{"id":"04771d06-1184-46ab-9e6e-31561b691661","arxiv_id":"2411.18778","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":12,"one_line_summary":"Bayesian model comparison on six infrared-faint radio sources with far-infrared detections strongly favors an AGN torus component, indicating that these sources are AGN-dominated and likely low-luminosity counterparts of high-redshift radio galaxies.","lead":"Using infrared and radio observations of 20 distant galaxies, this paper models their light across many wavelengths and finds that infrared-faint radio sources are probably powered by actively accreting supermassive black holes. The work is a step toward understanding an extreme and rare class of high-redshift active galaxies and their link to the most massive galaxies in the universe.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Bayes-factor case for AGN is weakened because the 'no-AGN' model lacks warm dust/PAH emission, so strong Table 4 evidence may only show that some mid-IR component is present; selection bias is secondary.","rationale":"The reader's weakest assumption is the representativeness of the six Herschel-detected IFRSs. I agree that this is a real limitation, but I do not think it is the most load-bearing. Even if the six are representative, the central Bayesian comparison is not a clean AGN-vs-no-AGN test. The two-component model (SSP+GB) cannot emit substantial mid-infrared continuum or PAH features, while the three-component model includes the CLUMPY torus. At z~2.4-3.6, the observed 24 μm points probe rest-frame 5-8 μm, where star-forming galaxies show PAH features; a cold graybody at 26-50 K is negligible there. Thus the enormous Bayes factors in Table 4 may simply quantify the need for any warm mid-IR component, not specifically an obscured AGN torus. A no-AGN model with PAH/warm dust (Draine & Li, Dale & Helou, or a starburst template) would be the correct baseline. If such a model also fits, the claim 'IFRSs are AGN' is not established by this SED analysis; the paper would need to rely on q24 and HzRG analogy, which are supportive but more indirect. The authors do flag the brighter-end selection in Section 5.4 and the K-S p-values in Section 2.3; these should be incorporated into the main claim. Internal typos (Section 4.1 comparing to the 'SSP+Torus model'; Figure 5 caption saying 25 rather than 20) do not affect the argument. As a result, I would keep the conditional verdict but add the model-comparison check as a required revision, and recommend softening the abstract and summary wording until that check is done.","tokens_in":17276,"tokens_out":9779,"duration_ms":94698,"concrete_test":"Refit the six Herschel-detected IFRSs with a no-AGN model that includes SSP + cold graybody + a warm dust/PAH starburst component (e.g., Draine & Li 2007 or Dale & Helou templates) using identical photometry and priors, and recompute ln(ev_noAGN_warm/ev_AGN). If the torus model still gives ln BF > 5 for all six sources, the AGN interpretation survives; if not, Table 4 does not establish AGN-heated dust, and the claims should be restricted to 'the mid-IR excess requires a warm component, plausibly AGN.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the model comparison in Section 4.1/Table 4. The no-AGN alternative is SSP + cold graybody (Equation 4), and the cold graybody temperatures inferred in Table 5 (26-50 K) emit negligibly at rest-frame 6-8 μm, where the observed 24 μm points fall. The SSP component supplies only stellar photospheric emission; it does not include warm dust or PAH features. High-redshift star-forming galaxies commonly show strong 6.2/7.7 μm PAH emission that can reproduce the mid-IR excess without an AGN torus. Therefore the three-component CLUMPY-torus model can win with huge Bayes factors (ln BF = 62-6831) even if the mid-IR emission is starburst-powered. The paper itself compares with Arp220/I19254 templates in Section 5.2, showing that starburst templates are relevant, but they were not allowed as the no-AGN model. The Section 5.4 caveat about selecting the brighter end is real and would restrict the population claim, but the more fundamental issue is that the evidence does not uniquely identify AGN-heated dust. Independent q24 and HzRG comparisons support AGN activity, but the headline Bayesian claim needs a realistic no-AGN baseline.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a UV-to-infrared SED analysis of 20 infrared-faint radio sources (IFRSs) with spectroscopic redshifts, using the Bayesian SED-fitting code BayeSED V3.0. The authors model each SED with a three-component model (stellar SSP, AGN torus via CLUMPY, and cold graybody dust) and, for the six sources with Herschel photometry, compare the Bayesian evidence of this model with a two-component model (SSP plus cold graybody). They report enormous log-Bayes factors in favor of the AGN-torus model (ln BF = 62 to 6831, Table 4), and from this conclude that IFRSs are most likely AGN. They further derive IR luminosities, disentangle AGN and star-formation contributions, estimate star formation rates, and compare dust temperatures and IR-radio ratios with high-redshift radio galaxies (HzRGs), hot DOGs, and QSO samples. The paper also identifies a dichotomy in the median SEDs of the sample, with one group resembling Type-1 QSOs and another resembling AGN-starburst composites.","tokens_in":17471,"tokens_out":7086,"duration_ms":61435,"significance":"If the central conclusion is correct, the paper would establish IFRSs as AGN-dominated, high-redshift sources that are low-luminosity analogs of HzRGs, using an appropriate Bayesian model-comparison framework. The compilation of multiwavelength photometry for a spectroscopically confirmed sample and the use of a public, tested SED code are strengths. However, the headline Bayesian claim is weakened by the specific model set used: the no-AGN comparison model omits warm dust and PAH emission from star formation, so the overwhelming evidence may reflect only the need for a mid-IR component, not uniquely an AGN torus. In addition, the generalization from six Herschel-detected sources to the whole IFRS class is not well supported by the sample-selection statistics. The paper's other diagnostics (q24, comparison with HzRGs, SED templates) provide supporting but not conclusive evidence for AGN activity. The manuscript is worth publishing after the model-comparison interpretation is substantially revised or the no-AGN baseline is extended.","major_comments":[{"comment":"The two-component comparison model (SSP+GB) contains no warm-dust or PAH component, so the reported Bayes factors (ln BF = 62.4 to 6831.1) only demonstrate that a mid-IR component is required to fit the data, not specifically that the component is AGN-heated dust. The cold graybody temperatures inferred in Table 5 are 26-50 K, which emit negligibly at rest-frame 6-8 um, exactly where the observed 24 um points fall (rest ~7 um at z~2.4), and the SSP component supplies only stellar photospheric emission. High-redshift star-forming galaxies commonly exhibit strong 6.2/7.7 um PAH features that can reproduce such a mid-IR excess without a torus. The paper itself shows in Section 5.2 that starburst templates (Arp 220, I19254) are relevant to the IFRS population, but these were not allowed as the no-AGN model. I therefore do not regard the abstract's claim that the Bayesian evidence 'suggest[s] that IFRSs are most likely to be AGN' as uniquely supported. The authors should add a comparison model that includes warm dust and/or PAH templates (e.g., a starburst template or an additional warm-dust component) or temper the interpretation to 'require a mid-IR component', with the AGN identification supported by the ancillary q24 and HzRG comparisons rather than the Bayes factor alone.","section":"Section 4.1, Table 4"},{"comment":"The conclusion is stated for IFRSs as a class, but only 20 of 145 spectroscopic-redshift IFRSs have the two or more IRAC/WISE bands required for SED fitting, and only six of those have Herschel detections. The K-S tests reported in Section 2.3 give p=0.066 (redshift) and p=0.105 (radio-to-IR ratio), which are only marginally above the usual 0.05 threshold, and Section 5.4 explicitly concedes that the analysis 'targets towards the brighter end of IFRSs'. The abstract and summary should therefore qualify the AGN claim to the FIR-detected subsample, and the discussion should state more prominently that the extension to the full IFRS population is an assumption rather than a demonstrated result.","section":"Abstract and Section 5.4"},{"comment":"The abstract states that 'our sample is likely AGN-dominated', but the luminosities in Table 5 do not support that characterization for the majority of the six FIR-detected sources. For sources 122, 154, 160, and 161, L_SF_IR is equal to or greater than L_AGN_IR (e.g., source 122: log L_SF_IR = 12.71 vs log L_AGN_IR = 11.89; source 160: 12.39 vs 12.07); only sources 136 and 162 are clearly AGN-dominated. Summary point 3 says the contributions are 'approximately equal'. Please reconcile these statements with the tabulated values, or rephrase the conclusion to indicate that the sample spans a range from star-formation-dominated to AGN-dominated systems.","section":"Abstract and Table 5 / Summary point 3"}],"minor_comments":[{"comment":"In the paragraph after Table 4, the text says the SSP+Torus+GB model has higher evidence than the 'SSP+Torus model'; this should read 'SSP+GB model' (the model without the AGN component).","section":"Section 4.1"},{"comment":"The caption states 'of 25 IFRS', but the text and the rest of the paper say the sample contains 20 IFRSs; the number should be 20.","section":"Figure 5 caption"},{"comment":"The text says the derived dust temperatures (26-50 K) are hotter on average than those of ULIRGs, SMGs, and DOGs (20-50 K), while Summary point 4 calls the temperatures 'relatively low'. Please clarify the comparison baseline so the description is consistent.","section":"Section 5.4 and Summary point 4"},{"comment":"The statement that the SFR of IFRSs shows no correlation with AGN luminosity is based on only six sources; this should be presented as a null result with limited statistical power, rather than a strong observational constraint.","section":"Section 5.3"},{"comment":"When reporting the K-S test results, the text should explicitly state the sample sizes of the two distributions being compared (20 selected versus 145 parent) so the reader can assess the sensitivity of the p-values.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is likely to be accepted by readers as the first Bayesian proof that IFRSs are AGN, but the model-comparison design is too weak to carry that claim: the no-AGN alternative lacks warm dust and PAH emission, making the huge Bayes factors unsurprising and uninformative about the heating source. The authors have clearly done careful work on the data compilation and the Bayesian machinery, and the q24 and HzRG comparisons point in the AGN direction, so the paper is salvageable with a revised framing or additional models. I would also encourage the editor to request that the abstract be brought into line with the evidence: the current abstract overstates both the generality ('IFRSs') and the one-sided conclusion ('AGN-dominated') relative to the six-source subsample and the tabulated luminosity ratios."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThe headline: the paper applies a proper Bayesian SED comparison to a small sample of infrared-faint radio sources, but the no-AGN model is missing warm dust and PAH emission, so the huge Bayes factors don't uniquely identify AGN. The work deserves a referee, but the abstract and summary need recalibration.\n\nWhat's actually new: this is the first time BayeSED with a CLUMPY torus plus cold graybody has been used on IFRSs with spectroscopic redshifts, and the first quantitative decomposition of AGN versus star-formation IR luminosity for these sources. The Bayesian evidence comparison is done correctly, with enormous log Bayes factors (62 to 6831) favoring the three-component model for the six Herschel-detected sources. The derived IR luminosities place these sources at the low-luminosity end of HzRGs, which is a plausible connection. The K-S tests on redshift and radio-to-IR ratio are a good-faith attempt to quantify selection effects.\n\nThe main soft spot is the model comparison itself. The no-AGN model is stellar photosphere plus a cold graybody at 26-50 K. That graybody emits essentially nothing at rest-frame 6-8 micron, where the observed 24 micron points land, and the stellar component is also faint there. High-redshift starbursts show strong 6.2/7.7 micron PAH features that would fill that gap. So the Bayes factor may only show that a mid-IR component is present, not that an AGN torus is required. The paper compares the median SED to Arp220 and I19254, which include PAH-rich starburst emission, but those templates were not allowed as the no-AGN alternative. The independent q24 ratios and HzRG overlap do support AGN activity, but the headline 'most likely AGN' goes beyond the evidence comparison.\n\nThe second issue is the scope of the claim. The model comparison uses six IR-bright sources, and the K-S tests (p=0.066 and 0.105) indicate marginal representativeness. The authors concede the sample targets the brighter end, yet the abstract and summary say IFRSs are 'AGN-dominated.' Table 5 shows a mix: source 122 is star-formation-dominated in the IR, while 136 is AGN-dominated by an order of magnitude. Calling the sample AGN-dominated is a stretch.\n\nThere are also minor typos, including 'SSP+Torus' instead of 'SSP+GB' in Section 4.1 and '25 IFRS' in the Figure 5 caption.\n\nThe paper is useful for anyone working on IFRSs, HzRGs, or Bayesian SED modeling. The model comparison is a clean illustration of how the choice of alternative models drives the conclusion.\n\nRecommendation: send to peer review, but require the authors either to add a no-AGN model with warm dust/PAH components or to reframe the claim as evidence for a mid-IR excess, and to restrict the population-level claim to the FIR-detected subsample.","headline":"A careful Bayesian SED comparison on six IFRSs, but the no-AGN baseline lacks warm dust and PAH emission, so the claim that IFRSs are AGN rests on an incomplete alternative model.","tokens_in":18147,"tokens_out":5067,"would_cite":false,"duration_ms":43923,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Bayesian SED fits show infrared-faint radio sources are AGN.","keywords":["infrared-faint radio sources","active galactic nuclei","spectral energy distribution","Bayesian model comparison","AGN torus","high-redshift radio galaxies","dust-obscured galaxies","star formation rate"],"falsifier":"Take the IFRSs that currently lack far-infrared detections and observe them with a deeper far-infrared camera, or stack their Herschel/SPIRE images, then repeat the Bayesian model comparison: if their evidence ratios no longer favor the AGN torus, the claim that IFRSs are generally AGN would be refuted for the class.","tokens_in":16976,"feed_emoji":"🔭","tokens_out":12473,"duration_ms":96953,"temperature":0.7,"pith_summary":"This paper asks what infrared-faint radio sources (IFRSs) are: galaxies so bright at radio wavelengths and so faint in the infrared that their radio-to-infrared flux ratios reach thousands. The authors fit the ultraviolet-to-infrared spectral energy distributions of 20 IFRSs with spectroscopic redshifts and compare two physical models. For the six sources with far-infrared photometry, adding an AGN-heated dust torus to a stellar-plus-cold-dust model raises the Bayesian evidence by a logarithmic factor between 62 and 6831, far above the standard threshold for strong evidence. The paper concludes that IFRSs are most likely AGN-dominated systems and that their infrared luminosities place them as low-luminosity counterparts of high-redshift radio galaxies. If true, this identifies IFRSs as an early, dust-obscured phase of AGN and massive-galaxy growth rather than a purely starburst population.","feed_headline":"Bayesian SED fits show infrared-faint radio sources are AGN","feed_subtitle":"Bayesian evidence for an AGN torus links these rare galaxies to high-redshift radio galaxies","key_machinery":"The machinery is Bayesian SED model comparison carried out with the BayeSED code. Each source is fit by a three-component model: a simple stellar population with an exponentially declining star-formation history, a graybody for cold dust that re-emits absorbed starlight under an energy-balance assumption, and a CLUMPY clumpy-torus AGN component. Against this, a two-component model omits the torus. Bayesian evidence, rather than a simple goodness-of-fit statistic, is used to decide whether the extra AGN component is required; this penalizes the three-component model for its extra parameters, so the large positive Bayes factors are the load-bearing result. The q24 mid-infrared-to-radio ratio and the $T_{\\rm dust}$--$L_{\\rm IR}$ relation are then used to place IFRSs against comparison populations.","core_discovery":"On its own terms, the paper claims that infrared-faint radio sources are active galactic nuclei. For the six IFRSs in the sample with Herschel far-infrared detections, the authors compare a two-component model (stellar population plus cold dust) with a three-component model that adds a CLUMPY AGN torus. The logarithmic Bayes factor in favor of the AGN component ranges from 62.4 to 6831.1, all far above the commonly used strong-evidence threshold of 5. The paper concludes that IFRSs are AGN-dominated, that their infrared luminosities ($L_{\\rm IR}>10^{12}\\,L_\\odot$) make them low-luminosity counterparts of high-redshift radio galaxies, and that their median SED resembles an AGN-starburst composite in the infrared.","pith_inferences":["Beyond the paper, if IFRSs are young obscured AGN, multi-epoch VLBI monitoring should reveal compact jets whose morphology or flux changes on decade timescales.","Beyond the paper, the lack of an AGN-luminosity/star-formation correlation could be tested by placing IFRSs and high-redshift radio galaxies on a common luminosity--SFR plane to see whether they trace a single evolutionary sequence.","Beyond the paper, deeper far-infrared observations of the IFRSs without Herschel detections would show whether the AGN-dominated conclusion extends to the infrared-fainter population or whether those sources form a separate, less dusty class."],"forward_implications":["IFRSs are AGN-dominated rather than extreme starbursts, so their extreme radio-to-IR ratios are a signpost of obscured AGN activity.","The six far-infrared-detected IFRSs have total infrared luminosities above $10^{12}\\,L_\\odot$ and occupy the low-luminosity extension of high-redshift radio galaxies.","Their star formation rates lie in the range 100--900 $M_\\odot\\,\\mathrm{yr}^{-1}$, and AGN luminosity shows no correlation with star formation rate in this sample.","The median SED splits into a Type-1-QSO-like group and an AGN-starburst-composite-like group, implying IFRSs are not a single homogeneous population.","Cold dust temperatures of 26--50 K with small effective radii suggest that the growth in infrared luminosity is driven by dust temperature rather than dust mass."],"supporting_citations":[{"why":"Defines the modified IFRS selection criteria and supplies the starting catalog of IFRSs with spectroscopic redshifts.","marker":"Collier et al. (2014)"},{"why":"Provides the spectroscopic redshifts for most of the 20 sources modeled.","marker":"Orenstein et al. (2019)"},{"why":"Supplies the CLUMPY clumpy-torus SED library used as the AGN component in the three-component fits.","marker":"Nenkova et al. (2008a,b)"},{"why":"Provides the Bayesian evidence model-comparison framework and the Occam-razor criterion used to justify preferring simpler models unless evidence is strong.","marker":"Han & Han (2019)"},{"why":"Supplies the BayeSED V3.0 code used for SED generation, posterior sampling, and evidence computation.","marker":"Han et al. (2023)"},{"why":"Supplies the Type-1 and Type-2 QSO, Arp 220, and I19254 templates used to classify the median and subgroup SEDs.","marker":"Polletta et al. (2007)"},{"why":"Provides the high-redshift radio galaxy comparison sample and the q24-redshift locus used to separate AGN from star-forming galaxies.","marker":"Seymour et al. (2007)"},{"why":"Earlier SED modeling that tied IFRSs to obscured high-redshift radio galaxy templates, which this paper extends with spectroscopic redshifts.","marker":"Herzog et al. (2015b)"},{"why":"Provides the high-redshift radio galaxy infrared luminosity and star-formation-rate comparison sample for the luminosity decomposition diagram.","marker":"Drouart et al. (2014)"},{"why":"Calibrates the Bayes-factor threshold of ln(evidence ratio) > 5 used to declare strong evidence for the AGN component.","marker":"Trotta (2008)"}],"fun_headline_variants":["Strong Bayesian support for AGN in infrared-faint radio sources","AGN torus favored by Bayesian evidence in infrared-faint radio sources","Bayesian SED analysis reveals AGN dominance in infrared-faint radio sources","Infrared-faint radio sources are AGN-dominated per Bayesian SEDs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument's load-bearing premise is that the six far-infrared-detected IFRSs represent the broader IFRS population, even though only 20 of 145 sources had enough infrared bands for SED fitting and the six with Herschel detections are the infrared-brightest cases.","fun_headline_variants_meta":{"raw":{"variants":["Strong Bayesian support for AGN in infrared-faint radio sources","AGN torus favored by Bayesian evidence in infrared-faint radio sources","Bayesian SED analysis reveals AGN dominance in infrared-faint radio sources","Infrared-faint radio sources are AGN-dominated per Bayesian SEDs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0011,"raw_usage":{"total_tokens":4621,"prompt_tokens":1011,"completion_tokens":3610,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":3531}},"tokens_in":627,"tokens_out":3610,"duration_ms":119689,"temperature":1.0,"reasoning_tokens":3531,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:53:33.958863+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the IFRSs that currently lack far-infrared detections and observe them with a deeper far-infrared camera, or stack their Herschel/SPIRE images, then repeat the Bayesian model comparison: if their evidence ratios no longer favor the AGN torus, the claim that IFRSs are generally AGN would be refuted for the class.","supporting_citations":[],"review_version":1}