{"id":"8c21db71-420f-4d7c-b169-ecaedc0baa6e","arxiv_id":"2506.12796","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Surprise Calibration uses the model's own surprise at each demonstration's label to dynamically correct class priors in in-context learning, improving accuracy on eight NLP benchmarks.","lead":"This paper introduces Surprise Calibration, a method that adjusts a language model's predictions during in-context learning by tracking how surprising each demonstration's label is to the model. It reports accuracy gains over existing calibration baselines on eight NLP benchmarks while avoiding the extra per-query inference cost of dynamic prior estimation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The surprise-to-prior-shift premise in Eq. 7 is not formally guaranteed, and Figure 2 only supports it conditionally; the SC model's reported gains do not independently validate this central interpretive claim.","rationale":"The reader flagged the surprise-prior link as the weakest assumption; I agree but sharpen why it is load-bearing. The theoretical derivation's denominator is not what surprise measures, and the empirical validation is explicitly conditional and unstable in aggregate. The end-to-end training means downstream accuracy cannot validate the intermediate premise. The paper's limitations section admits the framework may not fully capture ICL. Therefore the central claim of surprise as a reliable signal is conditionally supported at best. The concrete checks would either restore confidence (if unconditional correlations hold across all datasets) or require the authors to weaken the interpretive claim and present SC as a purely empirical learned calibration method. The main accuracy comparison also lacks error bars, but that is secondary; even with error bars, the premise question remains. Thus I keep the CONDITIONAL verdict.","tokens_in":24436,"tokens_out":17093,"duration_ms":200539,"concrete_test":"Independently re-run the Figure 2 protocol on all 8 datasets (not just four) and report unconditional Spearman rho between signed surprise and empirical class-prior shift, alongside the label-conditioned rho, with bootstrap 95% CIs. If the unconditional correlation is non-significant or changes sign on multiple datasets, the claim that surprise is a reliable prior-shift signal is not supported. As a second check, simulate Eq. 6-7 with a known p(z), p(e|z), p(y|e,z) where p(e|z) varies; if -log E[p(e_j,y_j|z)/p(e_j|z)] is not monotonically related to 1/E[p(e_j,y_j|z)], the Eq. 8 feature is theoretically unmotivated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 derives a Bayesian update whose amplification factor is the denominator E[p(e_j,y_j|z)] (Eq. 6), but the surprise feature is -log E[p(e_j,y_j|z)/p(e_j|z)] (Eq. 7). The authors concede the link is 'not formally guaranteed due to the normalization by p(e_j|D_{j-1})', yet Eq. 8 constructs the entire surprise vector from this unproven proxy. The only direct evidence, Figure 2, conditions each correlation on the inserted label's class, and the authors report that aggregate correlations are 'occasionally unstable' (attributed to anti-recency bias). Thus the paper's central interpretive claim—that surprise 'systematically covaries with prior adjustments'—is not established as a general property; it is a label-conditioned heuristic. Because the GRU is trained end-to-end on supervised labels, the accuracy improvements in Table 2 would obtain even if the surprise-prior premise failed, so they do not rescue the premise. This leaves the method's motivation supported only by a partial correlation analysis that the paper itself admits is unstable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper interprets in-context learning (ICL) as implicit sequential Bayesian inference, defines a per-demonstration surprise signal as the negative log probability of the observed label before it is revealed, and hypothesizes that surprise indicates a shift in the class prior. On this basis it proposes Surprise Calibration (SC), which feeds a sequence of surprise vectors into a GRU, decodes a prior-adjustment vector, and uses it to recalibrate the model's log-probabilities. The paper reports experiments on eight datasets with Qwen2.5-3B and Qwen2.5-7B, comparing SC against ICL, BC, LinC, CC+, BC+, and LinC+ across demonstration selection and ordering strategies, and claims that SC outperforms state-of-the-art ICL bias calibration baselines while requiring fewer inference iterations.","tokens_in":24690,"tokens_out":4277,"duration_ms":49098,"significance":"If the central interpretive claim held, the paper would make a useful conceptual and practical contribution: it connects a measurable quantity (surprise) to class-prior adjustment in ICL and offers a calibration method that avoids per-query repeated sampling. The Bayesian identities in Eqs. (5)–(7) are correctly derived, and the empirical evaluation is broad, covering two model sizes, eight datasets, multiple demonstration selection and ordering strategies, and several shot counts. The paper explicitly acknowledges limitations, including the heuristic nature of the surprise-to-prior link and the need for labeled training data. However, the load-bearing premise is not formally guaranteed and the direct empirical evidence is only partially supportive, while the headline superiority claim is not uniformly supported by the reported tables. The method appears competitive and often strong, but the paper's framing overstates what is established.","major_comments":[{"comment":"The central theoretical link is explicitly heuristic: the paper states that low p(y_j | e_j, D_{j-1}) 'may suggest' lower E_z[p(e_j,y_j|z)] but that this 'is not formally guaranteed due to the normalization by p(e_j | D_{j-1})'. Since Eq. (8) constructs the surprise vector entirely from this unproven proxy, the derivation does not establish that surprise 'systematically covaries with prior adjustments.' The authors should either provide a formal condition under which the proxy is valid (for instance, bounded or slowly varying p(e_j|z)), or clearly reframe the contribution as an empirically motivated heuristic rather than a theoretically grounded signal.","section":"Section 3.1, Eq. (7)"},{"comment":"The empirical validation reported in Figure 2 is conditioned on the inserted label's class, and the text admits that 'aggregate correlations occasionally showed instability' due to an anti-recency bias. The conclusion that the results 'robustly support' the theoretical assertion is therefore too strong: the scatter plots demonstrate label-conditioned monotonicity, not a general, unconditional relationship between surprise and prior shift. The authors should report the aggregate correlations with appropriate statistical tests across both insertion classes, or explicitly qualify the claim as label-conditioned evidence for a heuristic.","section":"Section 3.2, Figure 2"},{"comment":"The abstract's claim that SC outperforms state-of-the-art ICL bias calibration baselines is not uniformly supported by Table 2. On Qwen2.5-3B, BC+ outperforms SC on SST-2 (90.33 vs 89.68), MNLI (79.07 vs 76.33), and QNLI (79.53 vs 78.01); on Qwen2.5-7B, LinC+ outperforms SC on QNLI (81.01 vs 80.48) and WiC (63.92 vs 62.93). The paper should report win rates over all dataset-model combinations and use paired significance tests or confidence intervals before claiming general superiority over the best baseline.","section":"Section 5.2, Table 2"},{"comment":"The ablation and effectiveness analyses do not isolate surprise as the operative mechanism. The ablation in Table 3 only removes surprise magnitude, retaining the sign and the full GRU sequence, so it does not test whether the surprise values themselves are necessary; a control with random features or label-only features is needed to show that the GRU is not simply exploiting label order and frequencies. In addition, removing magnitude improves accuracy on MNLI (77.02 vs 76.33 with magnitude), contradicting the claim that surprise magnitude plays a 'crucial role' in all cases. The high R^2 in Figure 8 between SC and BC calibrated ratios may merely reflect that both methods are derived from the same labeled data and approximate the same empirical class-conditional distribution, rather than demonstrating that surprise drives the adjustment.","section":"Section 5.4, Table 3 and Figure 8"}],"minor_comments":[{"comment":"The text refers to 'the BSC model' after introducing SC; this appears to be a typo for 'the SC model.'","section":"Appendix A.2"},{"comment":"The table header spells 'Verbalizer' as 'Verbizer'; please correct the typo.","section":"Table 4"},{"comment":"The reference to 'Tefnik and Kadlcik' appears to be a misspelling of 'Teplitsky' or the intended author name; please verify the citation.","section":"References"},{"comment":"The text defines the recorded surprise as '-log p(y|e,z)', but the formal definition in Eq. (4) uses p(y_j | e_j, D_{j-1}); using the same notation would avoid confusion.","section":"Section 3.2"},{"comment":"The first row appears as 'BCT' rather than 'BC T' or a properly spaced entry; please format the table consistently.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical breadth is a strength, and the authors are honest about the heuristic status of their central premise. However, the abstract and several section conclusions overstate the results relative to what Table 2 and Figure 2 actually show. The revision should focus on reframing the contributions as an empirical heuristic with demonstrated competitiveness rather than a theoretically guaranteed or uniformly superior method. If the authors provide the suggested control experiments and significance tests, the paper could become a solid contribution to ICL calibration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one if you work on ICL calibration. The new thing is Surprise Calibration: before each demonstration label is revealed, read the model's predicted probability for that label, take negative log as a 'surprise' value, then feed the sequence of signed surprise vectors into a small GRU that outputs a per-class prior adjustment. That is genuinely new relative to CC, BC, and LinC, which either use a fixed prior or pay a per-query inference cost. SC avoids that multiplicative cost, and the gains in Table 2 are consistent and often large (e.g., +8–14 points on MNLI with the 7B model). The paper is also unusually honest: Section 3.1 explicitly says the link between surprise and prior shift is 'not formally guaranteed', and Section 3.2 admits aggregate correlations are 'occasionally unstable'. The ablation in Table 3 shows surprise magnitude matters, and Figure 8 shows SC's calibrated ratios linearly track BC's, which is a reasonable sanity check.\n\nSoft spots. The central interpretive claim—that surprise systematically covaries with prior adjustment—is not established as a general property. Figure 2 conditions every correlation on the inserted label's class, and the aggregate instability is real. The stress-test note is right that the GRU is trained end-to-end on ground truth, so the accuracy gains in Table 2 do not validate the surprise-prior premise; they only show the features are useful to the GRU. That is a weaker claim, and the paper could have framed it that way. Also: the main results have no error bars or significance tests (only the ablation has standard deviations over three seeds), and no code is released. These are addressable and not fatal.\n\nBottom line: this is a practical calibration method with broad evaluation, not a theoretical breakthrough. The Bayesian derivation in Eqs 5–7 is correct as far as it goes; the surprise-prior premise is a heuristic that is only partially supported. A serious referee should engage. I would accept for review, ask for code and error bars, and push the authors to either soften the interpretive claim or test it more directly—for example, by measuring surprise against actual prior shifts across contexts without conditioning on label class.","headline":"A practical ICL calibration method that trains a GRU on signed surprise values to adjust class priors per query; the empirical gains are real, but the paper's theoretical framing outstrips the evidence.","tokens_in":25167,"tokens_out":1800,"would_cite":true,"duration_ms":19654,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that the surprise a language model registers just before a label is revealed — the negative log probability of that label under the current context — is a reliable signal for how much and in which direction its class…","keywords":["in-context learning","bias calibration","class prior shift","surprise signal","sequential Bayesian inference","large language models","dynamic demonstration selection","recurrent calibration"],"falsifier":"Repeat the paper's Figure 2 protocol on datasets where the input likelihood $p(e_j \\mid D_{j-1})$ varies widely across demonstrations; if the label-conditioned Spearman correlation between surprise and measured prior shift is not consistently positive and statistically significant, the surprise signal is not reliable and SC's input features would be noise.","tokens_in":24253,"feed_emoji":"😲","tokens_out":9435,"duration_ms":97589,"temperature":0.7,"pith_summary":"This paper tries to establish that surprise — the negative log probability a language model assigns to a demonstration's true label just before that label is revealed — is a usable signal for detecting class-prior shifts during in-context learning, and that a calibration method built on it beats fixed-prior calibration with fewer inference passes. The authors view in-context learning as implicit sequential Bayesian inference, in which each demonstration updates the model's belief about an unobserved task concept, with surprise amplifying the update. From this they build Surprise Calibration (SC), which converts each demonstration into a signed surprise vector and feeds the sequence through a recurrent network whose output adjusts the model's log-probabilities for each query. This matters because in dynamic in-context learning every query can have its own demonstration context, so a single fixed prior is the wrong tool, while per-query prior estimation is expensive; SC is designed to give the benefit of per-query priors without that cost.","feed_headline":"Surprise reveals when class priors shift in LLM prompts","feed_subtitle":"By reading each label's pre-reveal probability, Surprise Calibration beats fixed-prior methods on 8 NLP datasets at lower inference cost.","key_machinery":"The central object is the surprise vector, defined from the model's own pre-label predictive probability: for a demonstration $(e_j, y_j)$, entry $c$ equals $(1 - 2\\delta_{c,y}) \\log p(y=c \\mid e_j, D_{j-1})$, where the sign encodes the expected direction of prior adjustment for that class and the magnitude encodes how surprising the revealed label was. Collecting these vectors over the demonstration sequence gives a surprise sequence, and a gated recurrent unit processes that sequence to produce a cumulative per-class prior adjustment $a$, applied as $-\\log p_{\\mathrm{calib}}(y' \\mid e, z) = -\\log p_{\\mathrm{orig}}(y' \\mid e, z) - a_{y'}$. The machinery carries the argument because it converts a theoretical claim about Bayesian belief updates — surprise amplifies the covariance-driven shift in priors — into a trainable, context-dependent correction that requires no extra per-query sampling.","core_discovery":"The central claim is that the probability the model assigns to a label before seeing it, $p(y_j \\mid e_j, D_{j-1})$, tracks how the class prior should move: a low probability (high surprise) for the label that actually appears signals a larger shift, with the direction set by whether the label matches the class in question. In the paper's Bayesian reading, the update to the prior for a class $y'$ splits into a baseline expectation plus a covariance term between $p(y' \\mid z)$ and $p(e_j, y_j \\mid z)$, and a lower average joint likelihood amplifies that term. SC operationalizes this by forming, for every demonstration, a surprise vector whose $c$-th entry is $(1 - 2\\delta_{c,y}) \\log p(y=c \\mid e_j, D_{j-1})$, feeding the resulting sequence through a gated recurrent unit, and decoding the final hidden state into a per-class adjustment that is subtracted from the original log-probabilities. Empirically, the paper reports that SC improves on vanilla in-context learning by an average of +4.59 points on a 3-billion-parameter model and +3.54 points on a 7-billion-parameter model across eight datasets and six NLP tasks, matches or exceeds the best fixed-prior baseline in most settings, and needs only a single inference pass per query plus a small labeled training set rather than repeated per-query sampling. The paper also shows its calibrated probability ratios track batch-calibrated priors with $R^2 = 0.838$ and $R^2 = 0.977$ on two datasets.","pith_inferences":["Because surprise is read from the model's own next-token probabilities before labels are revealed, the same signal is available for any autoregressive language model that exposes per-token logits, so the method should transfer beyond the two model sizes tested here.","The paper's observation that anti-recency bias shifts mean priors while preserving the surprise-prior relationship suggests surprise could be used to separate recency effects from genuine prior shifts, something fixed-prior methods cannot do.","A testable extension is to use the accumulated adjustment not only to correct final predictions but also to re-rank retrieved demonstrations, since demonstrations that provoke high surprise may be the ones that most strongly move the model's priors.","The reported saturation of gains beyond 10 demonstrations implies the surprise sequence could be truncated once the GRU's adjustment stabilizes, giving an even cheaper variant that stops reading new demonstrations once priors stop moving."],"forward_implications":["Per-query class-prior calibration no longer requires repeated sampling: once trained, SC costs one forward pass per query plus a small labeled training set, instead of the multiplicative inference count of dynamic baselines.","Across eight datasets and two model sizes, SC outperforms vanilla in-context learning by an average of +4.59 and +3.54 accuracy points and matches or exceeds existing calibration baselines in most settings.","In 2-shot settings, SC can match or beat other calibration methods in 3-shot settings and vanilla ICL in 5-shot settings, meaning shorter demonstration contexts can be used at the same accuracy.","Surprise magnitude carries information: ablating it to sign-only reduces accuracy on most datasets, so the strength of the surprise, not just its direction, is needed for prior estimation.","SC's improvements persist across demonstration selection strategies, ordering strategies, verbalizer choices, and contexts with 1 to 15 demonstrations, indicating the adjustment is not tied to one prompt format."],"supporting_citations":[{"why":"Provides the implicit Bayesian inference interpretation of in-context learning that the paper's surprise-driven belief update builds on.","marker":"Xie et al. (2021)"},{"why":"Establishes that autoregressive language models predict token by token, which is what makes the pre-label surprise measurable.","marker":"Kossen et al. (2024)"},{"why":"Defines the contextual calibration baseline and the fixed-prior approach SC is designed to replace.","marker":"Zhao et al. (2021)"},{"why":"Supplies batch calibration, used both as a main baseline and as the prior estimator in the surprise-prior correlation experiments.","marker":"Zhou et al. (2023)"},{"why":"Introduces linear probe calibration, the closest trainable competitor that estimates priors from labeled data.","marker":"Abbas et al. (2024)"},{"why":"Frames query-specific demonstration selection, the dynamic in-context learning setting where fixed priors fail.","marker":"Liu et al. (2022)"}],"fun_headline_variants":["Surprise-based calibration adapts to class shifts in LLM prompts","Surprise signal recalibrates priors for in-context learning gains","Adaptive surprise calibration outperforms fixed-prior LLM methods","Surprise-driven prior adjustment boosts ICL accuracy cheaply"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that high surprise — the true label looking unlikely before it is revealed — reliably signals a larger class-prior shift, a link the paper explicitly concedes is not formally guaranteed because the predictive probability is normalized by $p(e_j \\mid D_{j-1})$.","fun_headline_variants_meta":{"raw":{"variants":["Surprise-based calibration adapts to class shifts in LLM prompts","Surprise signal recalibrates priors for in-context learning gains","Adaptive surprise calibration outperforms fixed-prior LLM methods","Surprise-driven prior adjustment boosts ICL accuracy cheaply"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000265,"raw_usage":{"total_tokens":1661,"prompt_tokens":1054,"completion_tokens":607,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":670,"completion_tokens_details":{"reasoning_tokens":534}},"tokens_in":670,"tokens_out":607,"duration_ms":6830,"temperature":1.0,"reasoning_tokens":534,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:42:11.157668+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the paper's Figure 2 protocol on datasets where the input likelihood $p(e_j \\mid D_{j-1})$ varies widely across demonstrations; if the label-conditioned Spearman correlation between surprise and measured prior shift is not consistently positive and statistically significant, the surprise signal is not reliable and SC's input features would be noise.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the implicit Bayesian inference interpretation of in-context learning that the paper's surprise-driven belief update builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that autoregressive language models predict token by token, which is what makes the pre-label surprise measurable."}],"review_version":1}