hayaliv / SLM ideas

53 ideas for small language models

A map of research and product ideas for small language models (SLMs), grouped into eleven categories. For each idea: what the model learns, which kinds of task it performs, and what the source document says about data.

The data-availability labels and task tags are our own reading of a research idea list. They are interpretations, not guarantees.

Browse them in the interactive SLM book

A. Sales & Communication

• Sales Reply Intelligence SLM: Reply text + campaign context → intent label, objection, next step. Tasks: Classify, Extract, Decide. Data: Enron (~500k emails) has no outcome labels.
• Sales Call Objection & Commitment SLM: Call transcript → objections, conditions, promises, owners. Tasks: Extract, Classify, Summarize. Data: Needs private or partner transcripts.
• Human Negotiation Strategy SLM: Dialogue turns → offer, counteroffer, concession, agreement status. Tasks: Classify, Extract, Decide. Data: Deal or No Deal: 5,808 dialogues; B2B transfer unclear.
• Relationship-Aware Outreach SLM: Prospect context → personalized message, checked against facts. Tasks: Extract, Generate, Verify. Data: Profiles are public; outcome labels are not established.
• Meeting Commitment Auditor SLM: Meeting notes → decisions, owners, deadlines, open items. Tasks: Extract, Classify, Verify. Data: Needs authorized transcripts and completion labels.
• Customer Intent & Escalation SLM: Customer message → intent, urgency, frustration, escalate yes/no. Tasks: Classify, Detect, Decide. Data: Banking77 (13,083 queries, 77 intents), CLINC150.

B. Support & E-commerce

• Support Root-Cause SLM: Ticket text → recurring cause category with quoted evidence. Tasks: Classify, Detect, Extract, Summarize. Data: Banking77 for intents; root causes need own labels.
• Financial Complaint Intelligence SLM: Complaint narrative → topic, severity, owning team. Tasks: Classify, Extract, Decide. Data: CFPB Consumer Complaint Database.
• Return/Refund Policy Decision SLM: Request + order + policy → missing evidence, exception, human-review flag. Tasks: Classify, Extract, Verify, Decide. Data: Real cases are private; synthetic for prototypes.
• Product Review Root-Cause SLM: Review text → defect or feature-gap type, grouped by theme. Tasks: Classify, Detect, Extract, Summarize. Data: Amazon Reviews 2023.
• Supplier Catalog Intelligence SLM: Messy product title → brand, model, specs; record pair → match or not. Tasks: Extract, Match. Data: WDC Products: 11,715 offers, 2,162 entities.
• Voice-of-Customer Defect Mapper: Reviews, complaints, tickets → one linked defect report. Tasks: Match, Detect, Summarize, Generate. Data: Amazon + CFPB exist but are not linked per case.

C. Finance & Procurement

• Invoice Exception Resolution SLM: Invoice + PO + receipt → the discrepancy and why. Tasks: Extract, Match, Detect, Generate. Data: Private documents; synthetic for prototypes.
• Payment Dispute Evidence Auditor: Dispute file → missing evidence, timeline conflicts. Tasks: Extract, Detect, Verify, Decide. Data: No joined dispute-outcome data established.
• Tender Requirement Intelligence SLM: Tender document → criteria, required documents, deadlines. Tasks: Extract, Classify, Summarize. Data: EU TED open data; PDFs need parsing.
• Supplier Quote Comparison SLM: Quotes + requirements → differences in price, unit, scope. Tasks: Extract, Match, Generate. Data: Procurement + WDC partial; real quotes are private.
• Expense Policy Exception SLM: Expense claim + receipt + policy → inconsistency flags. Tasks: Extract, Verify, Detect. Data: No suitable open dataset established.
• Vendor Security Questionnaire Verifier: Questionnaire answer + evidence → supported or not, with citation. Tasks: Retrieve / rank, Extract, Verify, Generate. Data: Public frameworks; company evidence is private.

D. Legal & Compliance

• Contract Playbook Deviation SLM: Clause + playbook → deviation type, with the clause as evidence. Tasks: Extract, Classify, Match, Verify. Data: CUAD: 510 contracts, 13,000+ annotations, 41 clause types.
• Data Protection Obligation SLM: DPA text → retention, subprocessor, breach, deletion obligations. Tasks: Extract, Classify, Verify. Data: CUAD, ContractNLI; expert validation needed.
• Policy Drift & Contradiction SLM: Two policy versions → contradictions and outdated lines. Tasks: Detect, Extract. Data: Public regulations + synthetic version histories.
• Public Tender Compliance Auditor: Bid package + tender → missing documents and declarations. Tasks: Extract, Classify, Verify. Data: EU TED plus tender-document processing.

E. Cybersecurity

• Threat Report-to-ATT&CK SLM: Threat report → ATT&CK technique IDs with quoted evidence. Tasks: Extract, Classify, Match. Data: MITRE ATT&CK STIX data; reports need labeling.
• SOC Alert Triage SLM: Alert group → incident or noise, missing context, analyst brief. Tasks: Classify, Detect, Decide, Summarize. Data: Public network datasets; analyst outcomes hard to get.
• Phishing & Suspicious Email Triage SLM: Email body + headers → suspicious signals and safe next steps. Tasks: Classify, Extract, Generate. Data: Enron is not a modern phishing benchmark.
• Security Evidence Summarization SLM: Incident evidence → timeline mapped to techniques, as a report. Tasks: Extract, Detect, Match, Summarize. Data: ATT&CK data; real incident reports need partners.

F. Software & DevTools

• Bug Triage & Duplicate Issue SLM: Bug report → component, owner, duplicate candidates. Tasks: Classify, Match, Decide. Data: BugHub: Chromium, Mozilla Core, Firefox.
• AI Code Review Finding Validator SLM: AI review finding + diff → real issue or false positive. Tasks: Verify, Classify, Extract. Data: ReviewBench: 219 PRs with golden findings.
• CI/CD Failure Diagnosis SLM: Build log → likely cause class and next diagnostic step. Tasks: Classify, Detect, Generate. Data: No reliable log-to-cause benchmark established.
• AI Agent Tool-Call Validator SLM: Intent + proposed tool call → valid, missing parameter, or policy break. Tasks: Verify, Classify, Extract, Decide. Data: Berkeley Function Calling Leaderboard (BFCL).
• RAG Evidence & Hallucination Auditor SLM: Answer + retrieved passages → each claim supported or not. Tasks: Extract, Classify, Verify. Data: RAGTruth (~18k responses), FEVER (185k claims).
• Repo Issue-to-Fix Planner SLM: Issue + stack trace → likely modules and an investigation plan. Tasks: Retrieve / rank, Match, Decide. Data: BugHub + ReviewBench are partial.

G. Knowledge & Research

• Scientific Claim-to-Evidence SLM: Paper text → claims, evidence, unsupported conclusions. Tasks: Extract, Retrieve / rank, Verify, Summarize. Data: FEVER; science-paper transfer uncertain.
• Peer Reviewer Expertise Matching SLM: Paper + reviewer publications → expertise match score. Tasks: Match, Retrieve / rank. Data: FRONTIER-RevRec research dataset.
• Enterprise Knowledge Contradiction SLM: Wiki pages → conflicting or stale instructions. Tasks: Detect, Retrieve / rank, Extract. Data: Public docs for prototypes; histories are private.
• Research Reviewer Comment Quality SLM: Review comment → actionable, unsupported, or duplicate. Tasks: Classify, Match. Data: Corpora exist; quality labels need validation.

H. HR, Education & Health

• Job Description Skill Normalization SLM: Job post → skills mapped to ESCO taxonomy. Tasks: Extract, Match. Data: ESCO taxonomy, 28 languages.
• Student Misconception Diagnosis SLM: Student answer → misconception type and a hint, not the answer. Tasks: Classify, Generate. Data: MathDial: 2,861 teacher–student dialogues.
• Clinical Documentation Quality SLM: Clinical note → missing info and contradictions for review. Tasks: Extract, Detect, Verify. Data: MIMIC-IV-Note; training and data-use agreement.
• HR Policy Query & Resolution SLM: Employee request → category, policy section, grounded answer. Tasks: Classify, Retrieve / rank, Generate, Decide. Data: Public policies; real tickets are private.

I. Industrial & Logistics

• Maintenance Work-Order Intelligence SLM: Technician report → failure, repair, parts, next check. Tasks: Extract, Classify, Decide. Data: C-MAPSS is sensor data, not text reports.
• Manufacturing Quality Incident SLM: Defect report → failure mode and probable cause cluster. Tasks: Classify, Detect, Summarize. Data: Open text with verified root causes not established.
• Logistics Exception Explanation SLM: Shipment events + notices → delay reason, missing event, escalate?. Tasks: Classify, Extract, Detect, Decide. Data: Event logs partial; carrier notes private.
• Supplier Risk Evidence SLM: Notices + contract records → risk indicators with evidence. Tasks: Extract, Match, Classify, Summarize. Data: Public contract records; some risk indicators.

J. AI Systems & Agents

• Agent Tool-Selection SLM: Task + tool list → which tool, which required parameters. Tasks: Classify, Extract, Decide. Data: BFCL definitions, calls, evaluations.
• Model Output Quality Gate SLM: Big-model output + request → meets spec or not. Tasks: Verify, Classify. Data: Needs your own app-specific success labels.
• RAG Data-Quality Monitor SLM: Retrieved set → duplicate, conflicting, or irrelevant docs. Tasks: Classify, Detect. Data: RAGTruth + own retrieval logs.
• Small-Model Routing SLM: Task → small model, large model, or tool. Tasks: Classify, Decide. Data: Self-collected cost, latency, outcome logs.
• AI Failure Taxonomy SLM: Agent trace → failure type label and explanation. Tasks: Classify, Extract. Data: Needs structured trace logs; BFCL/RAGTruth partial.

K. Public Sector & Multilingual

• Citizen Complaint Routing SLM: Complaint → department, issue type, urgency. Tasks: Classify, Extract, Decide. Data: CFPB is US finance only.
• Multilingual Business Support SLM: Albanian–English message → intent, missing info, next action. Tasks: Classify, Extract, Decide. Data: Banking77 is English only; an Albanian–English set must be built.
• Accessibility Issue Triage SLM: Accessibility bug → WCAG criterion and fix hint. Tasks: Classify, Extract, Match, Decide. Data: ReviewBench has some; a dedicated benchmark may be needed.
• Public Procurement Transparency SLM: Contract record → unusual-pattern flag with reasons. Tasks: Detect, Classify, Match. Data: Global public procurement records.