This is a fuller, reusable version of your request. The guidance that follows answers it while keeping the proposed projects separate from completed research.
Copy and adapt this prompt
I am Rohan, an ML engineer at Altimetrik. I graduated with a BTech from SRM in 2019 and have worked in IT since 2019. I want to design, execute, write, and submit my first AI research paper as an independent solo author. I do not currently have an academic supervisor or university research affiliation. Help me choose a narrow, credible project that I can finish alongside my job.
Act as a careful research mentor, literature reviewer, experimental-design partner, and scientific writing assistant. You may help draft text and code, but I will steer the research, run and check experiments, make the scientific decisions, and take responsibility for the final paper. Do not present yourself as a coauthor. Treat all proposed findings as hypotheses until supported by actual results.
First, research primary papers released between 4 May and 4 October 2026. Give me ten distinct, realistic paper ideas inspired by this work. Separate each proposed paper title from the actual source-paper title. For each idea, cite the source URL and its verified first-publication or submission date; distinguish a new paper from a recent revision of older work. Do not claim that a gap is novel merely because you have not seen it. Explain what still needs to be checked in the literature.
For every idea, provide one or two substantial paragraphs explaining the question, motivation, possible contribution, method, expected paper contents, and what would count as a useful negative result. Add a compact experiment plan with public or appropriately licensed data, practical baselines, evaluation metrics, ablations, leakage controls, rough compute and storage needs, feasibility risks, and a plausible publication route. Rank the best three for a first solo paper and explain the tradeoffs.
My reported hardware is a Ryzen 7 5800H with eight cores and sixteen threads and 16 GB RAM. The latest hardware output confirms an NVIDIA RTX 3060 Mobile/Max-Q GPU with driver 610.57.04, but does not show VRAM. The 6 GB VRAM figure comes from earlier context; verify current usable VRAM and framework access before sizing experiments. The root and home directories share a 186.25 GiB filesystem with 175.79 GiB used, leaving roughly 10.46 GiB by subtraction. An external 4.55 TiB drive is reported, but its available space and suitability need checking. Design a CPU-first pilot, minimize downloads, and use small models, inference, evaluation, or parameter-efficient work only when appropriate. Separate measured requirements from estimates.
Explain the difference between an arXiv preprint, a peer-reviewed workshop paper, a conference paper, and a journal article. Compare realistic routes, including no-author-fee options, using current official sources. Distinguish submission fees, article-processing charges, mandatory registration, travel, and optional open-access charges. Explain submission steps, peer review, revisions, typical planning uncertainty, arXiv endorsement, independent-researcher affiliation, employer permissions, conflicts, data licensing, responsible AI use, and common first-author mistakes. Do not guarantee acceptance or assume a venue is suitable because it accepts submissions.
Finish with a practical eight-to-twelve-week plan, explicit phase gates, a defensible novelty test, a suggested paper structure, an experiment log and reproducibility checklist, a budget strategy, and the smallest useful first experiment. Label timelines and budgets as planning estimates. Use authoritative citations for changing rules and fees. Never fabricate results, citations, code execution, acceptance rates, or benchmark improvements. Deliver a readable, self-contained HTML guide with embedded styling and no external dependencies; do not publish a website.
02 · The practical answer
A credible first paper is within reach
Start with a small, carefully controlled empirical study. Your strongest initial advantage is engineering discipline: reproducible systems, honest measurement, useful error analysis, and a clear explanation of when a method works.
A first contribution can identify a failure condition, isolate a confound, measure a meaningful cost–quality tradeoff, or improve a benchmark. It must teach readers something beyond applying a popular method to another dataset. A modest result with strong controls is enough to investigate seriously.
Yes, you can submit as a solo independent researcher without a current university affiliation. A venue evaluates the manuscript under its own rules. A supervisor can provide useful feedback, but you do not need to acquire a university coauthor just to submit.
My default recommendation: choose one question, two small datasets or task families, a small set of fair baselines, and one primary outcome. Finish a pilot before committing to a model, venue, or expensive run. Keep the first paper’s scope small enough that you can personally audit the code and every central claim.
All titles below are proposals. They are not claims of original discovery, completed experiments, or publication readiness. Recent inspiration is a starting point; a broader prior-art search remains necessary.
03 · Feasibility before ambition
Design around the machine you can verify
Your Ryzen 7 5800H and 16 GB RAM are sufficient for a CPU-first evaluation pipeline, small classical baselines, dataset analysis, and lightweight local inference. Your latest hardware output confirms the NVIDIA RTX 3060 Mobile/Max-Q GPU and driver 610.57.04. It does not report VRAM; the 6 GB figure comes from earlier context. Verify current usable VRAM and framework access before choosing model size, precision, batch size, or context length. A “6 GB GPU” is not a guarantee that a particular six-gigabyte model will run: runtime buffers and attention caches also need memory.
The immediate constraint is storage. Root and home share the same 186.25 GiB filesystem, with 175.79 GiB used. The arithmetic difference is approximately 10.46 GiB, and actual available space may be lower because of reservations and filesystem details. Do not count root and home as separate pools. Verify the external drive’s free space, filesystem, mount, write access, and reliability before using it for datasets and caches. A 4.55 TiB capacity is not a measurement of free space.
CPU-first pilot: begin with a few hundred examples, cached text or embeddings, and a model that comfortably fits verified memory. Stream or subset data rather than cloning every benchmark.
Keep assets separate: place approved model and dataset caches on verified roomy storage; keep code, manifests, and final result tables small. Maintain a backup of irreplaceable experiment records.
Measure the first run: record peak RAM, peak VRAM if applicable, tokens per second, disk growth, elapsed time, and failures. Extrapolate from those measurements instead of guessing from parameter count.
Spend only after the pilot: use a fixed cap for optional cloud compute or APIs. Preserve raw responses where licensing permits so a rerun does not silently double costs.
Budget convention for the ideas: “local-first” means zero additional service spend if your current machine can finish the pilot; it still uses your electricity and time. The suggested optional budget bands are planning caps, not current provider quotes. Obtain an actual estimate before paid work. Avoid fine-tuning in the first experiment unless training itself is essential to the question.
04 · Ten candidate papers
Recent inspiration with a small experimental core
Every principal inspiration below was first submitted within 4 May–4 October 2026. These are proposed titles, not existing publications. The sources motivate the questions; their reported findings have not been independently reproduced here. Recency indicates research activity, not proven popularity. The stated gaps remain hypotheses to test against the full literature.
The sampled work points to active directions in small-model efficiency, agent and memory reliability, calibration, retrieval, and test-time adaptation. These are useful themes for choosing a project, not a measured ranking of the field’s most popular trends.
Ten titles, ranked by catchiness
Subjective title appeal only. This order reflects memorability, clarity, and curiosity—not scientific merit, novelty, feasibility, or likelihood of acceptance. All ten are proposed studies; the titles do not report experimental outcomes. The separate research feasibility top three is unchanged.
Memory with a Past: Updating Agent Facts Without Losing History
Agent memory should answer both “what is true now?” and “what was true before?” Test whether explicit validity intervals and source pointers reduce stale answers without erasing useful history, compared with timestamp-only replacement. The contribution would be a paired current-versus-historical evaluation covering corrections, temporary exceptions, and separate events. Existing memory work already evaluates updates, so merely adding timestamps would be an inadequate novelty claim.
The paper would introduce a small, transparently synthetic extension to the cleaned LongMemEval benchmark, compare memory policies, and analyze conflicting histories. Use a fixed small reader with raw-history retrieval, flat summaries, timestamp replacement, and versioned memory at matched token budgets. Keep answer keys out of memory construction and split entities and templates before generating histories. Manually audit ambiguous cases. A useful negative result would show that elaborate versioning adds cost without improving both present and historical questions. The enormous full-history LongMemEval-V2 setting is unnecessary.
Tiny pilot: 30 short histories with current, historical, and unanswerable questions.
Measure and ablate: current/historical accuracy, stale facts, false overwrites, abstention and memory tokens; remove intervals or source pointers.
Cost and route: local-first; optional USD 50–150 cap for independent judgments. Memory or agent-evaluation workshop.
Title appeal #02 · Low-cost language-model analysis
Echo or Answer? Detecting Unwanted Prompt Repetition
Repeating a prompt can be an unwanted failure, but copying is correct in quoting and extraction tasks. Test whether a task-aware detector distinguishes these cases better than a simple overlap threshold or repetition penalty. The candidate contribution is a contrastive evaluation set and an early-stop-and-retry rule that accounts for legitimate copying. Prompt overlap is not automatically evidence of training-data leakage, and prompt echoing itself is already studied.
The paper would pair benign extraction, quotation, and summarization prompts with diagnostic echo cases, then compare prefix matching, shared-span matching, repetition penalties, and a lightweight rule or logistic detector. Test compact models without modifying attention heads or fine-tuning. Separate naturally generated failures from constructed diagnostics, split by template family, and manually label a blinded sample. Analyze suppressed correct quotes as seriously as detected echoes. A balanced negative finding may reveal that detection costs more task accuracy than it saves; an echo-enriched dataset cannot estimate deployment prevalence.
At a fixed bit width, does a calibration corpus balanced across prose, extraction, and numeric fields preserve output meaning better than generic text? Test whether improvements remain when JSON syntax is already constrained. The contribution would be a semantic-versus-syntactic quantization audit and a small calibration-selection recipe. Domain calibration and structured generation predate this proposal; the useful result must establish their interaction and generalization on held-out tasks.
The paper would compare a floating-point reference, simple rounding, generic-calibration GPTQ, balanced-calibration GPTQ, and PiSO if compatible code is available. Keep bit width, group size, tokenizer, and inference backend fixed. Pair a small JSONSchemaBench syntax subset with extraction tasks that have exact field and numeric answers. Cross constrained decoding with calibration conditions instead of equating valid JSON with success. Separate calibration templates and values from test data, repeat calibration samples, and distinguish fake-quantization quality from actual deployment speed.
Tiny pilot: one 0.36B–0.6B model, two 128-example calibration sets, 100 held-out tasks; no custom kernels.
Measure and ablate: schema validity, slot accuracy, numeric correctness, perplexity, memory and latency; vary mixture and calibration tokens.
Cost and route: local-first if toolchain fits; optional USD 50–150 cap. Efficient-NLP, compression, or ML-systems workshop.
Title appeal #04 · Higher scientific risk
Check or Retry? Reasoning on a Fixed Budget
When is checking one intermediate claim better than generating another answer with the same compute? Test whether selective verification helps mainly when claims have inexpensive independent checks, while a weak verifier reinforces errors elsewhere. The candidate contribution is a task-stratified allocation policy and failure-boundary study. The recent inspiration already reports matched-budget claim checking, so it must be a direct baseline where implementable; using a smaller model alone is insufficient novelty.
The manuscript would compare one-shot answers, self-consistency, whole-answer critique, claim-level verification, and selective checking on arithmetic or constraint problems with executable answers. Distinguish programmatic checks from LLM-only judgments so a gain is attributed correctly. Count extraction and verification tokens, hold out templates, freeze thresholds, and report false rejection of correct answers. Include a self-verification condition. If the smallest models lack basic competence, verification comparisons may be uninformative; the pilot must detect that before an expensive sweep.
Tiny pilot: 50 questions and two budget allocations; scale only if the generator can solve a meaningful fraction.
Measure and ablate: exact answers, wrong-to-right corrections, false rejection, tokens and latency; vary verifier and claim-selection rule.
Cost and route: local-first; optional USD 50–200 cap for a stronger reference verifier. Reasoning or efficient-inference workshop.
Title appeal #05 · Strong applied ML fit
Can Calibration Backfire with Few Labels?
Good classification accuracy does not guarantee trustworthy probabilities. Ask how many calibration labels are needed before post-hoc calibration helps rather than overfits. Test whether temperature scaling conservatively shrunk toward no adjustment is more reliable than flexible calibration with only 20–100 labels. The contribution would be a sample-size, imbalance, and model-family regime map, including failure conditions. Calibration itself is established; renaming a regularizer is not enough.
Build the paper around repeated, controlled comparisons across small OpenML datasets. Compare an accessible pinned TabPFN version with CatBoost, logistic regression, and random forests; then compare no calibration, temperature scaling, Platt scaling, and isotonic regression. Keep calibration data separate from training and final testing. Aggregate evidence across datasets, not as if every row were an independent experiment. Include unstable calibrators and negative results. The recent TabFM paper adds motivation, but the experiment must not depend on weights whose availability or license has not been checked.
Tiny pilot: three datasets, two calibration-set sizes, mostly CPU; scale toward 15–20 small datasets only after timing.
Measure and ablate: log loss and Brier score first; ECE and AUROC second; vary label count, imbalance, and regularization.
Cost and route: local-first; optional USD 25–100 cap. Tabular/AutoML workshop; TMLR only after its AI-policy compatibility check.
Title appeal #06 · CPU-friendly protocol study
Late Labels, Shifting Forecasts: Testing Online Adaptation
Online forecasters cannot update from ground truth before it arrives. Ask whether delayed or bursty feedback changes the ranking of frozen, recalibrated, and adapting models. Test whether conservative updates outperform frequent updates when observations are stale. The contribution would be an event-time evaluation harness and a reproducible boundary map. Existing work already uses matured ground truth; simply warning about temporal leakage is not a new contribution.
The paper would compare seasonal-naive or ETS forecasts, a lightweight linear baseline, and a compact frozen forecaster with residual correction and suitable adaptation methods. Evaluate immediate, delayed, and missing-feedback schedules using a strict predict–reveal–update log. Exclude future normalization statistics and unmatured residuals. Use rolling-origin evaluation and block-level uncertainty, with synthetic shifts for controlled evidence. Public time series may overlap pretraining; do not claim that exposure is impossible. Interval results need dependence-aware caveats rather than distribution-free guarantees under arbitrary drift.
Tiny pilot: three series and two reveal schedules, mostly CPU; no large-model training.
Measure and ablate: MAE/MASE, recovery after shift and update cost; vary delay, missingness, drift, and calibration-window length.
Cost and route: local-first; optional USD 25–100 cap. Forecasting or applied-ML workshop.
Title appeal #07 · Real deployment costs
The Cost of a Second Model: Measuring Small-Model Routing
Does a small-model ensemble still help when only one model fits comfortably in memory and models execute sequentially? Test whether an early-exit cascade improves the accuracy–wall-clock tradeoff after charging for model loading, routing, and every generated token. The contribution would be a constrained-deployment evaluation that separates cached candidate selection from actual online execution. Adaptive routing is mature research; adding a router alone is not novel.
The paper would compare the strongest single model, fixed voting, self-consistency, a fixed cascade, and a validated early-exit policy on extractable-answer questions. Use small sequential models and a mixture of benchmark and fresh arithmetic tasks. Match token budgets and separately match measured wall-clock budgets. Tune thresholds on development tasks, then freeze them. Report cold and warm starts, correlated errors, and domains where routing is a net loss. Do not claim real-device speedups from cached results or broad hardware generality from one laptop.
Measure and ablate: accuracy, p50/p95 latency, loading time and memory; vary model diversity, confidence calibration, and cache state.
Cost and route: local-first; optional USD 25–100 cap. Efficient NLP or ML-systems workshop.
Title appeal #08 · Strong engineering portfolio value
Errors That Guide: Testing Feedback Transfer in Database Agents
Small language models often fail at tool boundaries even when the underlying question is answerable. Ask whether compact feedback identifying a violated field, type, or constraint helps recovery across unseen database schemas and tool wrappers. The useful contribution would be transfer across interfaces at matched retry and token budgets. “Clearer errors help” is already known, and repairing an ordinary server bug would not establish this research claim.
The manuscript would describe an isolated read-only benchmark, a taxonomy of failures, and a controlled recovery study. Use SQLite or DuckDB with held-out Spider schemas and answerable SELECT tasks. Compare plain errors, verbose replay, typed compact feedback, and typed feedback with a valid schema example. Keep contracts correct and include a no-injected-error condition. Split by database and wrapper family, verify query results directly, and report when extra detail increases repeated failures. Use only local benchmark data, read-only connections, and timeouts.
Tiny pilot: 30 tasks, one compact instruct model, two feedback formats; no production systems.
Measure and ablate: result accuracy, recovery, repeated-invalid-call rate, tokens and latency; vary error type, verbosity, and wrapper.
Cost and route: local-first; optional USD 50–150 cap for a hosted comparison. Agent-evaluation or AI software-engineering workshop.
Title appeal #09 · Best low-cost starting point
Growing Corpora, Stale Corrections: Testing Multilingual Retrieval
Multilingual embedding indexes can over-retrieve a few “hub” items. Ask whether a correction that helps a static index stays useful as the corpus grows or becomes language-imbalanced. Test whether refreshing a small, balanced reference reservoir preserves the benefit of full similarity-statistic recomputation at lower cost. The possible contribution is a corpus-growth benchmark and an accuracy–update-cost comparison; neither CSLS nor reservoir sampling would be presented as new.
The paper would explain the growth protocol, compare corrected and uncorrected retrieval, and map when stale statistics help or hurt. Use aligned FLORES sentences, unrelated distractors, and two small multilingual encoders. Compare cosine similarity, fresh CSLS, stale CSLS, and reservoir-refreshed CSLS. Split by source sentence and keep calibration queries separate from future test queries. Report language-specific failure cases and paired source-level uncertainty. A null result is useful if it shows when refreshing is unnecessary; sentence retrieval alone does not establish a RAG improvement.
Tiny pilot: 200 aligned query pairs, two encoders, two growth stages; frozen embeddings and no training.
Measure and ablate: Recall@1/10, worst-language recall, hub concentration, update time and RAM; vary reservoir size and refresh frequency.
Cost and route: local-first; optional USD 25–75 cap. Multilingual retrieval workshop, or a stronger ACL-family short paper after gap validation.
Title appeal #10 · More implementation-sensitive
How Many Views Are Enough? Calibrating Vision at Test Time
Extra augmented views can improve a vision model’s prediction but consume time and memory. Test whether stopping when views agree preserves calibration while reducing work on easy images. The contribution would be a compute–calibration tradeoff study and a conservative sequential stopping rule. The recent paper already proposes calibration-aware test-time prompt tuning, so that objective is a comparison point rather than a method to rename.
The paper would compare no adaptation, fixed view budgets, calibration-aware prompt tuning, and an adaptive view budget using a compact CLIP-sized model and small image benchmarks with corruption tests. Keep optimization settings matched and process views serially where needed. Tune on separate images and corruption types, reset adaptation between samples, and cluster uncertainty by original image. Include accuracy, uncertainty, and abstention behavior. Fewer views can damage both correctness and calibration; a credible negative conclusion needs more than one corruption or model setting.
Tiny pilot: 200 images; verify memory and runtime before expanding or enabling gradient-based adaptation.
Measure and ablate: accuracy, NLL, Brier score, ECE, risk–coverage, views and latency; vary stopping threshold and corruption severity.
Cost and route: local-first if hardware fits; optional USD 50–150 cap. Efficient-vision or robustness workshop.
All optional USD bands above are suggested spending caps, not vendor quotes. Pilot locally where feasible; larger runs remain conditional on verified hardware, storage, licenses, runtime, and a worthwhile gap. Older datasets and implementation links support reproducibility and are not counted as recent-paper discoveries.
05 · My recommendation
Research feasibility: the top three
This is the separate recommendation for a first solo project, based on cost, measurement clarity, and engineering fit. It is independent of the subjective title ranking above.
Growing Corpora, Stale Corrections: Testing Multilingual Retrieval
Best first choice for low cost and clear measurement. Frozen embeddings and exact alignment labels avoid generation and training. The challenge is establishing a useful gap beyond static hubness correction.
Best fit for your applied ML experience. Strong controls and dataset-level analysis matter more than GPU power. The contribution must be a useful regime map, not another calibration variant.
Errors That Guide: Testing Feedback Transfer in Database Agents
Best engineering showcase. Tool contracts and answer correctness are directly testable. It takes more implementation work, and cross-interface transfer must go beyond repairing one harness.
Novelty is a relationship to the literature, not an adjective in the title. Search for the precise intervention, failure mode, evaluation protocol, and adjacent terminology. Read the closest papers’ methods and limitations, not only abstracts. A literature summary alone is usually a weak fit for an original-research track; a systematic survey needs its own contribution and a venue that accepts it. Track the closest papers’ questions, datasets, models, controls, findings, limitations, and the difference your study would add.
Write one falsifiable sentence. For example: under a fixed context and latency budget, a specified retrieval policy improves a specified outcome on specified tasks. Avoid “better,” “robust,” and “efficient” without defining the measure.
Identify the closest three to five comparisons. Include a simple baseline and the most relevant existing approach. If you cannot explain why each baseline is fair, the experiment is not ready.
Remove easy confounds. Match token budgets, examples, prompts, corpus versions, model settings, and hardware where they affect the comparison. If a method receives extra information or compute, report it explicitly.
Freeze the evaluation. Use a development set to choose settings; hold out the test set. Keep near duplicates, template variants, and related source documents from leaking across splits. Record all exclusions before examining outcomes.
Estimate uncertainty. Choose a paired analysis when methods answer the same examples. Report confidence intervals and meaningful effect sizes. Bootstrap at the appropriate independent unit, such as document or task, rather than treating correlated variants as independent samples.
Look for alternative explanations. Ablate the ingredient you claim matters. Inspect failed examples and report when the effect disappears. Do not generalize from one model or one dataset to all LLMs.
A replication must resolve uncertainty or reveal a boundary condition. A negative result needs a design capable of detecting a meaningful effect. If a pilot merely reproduces an established finding, keep it as a learning artifact and refine the question.
Novelty gate: proceed only when you can state “Closest work establishes X; our study tests Y under Z; this matters because…” and back each part with sources. If the difference is just a new model name, expand the conceptual contribution or narrow the claim.
07 · Where and how to publish
Choose a route with eyes open
A preprint, workshop paper, conference paper, and journal article serve different purposes. A low-cost route is a complete manuscript on arXiv, where eligible, plus submission to a suitable fee-free peer-reviewed journal. The preprint and the peer-reviewed publication are separate statuses. Check the journal’s policies before posting, and choose the destination before relying on a particular AI-assisted writing or idea-generation workflow.
arXiv for a citable preprint
arXiv is free dissemination, not peer review or a journal acceptance. Submit a complete scholarly manuscript; moderation may decline or reclassify it. An unaffiliated registrant may use “Independent” as the organization. First-time submission in some categories requires endorsement. The current process does not guarantee access just because you have an institutional email: qualifying prior coauthored work or a personal endorser may be relevant. Follow the actual prompt in your account, and seek an appropriate qualified endorser with your real draft if required. Endorsement is neither coauthorship nor a quality certification.
TMLR has rolling submission and no author fees. Its criteria emphasize well-supported claims and findings that interest the ML community; a new architecture or a state-of-the-art score is not mandatory. That makes it worth investigating for a rigorous empirical project, not an easy acceptance. Surveys have been excluded since 1 September 2026. Its roughly nine-week review aim is not a guarantee; most submissions and reviews become public, including rejected or withdrawn papers.
Important for your planned AI workflow: TMLR’s current FAQ expects ideas, claims, and results to be human-sourced and requires a first-page footnote for ChatGPT-style writing assistance. Because this guide proposes AI-assisted research ideas, do not assume that choosing one and approving a draft automatically satisfies that requirement. Read the full policy and clarify compatibility with editors before adopting this route. Do not misrepresent the provenance of a hypothesis. JMLR also has no publication fees, but expects substantial ML contributions and is a demanding fit; simply applying ML to a different domain is outside its stated scope, and unsolicited surveys are not considered.
Conferences and workshops for a research community
ICML, NeurIPS, and ICLR have fixed cycles, anonymous review, rebuttals, and post-acceptance obligations. Their main tracks require mature evidence. As of 4 October 2026, ICML 2026 and NeurIPS 2026 submission deadlines have passed, and ICLR 2027’s full-paper deadline was 25 September 2026. Do not rush toward a deadline that has already closed or invent the next one. ICLR 2027 explicitly welcomes newcomers, but its reciprocal-review rules cap an author at one submission without a qualified reciprocal-reviewer coauthor.
For NLP, retrieval, and multilingual questions, the ACL family can be more relevant. ARR provides reviews; you then separately commit to a participating venue. ARR review itself is not publication. As of October 2026, a qualified designated service contributor, who can be a nonauthor, is needed for guaranteed review; otherwise allocation uses a lottery. Findings is an archival peer-reviewed publication associated with a conference, distinct from its main track. A focused workshop can offer a suitable audience for a smaller study, but check whether it is archival, what later submission is permitted, and what attendance costs apply. Workshop acceptance is not main-conference acceptance.
Name check: ICML is the International Conference on Machine Learning. ACM ICMR is the International Conference on Multimedia Retrieval, a different venue, appropriate only when multimedia retrieval is central. ICMR scope
What you might actually pay
Separate submission charges, publication or article-processing charges, compulsory author registration, and travel or accommodation. A paper being open access does not tell you who pays. The following are verified edition-specific examples, not future quotes. You graduated in 2019, so do not budget at student rates unless you independently meet the event’s current student eligibility.
Industry conference-only: USD 550 early / 650 late; virtual: USD 275
Proceedings-only option can use virtual registration; in-person papers require conference registration. Fees · author rules
NeurIPS 2026
Industry, Sydney conference plus tutorials: USD 650 early / 700 late
Workshops separate; virtual-only does not satisfy the main-track author registration rule. Fees · handbook
ICLR
Exact applicable price not verified
2026 accepted papers required one Conference and Workshops author registration. Do not infer a zero price or a 2027 amount. 2026 rule
ACL 2026
Industry presenting author: USD 1,350 / 1,500 in person; USD 600 / 750 virtual, early / late
Nonstudent ACL membership adds USD 100 if needed. Nonpresented Findings papers had a registration exception for this edition. Registration rules
ACM ICMR 2026
Early author: EUR 750 member / 850 nonmember; separate APC: USD 250 ACM/SIG member / 350 nonmember
One paper per full registration; same virtual/in-person registration price. APC coverage or waiver must be verified. Registration and APC
Fees checked 4 October 2026. Travel, lodging, currency conversion, and optional events are not included. Do not assume employer open-access coverage or a geographic waiver. No separate submission fee was identified in the reviewed flagship calls; check each future edition rather than treating that as a standing guarantee.
From finished manuscript to submission
Create accurate OpenReview and, if useful, ORCID profiles early; personal-email accounts may need moderation. Use your employment and education history truthfully. Choose the exact track, read its template and policies, prepare required checklists, and inspect the rendered submission. For anonymous review, remove identifying material from the paper and supplements, including repository metadata where relevant. Abstract and full-paper deadlines may differ.
Submit to one archival peer-review destination at a time unless the particular rules explicitly permit otherwise. Check preprint and publicity rules before posting or promoting a version. Answer reviews with evidence, revise clearly, and retain a record of changes. After acceptance, complete camera-ready files, publication forms, required registration, and presentation arrangements. Review duration and acceptance cannot be promised.
This is a planning estimate for a focused part-time project, not a promise of publication within three months. Review, revisions, deadline cycles, and venue scheduling may extend well beyond the research sprint. Keep the next gate small enough to reach before investing in the following phase.
When
Work and deliverable
Decision gate
Week 1
Select one question; read the closest work; verify hardware, data access, licenses, and venue fit. Write a one-page protocol.
One claim, identifiable gap, affordable test, no unresolved data or employer permission.
Week 2
Run a tiny end-to-end pilot. Include the simplest baseline, proposed intervention, timing, memory, and manual review of errors.
The pipeline runs, scoring is trustworthy, and scaled cost fits the cap.
Weeks 3–4
Reproduce key baselines; refine settings only on development data. Freeze prompts, splits, metrics, and exclusion rules.
No obvious leakage; comparisons have comparable information and budgets.
Weeks 5–6
Run the frozen evaluation, uncertainty analysis, ablations, and error taxonomy. Keep raw logs and configurations.
The main conclusion survives basic checks, or the negative finding is itself informative.
Weeks 7–8
Write methods and results from verified artifacts; build figures; audit every claim against evidence; assemble reproducibility files.
A reader can trace each table to a run and understand the limits.
Weeks 9–10
Seek feedback, improve exposition, rerun a clean setup, choose the final route, and check anonymity and policies.
No missing citation, broken dependency, confidentiality problem, or unsupported novelty claim.
Weeks 11–12
Keep buffer for failed runs or a targeted extension. Submit only if ready; otherwise reduce scope or continue.
Readiness determines submission. The calendar does not manufacture a result.
Keep an experiment ledger from day one
Each run should identify the code commit, dataset version and split, model revision and license, environment, random seeds, decoding settings, prompt version, hardware, start and end times, cost, and output location. Save per-example predictions and failures where permitted. Separate runs used for tuning from the final reported evaluation. If you change the protocol, record the reason and date; do not rewrite history to make exploratory choices look planned.
Write methods early to expose missing decisions. Leave result cells empty until runs exist. Draft the abstract last and check every claim against the evidence.
A sensible stopping rule
Stop an unworkable pilot and diagnose whether the implementation, data, or question needs changing. After a successful pilot, stop adding features once the frozen question has a clear, supported answer. Save unrelated ideas for another paper.
09 · Solo authorship and responsible work
You own the paper and its claims
You can lead the project as a single human author. Use an accurate affiliation, such as “Independent Researcher,” when that matches the circumstances and the venue permits it. Your former university is part of your education, not automatically a current research affiliation. Your current employer is not automatically the correct affiliation for private work either. Check employment terms, intellectual-property obligations, outside-work rules, and any approval requirements before using company time, infrastructure, code, data, or branding. Personal time and a personal laptop do not by themselves settle publication rights.
Keep employer and client material out of the research unless you have explicit permission. Prefer public, properly licensed datasets and open models. Publicly accessible content is not necessarily freely redistributable. Check dataset conditions, model use and redistribution terms, benchmark rules, and software licenses. Avoid personal or sensitive records unless there is a legitimate research basis, suitable approvals, and a data-protection plan. A first paper is easier to execute with low-risk public data.
AI can help with literature, code, critique, and drafting, but cannot take accountability or be a human author. Verify citations, code, and analyses, and follow the venue’s disclosure policy. Never let a draft imply that unrun experiments produced results.
If another person makes a substantial intellectual contribution, reassess authorship fairly. Record contributions and follow venue rules; routine feedback may warrant acknowledgment instead. Never buy authorship or add a prestigious name without a real contribution.
Current policies differ materially. ICLR 2027 requires an AI-use section and form disclosure; research and writing uses have different requirements. ARR asks for relevant writing/coding disclosures. NeurIPS 2026 distinguishes methodological use from basic editing. TMLR has the stricter human-sourced-ideas caveat discussed above. Follow the exact destination’s current rules rather than a generic disclosure template. Affiliation and rights: arXiv identity policy · rights and AI guidance.
First submission mistakes that are easy to avoid
Writing “first” or “state of the art” without a thorough comparison and a clearly bounded claim.
Using test examples to tune prompts, selecting only favorable runs, or hiding failed conditions.
Comparing methods with different token budgets or allowing one method access to answer-bearing evidence.
Relying solely on an LLM judge without a scoring audit or reporting accuracy without uncertainty.
Submitting to more than one peer-reviewed venue at the same time when their policies prohibit it.
Missing an anonymity rule, page limit, supplementary-material deadline, review-response request, or required registration.
Paying a venue because it promises fast acceptance, without checking its proceedings, editors, review process, and indexing claims.
10 · The writing blueprint
Build the manuscript around one result
Title and abstract: name the actual phenomenon, setting, and contribution. Summarize verified evidence and its limits without turning the abstract into marketing.
Introduction: explain the practical problem, what is unresolved, the research question, and a short list of contributions that the experiments genuinely support.
Related work: organize the closest evidence around the question. Explain the difference respectfully and precisely; a catalog of paper summaries is not enough.
Method or study design: define the task, data, intervention, baselines, assumptions, and evaluation protocol so another researcher can reconstruct the comparison.
Results and analysis: begin with the primary outcome, then ablations, costs, uncertainty, failure cases, and sensitivity to the main design choices.
Limitations and ethics: state exactly where conclusions may fail, what you did not test, and any risks or access constraints.
Conclusion and reproducibility material: restate what the study establishes, then provide permitted code, environment files, run instructions, and configuration manifests.
Use the venue’s current template and page limit. Keep a fuller technical draft during research, then trim it without losing core evidence.
Submission readiness checklist
Every main claim maps to a real table, figure, derivation, or cited source.
The closest prior work is represented accurately and citations open to the correct papers.
The final test set stayed isolated from tuning, and uncertainty is reported appropriately.
A clean environment can reproduce the principal result, or its access limitations are clearly documented.
Data, code, and model licenses permit the intended use and release.
The manuscript satisfies the target’s scope, template, anonymity, authorship, disclosure, and concurrent-submission rules.
All required fees and registration obligations are understood before commitment.
11 · Your first decision
Choose the question before writing the paper
Next, write a one-page protocol and run a tiny reproducible pilot. Pick one idea, verify hardware and storage, and define its primary metric and simplest baseline.
First, read the inspiration paper and two close predecessors, define the unresolved question, and identify a licensed data slice. Produce one pilot comparison table, inspected errors, and measured runtime and memory; then decide whether to proceed or change course.
The objective is a trustworthy first research contribution. A polished draft comes after a sound question and auditable evidence. You steer the work and make the scientific judgments; the assistant helps you move from idea to protocol to verified analysis to a manuscript you can stand behind.