{"id":"e30bb1f6-08b1-4523-8f23-4586888b0487","arxiv_id":"2607.19355","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"LLMs update their stated numeric beliefs almost regardless of source reliability or whether a claim moves them closer to the truth, performing near chance on both dimensions.","lead":"This paper measures whether LLMs update their answers more when information comes from reliable sources and when claims move them closer to the truth; across 13 models and nearly 670,000 trials, they mostly don't. The authors release a benchmark and a user survey showing people expect these behaviors, making it a testbed for a core AI alignment property.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline near-chance discernment may be partially an artifact of treating single numeric tokens as stable beliefs; P1/P2 lack test-retest or calibration evidence.","rationale":"The reader's weakest assumption—that single numeric outputs are treated as faithful beliefs without calibration/test-retest support—is the most load-bearing threat to the central claim. If P1/P2 are unstable, then Δ is contaminated by measurement error, attenuating the source and truth discernment correlations toward zero and making the balanced 'appropriate/inappropriate' updates in Table 3 look like indiscriminate updating when the model may simply be guessing. I agree with the reader's identification. I considered alternative concerns (e.g., the chance baseline for ρ(Δ,I) under a fixed-fraction-update null is not obviously 0; PC1 reliability is claim-independent), but the probe-stability issue is more fundamental because it undermines the dependent variable itself. The paper has independent support that partially mitigates this: the Reliable-prompt control shows ρ(Δ,R_LLM)=0.46, demonstrating that the probe can detect systematic updating under a different prompt, and the high Correct Defense values show models are stable when they know the answer. These internal controls keep the concern from being fatal, but the baseline measurement remains unvalidated. A concrete test-retest experiment would settle whether the near-chance findings are real or an artifact of output noise. Therefore the reader's CONDITIONAL verdict is appropriate; no change is needed. The paper is transparent about limitations and provides substantial scale and a pre-registered user study, so unconditional acceptance is not yet warranted, but rejection is too harsh given the internal controls.","tokens_in":31990,"tokens_out":10584,"duration_ms":129723,"concrete_test":"Run a test-retest experiment on a random subset (e.g., 500 trials per model × 13 models). Elicit P1 twice without any claim and P2 twice with the same source/claim. Compute (a) mean absolute difference |P1_1−P1_2| and |P2_1−P2_2|, and (b) intraclass correlation. If these baseline differences are comparable in size to the mean Δ from the main experiment, or if recomputing source/truth discernment using the mean of repeated P1/P2 shifts the near-zero values by more than ~0.05, the headline finding is at least partly a probe artifact. If the near-zero correlations persist with averaged beliefs, the concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 and Appendix F.1-F.2 elicit P1 and P2 as a single numeric token ('Give a numeric answer and nothing else'). All four headline metrics (source discernment, truth discernment, reliability propensity, truth propensity) derive from Δ=|P1−P2|. The central claim—that models update roughly equally for reliable/unreliable sources and for helpful/harmful claims—presupposes that P1 and P2 are faithful, stable point estimates of the model's belief. The paper does not report calibration, test-retest reliability, or a control condition in which the same question is asked twice without a source/claim. On uncertain numeric items (common in this benchmark), an LLM can emit different plausible numbers across calls, so Δ would partly measure output instability rather than belief revision. This measurement error attenuates ρ(Δ, S) and ρ(Δ, I) toward zero and would also make the 'appropriate vs inappropriate update' split in Table 3 appear near 50/50. The Reliable-prompt control (ρ(Δ,R_LLM)=0.46, Figure 6) shows the probe can detect systematic updating under a different prompting regime, which mitigates the concern but does not validate the baseline P1/P2. Without a stability check, the near-chance baseline result is not separable from probe noise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a normative framework, 'information discernment', with three axioms: source discernment (update more for higher-reliability sources), truth discernment (update more when a claim moves the prior closer to ground truth), and correct defense (do not abandon a correct prior). The authors introduce Learn2Discern (L2D), a benchmark combining numeric-answer questions from existing and new datasets with 132 real web sources stratified by reliability and popularity, and claims generated by multiplicative perturbations of the ground truth. They report experiments across 13 LLMs and nearly 670K trials, finding near-chance source and truth discernment, stronger sensitivity to popularity than reliability, no consistent improvement in source discernment with scale/recency, and some benefit from inference-time prompting interventions. A pre-registered, quota-matched user study (n=299) is used to validate that LLM users endorse the axioms and report that violations reduce trust and usage intent. The paper releases the dataset, prompts, and survey.","tokens_in":32345,"tokens_out":7598,"duration_ms":91077,"significance":"If the central empirical claim holds, the paper identifies a practically important blind spot in current LLMs' use of external information, with direct relevance to RAG and misinformation propagation. The contribution is potentially significant: the benchmark is large and diverse, the measurement framework is simple and interpretable, the user study anchors the metrics in stated user preferences, and the data are released for reuse. The scale (13 models, 670K trials, 95% CIs), the pre-registration of the user study, and the explicit attention to contamination via novel GSS/World Bank items are clear strengths. However, the headline 'near chance' conclusion rests on an unvalidated assumption that a single numeric token elicited from an LLM is a stable, faithful point estimate of its belief. The absence of a test-retest or calibration control means the central null result is not yet separable from measurement noise. The paper also makes several categorical claims that are not fully supported by its own tables.","major_comments":[{"comment":"The core metrics all derive from Δ = |P1 − P2|, where P1 and P2 are single numeric tokens elicited by 'Give a numeric answer and nothing else.' The paper gives no test-retest reliability, no calibration evidence, and no control condition in which P1 and P2 are elicited on the same question without a source/claim. On uncertain numeric items, output instability will contribute to Δ; this attenuates ρ(Δ, S) and ρ(Δ, I) toward zero and also corrupts the taxonomy in Table 3 (a noisy P2 would turn 'correct hold' into 'sycophantic cave'). The Reliable-prompt control (Figure 6, ρ(Δ, R_LLM)=0.46) shows that the probe can detect a signal under a different prompting regime, but it does not validate the baseline P1/P2. Please add a no-information stability control and, if possible, a calibration check, or apply a measurement-error correction before reporting near-chance discernment.","section":"§5 (Algorithm 1), §6 (Δ definition), Appendix F.1–F.2"},{"comment":"The claim that 'all models perform near chance across all metrics except Correct Defense' is contradicted by Table 2. For example, Claude-3.5-Sonnet has truth discernment −0.25, Gemini-2.0-Flash −0.07, and GPT-3.5-Turbo −0.043. A Spearman correlation of −0.25 is not 'near chance'; it indicates systematic updating away from the truth. Conversely, source discernment values around 0.02–0.06 may be statistically nonzero at N≈670K, so 'near chance' needs a formal justification. Please report equivalence tests (e.g., TOST with pre-specified bounds) or reframe the headline as 'small / inconsistent' rather than 'near chance'.","section":"§9, Table 2"},{"comment":"The text states that bigger and newer models 'did not improve on reliability-related metrics (source discernment and reliability propensity).' This is not supported by the authors' own tables. Table 11 shows gpt-4o-mini→gpt-4o source discernment +0.028 (p<.001) and gpt-4.1-mini→gpt-4.1 +0.022 (p<.001); Table 12 shows gpt-3.5-turbo→gpt-4o +0.032 (p<.001). The pattern is mixed rather than absent. If the intended claim is 'no *consistent* or *practically meaningful* improvement', that should be stated with an effect-size criterion, and the abstract's 'not source discernment' should be softened accordingly.","section":"§9, Appendix J (Tables 11–12)"}],"minor_comments":[{"comment":"The CoT prompt contains a template artifact: 'FINAL_ANSWER: Give a numeric answer and nothing else.' is fine, but the RATIONALE line includes '$produce' which appears to be a LaTeX/formatting error. Please fix.","section":"Appendix F.3"},{"comment":"The captions are garbled: the same '(a) OECD vs Non-OECD' and '(b) Pre- vs Post-Internet' labels appear under both panels, while the text refers to 'update quality ratio' and 'median relative prior error.' Relabel the subfigures so they match the text.","section":"Figure 4"},{"comment":"The 'Direction' column uses Good/Neutral/Bad but the definitions of 'Appropriate Update' vs 'Inappropriate Update' depend on the prior being wrong; this is clear in the descriptions, but adding a sentence on how ties are treated (e.g., when P2 equals P1 within tolerance) would improve precision.","section":"Table 3"},{"comment":"Typo: 'Proceedure' should be 'Procedure'. Also, the survey wave handling for Source Discernment (Wave 2 only) is described; consider adding one sentence in the main text to preempt concerns about using a subset of waves for one of the three axioms.","section":"Appendix I.2"},{"comment":"The paper already acknowledges that source credibility may be claim-dependent; this is important because the source-discernment metric uses a claim-independent PC1 reliability label. I suggest adding a sentence in the discussion noting that the near-zero source discernment should be interpreted relative to this global, topic-independent notion of reliability.","section":"§10, Limitations"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the unvalidated P1/P2 probe. If the authors can supply a no-information stability control and a calibration check (or a convincing measurement-error argument), I would be willing to support acceptance after the 'near chance' and scale/recency claims are made precise. The negative-existence claim in the abstract is stronger than the evidence currently supports; the paper would be stronger as a benchmark-and-measurement contribution with a cautiously worded empirical summary. No concerns about authorship or code/data provenance were apparent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this paper deserves a serious referee. It gives the field a concrete framework and benchmark for information discernment, and the headline finding — models weight popularity over reliability and update about equally for truth-improving and truth-worsening claims — is real enough to engage with. The reader's main worry about single-token belief probes is on target, but it is partly mitigated by the paper's own controls, so I would frame it as a required addition rather than a fatal flaw.\n\nWhat's new: L2D formalizes source and truth discernment with continuous reliability and continuous falsity, which is a real step beyond the binary correct/incorrect framing in prior knowledge-conflict work. The pre-registered user study (n=299) grounding the axioms in actual user expectations is a nice touch that most benchmarks skip. The empirical scale is solid: 13 models, 670K trials, 95% CIs, power analyses, novel longitudinal data to dodge contamination, and a QA robustness check (r=0.99). The M1/M2 decomposition is genuinely useful: models can rate source reliability moderately well (ρ=0.41) but fail to apply it in updates (ρ=0.46 vs PC1), which separates estimation failure from application failure. The popularity-vs-reliability ablation and the size/recency asymmetry (truth discernment improves, source discernment doesn't) are new findings.\n\nSoft spots: the biggest is the one the stress-test flagged. Algorithm 1 reads a single numeric token as P1/P2, and all four headline metrics derive from Δ=|P1−P2|. No test-retest or calibration check is reported. On uncertain numeric items, single-token outputs may be noisy, which would attenuate the discernment correlations toward zero. However, the Reliable-prompt control (ρ(Δ,R_LLM)=0.46) shows the probe can detect systematic updating, and the popularity result shows Δ responds to source features, so pure noise is not a convincing explanation for the near-chance baseline. Still, the paper would be much stronger with a stability analysis — e.g., eliciting the same prior twice without a source, or reporting variance across decoding temperatures. The PC1 claim-independence issue is real but the authors acknowledge it and restrict to news sources, so I'd call it minor. The 'near chance' phrasing overstates a bit: some correlations are small but statistically positive, and the popularity effect shows models do respond to something. That's a framing issue, not a substantive one. The Wave 2-only source-discernment result is disclosed and consistent across waves; fine.\n\nWho it's for: anyone working on RAG safety, knowledge conflicts, or LLM trustworthiness. The benchmark is a contribution even if the headline result needs tightening. If I were the editor, I'd send it to review with a request for the stability analysis.","headline":"A genuinely useful benchmark for source/truth discernment in LLMs; the near-chance headline is plausible but needs one stability analysis before it is fully secure.","tokens_in":32820,"tokens_out":2548,"would_cite":true,"duration_ms":33218,"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":"Current LLMs update beliefs as much for unreliable sources as for reliable ones, and as much for claims that move them away from the truth as toward it.","keywords":["information discernment","source discernment","truth discernment","knowledge conflicts","retrieval-augmented generation","LLM belief updating","misinformation","benchmark"],"falsifier":"Run a version of the experiment where prior and posterior are elicited as 10 independent samples (or as full probability distributions) for a subset of trials; if average update magnitude then correlates strongly with source reliability (say ρ > 0.2), the near-zero correlation in the single-token probe would be an artifact, not a genuine absence of source discernment. Conversely, finding any unprompted model that scores above 0.2 on source discernment across datasets would directly refute the claim that all models perform near chance.","tokens_in":1322,"feed_emoji":"🤖","tokens_out":2266,"duration_ms":61738,"temperature":0.7,"pith_summary":"This paper asks whether LLMs treat external information the way a careful reader would: updating more for reliable sources and more when a claim moves them closer to the truth. Across 13 models and nearly 670K trials, it finds models do neither: update magnitude correlates near zero with source reliability and with truth improvement, and models actually weight source popularity twice as strongly as reliability. A pre-registered user study of 299 LLM users confirms that people endorse these norms and say violations would reduce their trust and usage intent. The paper releases a benchmark for measuring this 'information discernment' and shows that simple prompt modifications can partially fix the deficit.","feed_headline":"LLMs update for popularity, not reliability, in 670K trials","feed_subtitle":"A new benchmark shows near-chance sensitivity to source quality across 13 models; simple prompts partially fix it.","key_machinery":"The experimental machinery is a before/after probe: ask a model a numeric question to get a prior P1, present a claim C = T × f from a source S with known reliability and popularity, elicit posterior P2, and define update magnitude Δ = |P1 − P2|. Source discernment is the Spearman correlation ρ(Δ, reliability); truth discernment is ρ(Δ, I) where I = |P1 − T| − |C − T| is how much the claim improves on the prior. The framework also defines a 'reliability propensity' probability and a 'correct defense' stickiness measure, and a decomposition of low source discernment into mis-estimating reliability vs failing to apply it (LLMs estimate reliability moderately, ρ=0.41, but applying their own est","core_discovery":"The paper's central claim is that current LLMs systematically fail at information discernment: when told '[Source] says the answer is C', models shift their numeric answers just as much for a fringe outlet as for a highly reliable one, and just as much when the claim moves them toward the ground truth as when it moves them away. The exception is 'correct defense': models usually stick with an initially correct answer. Bigger and newer models show better truth discernment but no better source discernment, and models integrate external information best on datasets where their priors are already accurate. The failures are measured with Spearman correlations between update magnitude and reliabil","pith_inferences":["A direct testable implication is that near-zero source discernment is not a hard ceiling: since models can rate reliability moderately well when explicitly asked, fine-tuning or prompting strategies that bind reliability judgments to update sizes may close the gap.","The probe uses a single numeric token as the model's belief; if that token is noisy under knowledge conflict, the true belief-update distributions might be more discerning than reported—re-running with full probability distributions or repeated sampling would settle this.","The popularity-over-reliability result suggests a 'trust equals fame' shortcut learned from pretraining; if so, interventions that break the popularity–reliability confound in training data (not just prompts) may be needed to fix source discernment.","Because the design isolates belief updating from retrieval, the results bound the best-case performance of RAG: retrieval improvements alone cannot correct failures at the updating stage."],"forward_implications":["If the central claim is right, retrieval-augmented systems built on today's LLMs can be led astray by low-quality sources, since belief updating does not track source quality and instead tracks popularity.","Scaling model size and recency will not by itself fix source discernment: the paper finds no improvement on that axis across matched pairs of bigger/newer models.","Simple inference-time instructions—explicitly rating source reliability, reminding the model to defend a correct prior, chain-of-thought—improve facets of discernment and are cheap, actionable mitigations.","External knowledge integration is most effective where it is least needed: models update well from evidence only on datasets where their priors are already accurate, limiting how much web augmentation can equalize knowledge gaps.","If the user study's stated norms reflect real adoption behavior, low discernment is likely to erode trust and usage of LLM-based search as these systems replace traditional search."],"fun_headline_variants":["LLMs can't tell reliable sources from unreliable ones","LLMs update for popularity, not truth, in 670K trials","Source quality ignored by LLMs: near-chance discernment","New benchmark: LLMs fail to weigh source reliability"],"cache_read_input_tokens":34176,"weakest_assumption_plain":"The experiment treats the single numeric token a model emits when asked 'Give a numeric answer and nothing else' as its genuine belief, before and after the claim; if those tokens are noisy or arbitrary under knowledge conflict, the near-zero discernment correlations would be an artifact of the probe rather than proof that models fail to discriminate.","fun_headline_variants_meta":{"raw":{"variants":["LLMs can't tell reliable sources from unreliable ones","LLMs update for popularity, not truth, in 670K trials","Source quality ignored by LLMs: near-chance discernment","New benchmark: LLMs fail to weigh source reliability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1571,"prompt_tokens":747,"completion_tokens":824,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":764}},"tokens_in":491,"tokens_out":824,"duration_ms":9866,"temperature":1.0,"reasoning_tokens":764,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T13:19:07.534499+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a version of the experiment where prior and posterior are elicited as 10 independent samples (or as full probability distributions) for a subset of trials; if average update magnitude then correlates strongly with source reliability (say ρ > 0.2), the near-zero correlation in the single-token probe would be an artifact, not a genuine absence of source discernment. Conversely, finding any unprompted model that scores above 0.2 on source discernment across datasets would directly refute the claim that all models perform near chance.","supporting_citations":[],"review_version":1}