acceptodds
Under review as a conference paper at ICLR 2027

Where the Risk Goes: Error Allocation in LLM Agents under Conformal Risk Control

Abstract

Conformal risk control can calibrate a commit-or-abstain rule so that an LLM agent's marginal probability of committing to a wrong answer stays below a chosen level. This guarantee does not control where those errors occur: the same overall risk can be distributed very differently across trajectories of different lengths. That distribution depends on the trajectory score the rule uses, while common summaries can obscure where uncertainty occurs or how uncertainty across steps combines. We introduce trajectory spatial contrast (TSC), a family of scores that treats the sequence of per-step uncertainty values as an ordered multivariate object. Each path is normalized to a fixed grid and scored by contrasting its spatial ranks against reference trajectories from correct and incorrect episodes. On synthetic examples designed to isolate trajectory structure, TSC separates failure patterns that no symmetric summary distinguishes and no additive score separates perfectly. Because every score is fitted on a separate reference split, conformal risk control gives TSC and competing scores the same marginal wrong-commit guarantee. Evaluating twelve scores across three benchmarks and three model families, we show that accumulating scores can concentrate almost all permitted wrong commits on the shortest trajectories, making the rule sensitive to shifts in the trajectory-length distribution. By contrast, the TSC family places errors most evenly across length groups in most benchmark–model pairs, at a modest cost in commit rate: it commits most in a minority of pairs and trails the best alternative by a few percentage points on average elsewhere.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Conformal risk control, LLM agents, uncertainty quantification, selective prediction, trajectory scoring, error allocation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_jeZkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how the trajectory score used in conformal risk control allocates wrong commits across trajectory-length groups. It introduces trajectory spatial contrast (TSC), which interpolates step uncertainty onto a fixed grid and contrasts spatial-rank norms relative to correct and incorrect reference trajectories, and a length-local variant, L-TSC. A separate calibration split supplies the standard marginal wrong-commit guarantee. Experiments compare twelve scoring rules across nine benchmark–model populations, finding that accumulating scores often concentrate wrong commits on short trajectories, while the TSC family generally reduces length-group concentration at some cost in acceptance rate.

Critical strengths
  1. The distinction between marginal risk and its allocation across lengths is useful and supported by concrete, consequential diagnostics. For example, Table 2 (p. 8) shows pseudo-Noisy-OR concentrating 94–100% of wrong commits in one length group in several populations, whereas L-TSC substantially reduces that concentration. Reporting both the largest error share H and the largest within-group wrong-commit rate W avoids conflating equal error mass with equal group risk. The mixture-shift analysis in Appendix B (pp. 21–22) makes the deployment implication precise under explicitly invariant within-group laws.

  2. The calibration scope and experimental controls are carefully specified. Section 2.3 and Theorems A1–A2 (pp. 5, 18–19) separate reference fitting from calibration, handle strict thresholds and ties, and state an expectation guarantee for marginal wrong commits rather than claiming conditional accuracy or groupwise control. The shared-representation experiment, extra reference draws, label-budget control, and signal/geometry sensitivity analyses help assess whether the observed trade-off is stable. These are substantial strengths even though the CRC guarantee itself is inherited.

Critical weaknesses
  1. The evidence is less conclusive about the specific value of spatial-rank contrast than about the general diagnosis. In the common-representation study, TSC uses correctness-labeled reference clouds, while all five competing aggregations of the same U fit nothing and consume no reference labels (Appendix C, p. 22). The equal-label-budget control instead gives those methods more calibration examples (pp. 27, 30), so it does not distinguish spatial-rank geometry from the benefit of learning a nonlinear correctness score on the four coordinates. A single conventional nonlinear score fitted only on the same reference split, with the same CRC calibration, would directly address that attribution. This limits the methodological contribution established by Figure 1/Table 9; it does not invalidate the length-concentration diagnosis or the stated guarantee.

  2. The strongest acceptance-rate results for L-TSC occur in two WebShop populations with problematic agent–environment interaction. Appendix D.1 (p. 26) states that Gemma fails the pilot parsing/length checks and that 95% of Mistral test steps are unparsed; both populations are retained in the main comparisons and nine-population summaries. Table 2 (p. 8) gives L-TSC its acceptance-rate leads on these two populations. Transparent reporting is welcome, but it remains unclear how much those gains reflect handling malformed-action loops rather than uncertainty in functioning agents. The other populations still support the concentration finding, so this primarily limits the breadth of the practical utility evidence.

Questions for authors
  1. Can the authors compare TSC with one reference-fitted nonlinear correctness score on exactly the same U (and, for the L-TSC comparison, the same available length information), preserving the reference/calibration/test separation? This would clarify whether the favorable acceptance/concentration trade-off depends specifically on spatial-rank contrast.

  2. For the two WebShop populations discussed on p. 26, how many episodes contain parsing failures, and how do the acceptance and wrong-commit differences break down between trajectories with and without those failures? This would clarify the source and practical scope of the L-TSC acceptance-rate gains.

1 comment

I have read the rebuttal and decided to maintain my score.

Reviewer_dR3FOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Conformal risk control (CRC) can calibrate a commit-or-abstain rule so an LLM agent's marginal probability of committing a wrong answer stays below a target α — but it says nothing about where those errors land across trajectories of different lengths. The paper diagnoses that accumulating scores (e.g. pseudo-Noisy-OR) mechanically penalize long trajectories and therefore concentrate almost all permitted wrong commits on the shortest ones, leaving the rule fragile to shifts in the length distribution. It proposes TSC (trajectory spatial contrast): interpolate the per-step uncertainty path onto a fixed D=4 grid, compute spatial-rank norms against correct vs. incorrect reference clouds, and score by their difference in [−1,1]; plus L-TSC, a length-kernel-weighted variant. Both are fit on a separate reference split so CRC still grants the same marginal guarantee. Across 12 scores × 3 benchmarks × 3 models (9 populations, 50 splits, α=0.10), TSC/L-TSC spread accepted errors most evenly across length groups on most populations, at a ~3 pp cost in commit rate. The central tension: the method optimizes an uncontrolled, descriptive allocation metric, and the empirical edge on two of three benchmarks rests on overlapping resamples whose confidence intervals understate uncertainty.

Critical strengths
  • the diagnosis is clean and the control is sharp. the 1-(1-u)^T monotonicity argument explains error-on-short-trajectories without handwaving, and the length-only gate (score = log T, ignore all signals) is a lovely ablation: it recovers 75-88% of pseudo-noisy-or's acceptance on 6/9 populations, showing most of the "uncertainty" signal is just trajectory length wearing a hat.

  • honest conformal hygiene. every score including TSC is fit on a separate reference split so CRC's marginal guarantee holds uniformly, the order-statistic threshold is standard and correct, and the paper repeatedly says W and H arent controlled instead of pretending they are. correct scientific posture!

Critical weaknesses

Coverage assessment

The conformal-prediction and uncertainty-aggregation lineage is covered well and fairly. The gap is on the "beyond-marginal / stratified risk control" branch that is closest to the paper's actual thesis: Role-Stratified CRC (2607.24343) and the certification-bounds work (2606.29054) pursue the same "marginal is not enough" program along a different stratification axis and should be discussed as the nearest prior art — both to position C1 honestly and to clarify that this paper's novelty is the ordered-sequence score, not the observation that marginal control leaves allocation free.

Missing citations

  • arXiv:2607.24343 (Role-Stratified CRC) — nearest conceptual neighbor; must be cited and contrasted (their axis = role/field; yours = trajectory length).

  • arXiv:2606.29054 (CRC certification bounds / impossibility) — supports and sharpens the commit-rate-vs-allocation trade-off argument.

  • arXiv:2604.09155 (CORA) — adjacent CRC-for-agents instantiation.

  • the headline sells an uncontrolled quantity. the advertised win is low W/H, which carry no guarantee, so the central claim is empirical and draw-dependent. The body is a trade-off rather than a dominance: TSC/L-TSC trail the per-population commit leader by ~3pp and on some populations (bfcl/muse) lose on both allocation axes. the abstract should say "more even allocation bought with ~3pp commit rate", not imply it just wins.

  • pseudoreplication hides the weak wins. overlapping-rotation bootstrap CIs understate uncertainty, so every sub-1pp gap on musique/bfcl is unsupported, and the "lowest W on 6/9, lowest H on 7/9" tally is just best-cell counting with no paired significance test. strip it to webshop (fixed test) and the robust story shrinks a lot. plus "where the risk goes" is only ever studied along length terciles, never task type or tool or difficulty, so the broad title oversells a one-axis result. and TSC quietly needs labeled reference clouds that the baselines dont, with no label-noise test despite soft checkers (MuSiQue F1≥0.5).

Questions for authors

can you address weaknesses from above?

  • if you redo the error bars as paired per-episode diffs with cluster/subsampling over the fixed pool, do any of the sub-1pp W/H wins on musique/bfcl survive?
  • if you restrict the strong claims to webshop (the only fixed-test benchmark), how much of the headline remains?
  • group the errors by something other than length (tool count, task category, difficulty): does the allocation story still hold, or is it length-specific?
  • inject x% label flips into the reference clouds: how fast do TSC's spatial ranks and the W/H wins degrade?
Reviewer_v33uOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Under conformal risk control (CRC), an agent's commit-or-abstain rule bounds the marginal probability of committing to a wrong answer at α, but the allocation of accepted errors across trajectory lengths is left to the score. On a grid of 12 scores across nine benchmark–model populations (MuSiQue, BFCL v3 multi-turn, WebShop × three models; 50 splits; α = 0.10), pseudo-Noisy-OR — whose score grows mechanically with length T — concentrates ≥91% of accepted errors in the shortest tercile on 7/9 populations, and a length-only gate recovers 75–88% of its commit rate on 6/9. The paper proposes TSC, which contrasts spatial-rank norms of the interpolated uncertainty path against correct/incorrect reference clouds (Prop. A1 gives a witness where additive and symmetric scores fail). TSC/L-TSC attain the lowest error concentration on 7/9 populations at ~3.1 pp lower commit rate than the per-population best, and Corollary A1 bounds the wrong-commit rate under a TV-bounded shift of the length mixture.

Critical strengths
  1. A consequential, well-isolated diagnosis for the CRC-for-agents literature. The finding that a satisfied marginal cap can coexist with near-total error concentration in one length regime — and that an accumulating score acts largely as a length classifier — is unexamined in the conformal-abstention-for-agents work the paper builds on (Abbasi-Yadkori et al. 2024; Opoku & Banahene 2026) and should influence how trajectory scores for certified abstention are chosen and reported.
  2. Methodological hygiene that makes the comparison meaningful. All twelve scores share the same recorded trajectories and signals; the reference set is separated from calibration so every score's threshold retains the standard CRC guarantee (Theorems A1/A2); every baseline adaptation is documented (Table 3); and the mixture-shift analysis comes with an honest counterexample. The paper never claims more than the marginal certificate provides, which is why I trust the grid results despite reservations about their interpretation.
Critical weaknesses
  1. The objective is under-justified and the natural competitor is missing. Even allocation is motivated by a silent-failure scenario and the TV-shift bound, but no concrete harm of concentration is measured, and stratified (Mondrian-style) calibration by length tercile — discussed in the paper's own related work as the family that controls specified groups — would directly bound per-group wrong-commit rates and is the obvious alternative to allocation-by-score-design. Without that comparison, the paper has not established that score design is the right lever for the problem it poses, and the ~3-point commit-rate cost buys an unexamined trade-off.
  2. The advantage is modest and unguaranteed. TSC leads on commit rate in no population of the full grid (Table 2); the evenness property is empirical — one rollout per task at temperature 0.7, within-group laws assumed fixed — and Proposition A1's exact separation holds only on idealized clouds, degrading to a graded ranking under any perturbation. Margins over the best non-TSC competitor are a few points of H in most cells, and simpler fixes already in the grid (length-normalized scores) may capture most of the practical benefit.
Questions for authors
  1. Why should a practitioner prefer even allocation to concentration, concretely? Can you exhibit or simulate a deployment shift in which pseudo-Noisy-OR's concentrated allocation causes measurable harm that TSC avoids? This is the crux of whether the commit-rate cost buys anything.
  2. How does a Mondrian-style rule (separate CRC thresholds per length tercile, each calibrated at α) compare on commit rate and allocation? If it is competitive, how does that change the paper's recommendation?
  3. Given one generation per task, how stable are the W/H orderings across regenerated trajectories (not just re-splits)? One regeneration on a subset of one population would help.

Comments on the paper

1 comment

Dear Reviewers,

This is a friendly reminder to engage in the rebuttal discussion and review the authors' responses to your comments. Please assess whether their clarifications and additional evidence adequately address your concerns.

If any concerns remain unresolved, we encourage you to follow up with the authors during the discussion period.

Thank you for your time and contribution to the review process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.