AI Tech Inspire

ICLR 2026 Clarifies Score Reversion: What Authors Can Still Do

ICLR 2026 made an unusual mid-cycle decision: revert review scores to their pre-discussion state after a leak. If you authored a submission, the obvious worry is whether all that careful rebuttal work just evaporated. Short answer: it didn’t. The program chairs clarified what changed, what didn’t, and how area chairs will evaluate papers from here.

Key points at a glance

  • An email was sent to ICLR 2026 authors explaining a rollback of review scores to their pre-discussion state.
  • Only the Official Review scores were reverted; author responses and prior reviewer discussions remain available to area chairs (ACs).
  • Authors can still post additional comments; a concise summary comment to the AC is encouraged.
  • ACs get more time for meta-reviews and are instructed to consider author responses and previous discussions, and to estimate how impressions might have evolved without the cutoff.
  • Emergency ACs will be recruited to manage workload.
  • ACs have broad discretion; scores are only one signal among several.
  • Reason for rollback: a leak occurred as early as November 11, before the discussion phase, implying possible collusion during discussions; reverting to pre-discussion was deemed fairest.
  • No further reviewer discussions or public comments are allowed; author additional comments are permitted.

What actually changed—and what didn’t

The headline is easy to misread. The revert applies to only the Official Review scores. Your rebuttal text, clarifications, experiment updates, and any prior reviewer exchanges are still visible to the AC handling your paper. In other words, the record of your engagement lives on—and it matters.

What’s paused is new public discussion. Reviewers aren’t continuing the back-and-forth, and public comments are closed. However, authors can still post an additional comment. The chairs even suggest using this window to drop a targeted summary addressing what changed during and after rebuttal.

Key takeaway: your responses will not go to waste. The signals you’ve already provided are in play—now amplify them with a focused summary.

Why the rollback happened

The chairs point to a leak dated as early as November 11—before the discussion period began. If a leak compromises anonymity or process integrity, anything that happened afterward could be influenced, at least in principle. That’s why the decision was to revert to pre-discussion scores. It’s a blunt instrument, but the stated goal is consistency and fairness across all papers.

At AI Tech Inspire, this lines up with a broader theme in scientific venues: preserving the credibility of peer review while minimizing advantage from any information asymmetry. It’s not comfortable, but conferences typically prefer conservatism over uncertain adjustments when the process is compromised.

How ACs will evaluate from here

ACs are getting a longer runway to write meta-reviews. Their guidance includes three directives:

  • Explicitly factor in author responses and past discussion threads.
  • Estimate how reviewer impressions might have evolved had discussion continued uninterrupted.
  • Exercise discretion—scores are one signal, not the decision.

Emergency ACs will be recruited to manage load, reducing rushed judgments. Historically, AC discretion has been a backbone of decision-making: they weigh scores, response quality, reviewer engagement, and their own technical assessment. This time, that discretion is even more central.

Actionable moves for authors—today

If you plan to use the additional comment option, think of it as a meta-rebuttal targeting a single reader: the AC. It’s not a place to rehash; it’s where you compress signal. Consider this structure:

Title: Summary for AC — Post-Discussion Clarifications and Evidence

1) Core claim and contribution (1–2 sentences)
2) What changed due to rebuttal (bulleted):
   - Resolved misunderstanding X with experiment Y
   - Added ablation Z; delta +3.2% over baseline
3) Evidence snapshot:
   - Link to code artifact and exact commit hash
   - Repro instructions and config file used
4) Risk/limitation acknowledged and bounded
5) Why the work remains valuable even if assumption A is weak

Tips that help ACs quickly verify credibility:

  • Post immutable references: commit hashes, exact seeds, and configuration files. If applicable, provide a minimal requirements.txt and a one-line bash command.
  • Include a compact table of results with deltas vs. the best baseline, and a pointer to where the review misunderstanding was addressed.
  • Offer a short reproduction path: for example, a small dataset slice and script that reproduces the main figure end-to-end.
  • If you host artifacts, a well-structured Hugging Face space or model card helps ACs validate quickly.

Remember, ACs are asked to estimate how impressions might have shifted with more time. Make that estimation easier with concrete, checkable updates.

Why this matters to developers and engineers

Even if you’re not submitting to ICLR, the conference’s decisions shape the ideas that reach production stacks. Many techniques that ripple into PyTorch implementations or model hubs start as ICLR papers. If the signal the community receives shifts—say, because rebuttal-driven score boosts are discounted—then which ideas get early attention can change.

For practitioners, this is a nudge toward rigorous, verifiable artifacts. Whether you build with TensorFlow or coordinate training with CUDA, the papers you adopt fastest tend to have clean repos, clear configs, and reproducible numbers. The current situation raises the bar for that kind of clarity.

Rethinking your paper’s “deployment readiness”

Think of your submission like shipping an internal library. If a process disruption occurs, can a reviewer or AC still audit your claims quickly? A “deployment-ready” paper makes this trivial:

  • Single-command run script, e.g., bash run_experiment.sh --config configs/base.yaml.
  • Explicit hardware assumptions (A100 vs. V100), batch sizes, and expected wall-clock.
  • Seeded evaluations with variance intervals.
  • Concise rationale for design choices—why this loss, why this augmentation, why this scheduler.

In industry terms, your meta-comment is like a change log for the AC: what changed, why it matters, and how to verify fast.

Fairness vs. fidelity: the open questions

Rolling back scores to pre-discussion is designed to reset the field. But it also reduces the visible impact of author–reviewer convergence that happened during discussion. A few questions are worth watching:

  • Will AC guidance lead to consistent estimation of “what would have happened” if the discussion continued?
  • Does the added time and emergency AC support translate into more careful meta-reviews?
  • How will authors’ new summary comments affect borderline decisions?

From a community perspective, there’s a tension between process integrity and the value of iterative debate. The chairs’ message signals an attempt to preserve both: freeze the potentially compromised signal, but keep the richer context in play for ACs.

A practical checklist for your summary comment

  • Lead with the core contribution: one sentence the AC can quote in a meta-review.
  • State the precise misunderstanding(s) you resolved and how.
  • Include a tiny, verifiable win: a small script or notebook that reproduces the main result on a subset in minutes.
  • Add a table of deltas vs. the strongest baseline; keep it compact and scannable.
  • Link to a stable artifact (commit hash, tag) and specify environment versions.
  • Name the limitation you accept, and bound its impact.

Consider even a minimal code box in your comment:

# Reproduce Table 2 (subset, ~8 mins on 1xA100)
conda env create -f env.yml
python train.py --config configs/paper.yaml --fast_eval --seed 42
python eval.py --checkpoint ckpt/subset.ckpt --report table2_subset.json

That level of clarity helps ACs do their job under time pressure and aligns with best practices many teams already use for shipping internal ML tooling.

The bottom line: while reviewer score rollbacks feel jarring, the chairs clarified that your work in the rebuttal still counts. ACs will read it, factor it in, and have more time to do so. Use the author comment option to surface the most checkable, decision-relevant evidence. As always, the strongest signal in research—and in engineering—is high-quality, verifiable artifacts. At AI Tech Inspire, this is the kind of process twist that nudges the ecosystem toward more reproducible, developer-friendly science.

Recommended Resources

As an Amazon Associate, I earn from qualifying purchases.