acceptodds
Withdrawn from ICLR 2027

LicenseHarness: Preserving Evidence Boundaries in Autonomous Research Writing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language model research agents always suffer from defensive writing, cluttering generated papers with redundant disclaimers. Existing prompt engineering and rule-based linters fail across long interactions because debugging histories mix factual updates with conversational pressure, inevitably causing defensive relapse. We formalize defensive writing as unlicensed hedges that empirical evidence does not support, proving that removing defensiveness while preserving legitimate scientific boundaries is a unified constrained optimization problem. We introduce **LicenseHarness**, a framework that externalizes evidence into a typed state to block narrative pressure, combined with history invariance distillation and constraint-aware reinforcement learning that internalizes licensing directly into the model weights. Evaluating across 210 empirical machine learning bundles, four base models and production agents, **LicenseHarness** reduces conversational drift by 79 times while maintaining 93.8% epistemic boundary coverage. Double-blind human reviewers consistently prefer its outputs in persuasiveness and acceptance intent.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Defensive Writing, Preference Model Bias, RLHF, Alignment, Policy, Harness

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3LszOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper tackles defensive writing in LLM research agents: hedges and disclaimers that pile up as adversarial history accumulates. It formalizes defensiveness as unlicensed hedges, qualifiers the evidence behind a claim does not support. Each claim gets a warrant w(c) and a boundary set B(c) computed by deterministic rules over six evidence dimensions, and text is scored by defensive mass D, overclaim O, boundary coverage Cov, and repetition Rep. The system works at two levels: a harness that keeps evidence in an external typed state so adverse feedback updates the license only through a finite-alphabet channel, with an information-theoretic leakage bound, and a distilled policy trained with history-invariance SFT plus constraint-aware GRPO. On 210 evidence bundles (60 eval, 360 expert-annotated claims), LicenseHarness-RL on Llama-3.1-70B reaches D = 2.35 with 93.8% coverage, holds drift to +0.07 under adversarial history against +5.54 for the strongest baseline, transfers to Claude Code and OpenCode ports, and wins a 10-reviewer double-blind preference study. Validation includes mechanism-isolating ablations, judge calibration against expert annotation (kappa = 0.88), and a human preference study.

Critical strengths
  • The mechanism is tested, not just proposed. The ablations isolate each claimed cause of relapse: removing the typed channel drives drift from +0.18 to +3.84, lifting the finite-alphabet bound to +4.12, keeping multi-turn context to +1.95, mapping directly onto Proposition 2's assumptions. Combined with three-way provider separation (writer, verifier, judge from different providers), 5 seeds on the causal tier, Holm-Bonferroni correction, and working ports into Claude Code and OpenCode (D down over 6 points, Cov up 17), the evidence carries the central claim unusually well for this area.
Critical weaknesses
  • The evaluation loop grades its own homework. Every headline metric is computed by a learned judge applying the paper's own licensing taxonomy, the judge's human validation measures agreement with annotations of the same author-defined constructs, and the human study measures persuasiveness rather than factual alignment. Nothing external checks whether LicenseHarness text is actually more truthful. This is consequential because the system is trained, gated, and scored against one taxonomy, so the current evidence cannot rule out that it optimizes the paper's own definition of good writing. Concurrent work shows this divergence is not hypothetical: Liao [1] finds that AI reviewers score hedged rewrites higher and drop them by 0.82 points overall (1.12 on soundness) when only the qualifications are removed, while human readers find the same rewrites harder to read, and most of the rewarded qualifications are redundant. An LLM judge scoring qualification content thus embeds preferences human readers do not share, in either direction, which is exactly why an external, human-grounded check is needed here. An independent fact-check of claim-evidence alignment would close the loop, which is why I rate this fixable rather than fatal.
  • The warrant lattice is the load-bearing piece and the weakest measured one. Everything downstream (licenses, gates, rewards, metrics) depends on (w, B) from six hand-thresholded rule dimensions, yet the lattice agrees with experts at tau_b = 0.78 / kappa_w = 0.79, below the 0.84 inter-expert ceiling, and Proposition 1 says all residual trade-off comes from exactly this error. There is no sensitivity analysis showing the main rankings survive the ordinal fallback model or perturbed thresholds. Compounding this, all 210 bundles are empirical ML papers while the paper's own Figure 6 shows alignment dropping to 55% of bundles above 0.95 in Social and Cognitive Sciences, so the "research writing" framing outruns the evidence.

References

[1] Liao, Junchi. "Writing for the Reviewer: Defensive Writing in GPT Models." arXiv preprint arXiv:2610.11355 (2026).

Questions for authors
  • Can you re-run the Tier (iii) main tables with the warrant estimated by the ordinal fallback model (App A.1), or with perturbed thresholds, and report whether the rankings and the 79x drift claim survive? Given Proposition 1, this is the single most important robustness check.
  • Can you add one evaluation against a construct you did not define: an independent expert fact-check of claim-evidence alignment on a sample of LicenseHarness versus Hook outputs, or factual-consistency scoring against the source result tables?
Reviewer_5nmoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes LicenseHarness, a framework for reducing defensive writing in autonomous research agents while preserving scientifically necessary evidence boundaries. It introduces evidence-induced assertion warrants, a typed external state that separates factual updates from conversational pressure, verifier-gated generation, and history-invariance distillation with constraint-aware GRPO. Experiments on 210 evidence bundles, multiple foundation models, and production agents report substantial reductions in defensive writing and conversational drift while improving boundary coverage. Human evaluations further indicate improved persuasiveness and reviewer preference.

Critical strengths
  1. The paper addresses a meaningful problem by distinguishing redundant defensive language from scientifically necessary evidence boundaries. This formulation goes beyond conventional style suppression and explicitly considers epistemic correctness.

  2. The typed-state and fresh-context architecture is well motivated, and the adversarial-history experiments demonstrate substantial robustness improvements. The production-agent integrations and human evaluations further support its practical relevance.

Critical weaknesses
  1. Questionable problem formulation. The proposed warrant formulation assumes that each claim has a maximum justified assertion strength and treats weaker expressions as unlicensed, although legitimate scientific caution does not necessarily constitute defensive writing. Moreover, the fixed thresholds across six evidence dimensions may not reflect the different evidentiary requirements of diverse scientific claims.

  2. Insufficient novelty and unclear necessity of the proposed framework. The method combines existing techniques, including structured evidence states, fresh-context generation, verifier-guided filtering, and policy distillation. The experiments do not sufficiently establish whether the proposed licensing framework offers meaningful advantages over simpler alternatives using structured evidence summaries, fresh-context drafting, and equivalent verification budgets.

  3. Limited evaluation realism and generalizability. Most experiments rely on preconstructed evidence bundles and controlled adversarial histories rather than end-to-end autonomous research workflows, leaving the effectiveness under realistic, evolving scientific evidence unclear. Additionally, the apparent use of evaluation-set expert annotations for warrant calibration raises concerns about evaluation independence.

  4. Overstated theoretical guarantees and experimental inconsistencies. The theoretical results rely on restrictive assumptions or standard inequalities and do not establish reliable evidence licensing under imperfect estimation and verification. Furthermore, the gated and ungated RL results across Tables 1, 2, and 4 require clarification, particularly because Table 2 uses the gated score while discussing gate-free history invariance.

Questions for authors
  1. How do the authors justify treating expressions below the maximum warranted strength as unlicensed hedging? How does the framework distinguish legitimate scientific caution from unnecessary defensiveness across different claim types?

  2. Can the authors clarify whether simpler approaches combining structured evidence summaries, fresh-context generation, and verifier-guided filtering were evaluated under matched inference budgets? What evidence establishes the necessity of the proposed licensing and training mechanisms?

  3. Could the authors clarify the inconsistent gated and ungated RL results across Tables 1, 2, and 4? Additionally, were the expert annotations used for warrant calibration completely independent of the final evaluation set?

Reviewer_7a3yOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses a recent writing problem of LLMs, defensive writing with a new method, liscense harness. It attributes defensive writing to the existence of unsupported hedges and uses an evidence bundle to define and evaluate defensive writing. The authors also introduce a posttraining method to distill the capability into the LLM.

Critical strengths
  1. The problem that LLMs may overly hedge their claims and lead to unusable outputs is an interesting problem worthy of studying.
  2. The paper's experiments demonstrate significant emperiment improvements on their own metrics.
Critical weaknesses
  1. The major problem the authors claim to address, defensive writting, does not seem to have a proper definition in the paper. The explanation in the introduction about what the problem is is unclear, and the Definition 3.1 is actually a characterization based on the licensing map λy instead of an intuitive definition only dependent on the model behavior. It's arguable that this "formalization" is biased and cannot be used elsewhere and therefore an invalid contribution.
  2. The problem identification utilized in this paper is dependent on a set of predefined evidences. If the evidence set is not fully identified, we may falsely identify proper hedges as defensive writing. Furthermore, the paper does not seem to have an explicit discussion about this false positive possibility.
  3. The qualitative examples shown in the paper does not give convincing support. After seeing the examples, it's hard to say which ones are more preferable in general. It is possible that the authors are optimizing their model to better fit some kind of biased human preferences that do not generalize.
  4. Many concept formulation part of the paper are presented in the Method section. Mixing the method and the problem formulation makes both unclear and uneasy to understand.
  5. The mathematical theorems in the paper are poorly motivated. It's hard to see what is the connection between the main problem and the theorems.
Questions for authors
  1. Could the authors report the proportion of hedges that human experts consider legitimate but that the system classifies as ⊥, broken down by cause (information missing from the bundle, boundary type not in 𝔅, or wording issues)?
  2. If the evidence bundle is deliberately reduced (for example, by removing some logs or configuration files), how do D and Cov change? This would quantify the method’s sensitivity to the completeness of the evidence.
  3. Could the authors explain how are the theorems connected to the central problem?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.