sub-target: <sub-target> | Proposed split: component A = <...> | component B = <...>
• in standalone use, report the proposed split and continue with the split used in this run
• when invoked inside the public controller workflow, report the split and continue
• if the controller or user later edits the split and reruns the skill, use the edited split in that rerun

Classify theory papers by their actual theorem object and theorem goal after reading them, not by title wording alone. Do not force a paper into the nearest listed sub-target or family when its main formal claim is a nonstandard theorem property; use the appropriate Other bucket and name the property explicitly. When a paper could fit multiple generalization sub-targets, classify it by its main formal quantity and main claim rather than by surface terms such as learning curve, test risk, or asymptotic characterization.

Direct Theory

• scope: the full $ARGUMENT
• purpose: find direct prior theory and direct gaps
• for a single-object input, explicitly analyzing the named object under the current sub-target counts as direct prior work
• budget: MIN_DIRECT_THEORY_PAPERS to MAX_DIRECT_THEORY_PAPERS
• for multi-object inputs, a paper counts as Direct Theory only if it analyzes all named objects in the full current scope; papers covering only a strict subset must not be classified as Direct Theory

Foundation Theory

• use recursive Component Theory or Testbed Theory after zero-hit scopes
• allow lightweight foundation search when Direct Theory is sparse or technique coverage remains thin

Component Theory

• scope: one selected component at a time
• use after zero-hit decomposition, or in a lightweight foundation pass for sparse multi-object direct theory
• if the component is a method family, framework family, or problem family, search aliases, renamed variants, derived methods, and theoretically motivated descendants
• if the component still has zero qualifying papers and can split further, recurse again (recurse only on zero-hit scopes)
• if the component is no longer splittable and still zero-hit, fall back to Testbed Theory
• if a foundation paper still names, depends on, or specializes to an original object or component from $ARGUMENT, it is component-level foundation
• budget: MIN_FOUNDATION_PAPERS_PER_OBJECT_OR_TESTBED to MAX_FOUNDATION_PAPERS_PER_OBJECT_OR_TESTBED

Testbed Theory