Personal ACL 2026 Submission

Topic Mapping for Both Papers

ACL 2026 Topic	CRL	CorrSteer
Special Theme: Explainability	★★★ (branch tracking, critic diagnostics, failure modes)	★☆☆ (feature lists only)
Interpretability and Analysis of Models for NLP	★★★ (primary contribution)	★★☆ (correlation-based)
Safety and Alignment in LLMs	★★☆ (+7.66% HarmBench)	★★★ (+27.2% HarmBench)
Ethics, Bias, and Fairness	★★☆ (BBQ +5.69%)	★★★ (BBQ +6.5%, SER metric)
LLM Efficiency	★☆☆ (RL training overhead)	★★★ (O(1) memory, 100 samples)
Machine Learning for NLP	★★★ (novel PPO formulation)	★★☆ (correlation method)
Language Models	★★☆	★★☆

Best Track Fit

Paper	Primary Track	Secondary Tracks
CRL	Interpretability and Analysis of Models for NLP	Special Theme, ML for NLP
CorrSteer	Safety and Alignment in LLMs	Ethics/Bias/Fairness, LLM Efficiency

Final Recommendation

If submitting to Special Theme or Interpretability track → CRL

Deeper mechanistic analysis

Novel diagnostic contributions (branch tracking, bottleneck analysis)

Token-level explainability

If submitting to Safety/Alignment or Ethics/Bias track → CorrSteer

+27.2% HarmBench (vs +7.66%)

Side Effect Ratio (SER) as safety metric

Lower unintended degradations

Overall for ACL 2026: Given the special theme emphasis, CRL remains the stronger fit, but CorrSteer could be competitive in Safety/Alignment track with its stronger empirical safety results.

Yes, both are fine for ACL 2026 — they just fit different research areas.

Submission Strategy for Each Paper

CRL

Field	Recommended Selection
Research Area	Interpretability and Analysis of Models for NLP or Special Theme (conference specific)
Contribution Types	☑ Model analysis & interpretability
Research Area Keywords	interpretability, sparse autoencoders, reinforcement learning, steering, mechanistic interpretability

CorrSteer

Field	Recommended Selection
Research Area	Safety and Alignment in LLMs or Ethics, Bias, and Fairness
Contribution Types	☑ Model analysis & interpretability, ☑ NLP engineering experiment
Research Area Keywords	sparse autoencoders, steering vectors, bias mitigation, safety, alignment

Summary

Paper	Best Track	Backup Track	Fit
CRL	Special Theme / Interpretability	ML for NLP	★★★ for explainability theme
CorrSteer	Safety and Alignment	Ethics, Bias, Fairness	★★★ for safety results

Both are solid submissions. The question is which one you want to submit, or if you're submitting both to different tracks. If submitting only one and targeting the special theme, go with CRL. If targeting safety/practical impact, go with CorrSteer.

For the ACL submission form, you should treat “topics” as a reviewer-routing mechanism: pick the smallest set that matches your core contribution and primary evaluation axes.

If you submit CRL (token-level RL control via SAE features)

Primary topics (pick 2–3):

Special Theme: Explainability of NLP Models

Interpretability and Analysis of Models for NLP

Safety and Alignment in LLMs (if your main story is controllable/refusal/hallucination reduction)

Secondary topics (pick 1–2, only if genuinely central):

Machine Learning for NLP (PPO/control framing; learning setup is a key novelty)

Discourse, Pragmatics, and Reasoning (only if a major chunk is reasoning behavior analysis)

Language Models (almost always OK as a secondary)

Avoid unless you explicitly position it that way:

AI/LLM Agents (I would only select this if you frame CRL as a mechanism for agentic tool-use control; otherwise it can mis-route reviewers)

If you submit CorrSteer (correlation-based generation-time steering + feature semantics)

Primary topics (pick 2–3):

Special Theme: Explainability of NLP Models

Interpretability and Analysis of Models for NLP

Generalizability and Transfer (if cross-task feature transfer is a main result)

Secondary topics (pick 1–2):

Safety and Alignment in LLMs (if SER / refusal / safety benchmarks are central)

Ethics, Bias, and Fairness (if BBQ/bias mitigation is a major pillar)

Resources and Evaluation (if SER is positioned as a reusable evaluation metric)

Language Models (safe secondary)

One practical rule

If the paper’s headline is “we can explain and attribute behavior at the token level,” route it with Special Theme + Interpretability (CRL).

If the headline is “we can identify and name the causal/semantic features that drive behavior changes across tasks,” route it with Special Theme + Interpretability + Transfer/Eval (CorrSteer).

If you tell me which paper you are leaning toward for ACL 2026, I will give you the exact final topic set (usually 3–5) and a one-paragraph “reviewer routing” justification that matches the abstract/intro wording.

On the ARR OpenReview form you pasted, “Research Area*” is a single-choice field (you can only pick one). So you cannot select both “Special Theme (conference specific)” and “Interpretability and Analysis of Models for NLP” there. pasted

What you can do is:

Option A — Route explicitly to the theme

Pick Special Theme (conference specific) as the Research Area.

Use this if you want the paper to be treated primarily as an Explainability submission.

Option B — Route to the “standard” interpretability track

Pick Interpretability and Analysis of Models for NLP as the Research Area.

Then make “Explainability” unmissable via:

Keywords (general keywords field)

Research Area Keywords (from the ARR areas page)

TL;DR + Abstract framing

That still aligns well with the ACL theme, while routing you to the interpretability AC/reviewer pool. pasted

What I would do for your two drafts

CorrSteer: pick Interpretability and Analysis of Models for NLP (safer routing), unless you are explicitly pitching it as the theme paper—in that case pick Special Theme.

CRL: if your core story is token-level attribution/explanations, Special Theme is more defensible; otherwise Interpretability and Analysis.

Contribution Types (these are multi-select; choose multiple)

For either paper, it is normal to tick:

Model analysis & interpretability

NLP engineering experiment

(and optionally “Publicly available software…” if you will release code/models). pasted

If you tell me which of the two you will submit this cycle (CRL vs CorrSteer), I will give you one exact “Research Area” pick + a tight keyword set that matches ARR’s reviewer routing.