{"id":"1d4ed31c-0f40-4c40-b5dc-d2e802a31b23","arxiv_id":"2412.08317","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"GPT-3.5 with chain-of-thought prompting still fails often on multi-hop reasoning with external knowledge, especially with distractors, counterfactual facts, non-sequential proof structures, and higher hop counts.","lead":"This paper tests GPT-3.5 on four question-answering benchmarks and finds that the model still makes frequent errors when a task needs several pieces of knowledge, when reasoning is not a straight line, and when the number of steps grows. The experiments map where current large language models fail even when they are prompted to think step by step.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The non-sequential reasoning conclusion rests on an unvalidated proxy; a direct human or model comparison on the same samples would settle whether the depth-length criterion isolates genuine structural difficulty.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the depth-length mismatch proxy for non-sequential reasoning is unvalidated and underpins the Section 5 claim. I agree with the reader's overall conditional verdict. The paper's core qualitative findings are consistent with existing literature and are supported by multiple independent experiment families: external knowledge improves accuracy (Tables 2-3), distractors degrade knowledge selection and accuracy in a graded way, counterfactual knowledge causes measurable confusion (Tables 4-5), and higher-hop exemplars do not produce consistent hop-wise generalization (Table 6). These findings do not collapse even if the non-sequential proxy is flawed, because they are based on direct comparisons of conditions. However, the Section 5 conclusion is a distinct and headline claim, and the paper's own limitations section acknowledges the small sample sizes and single model, which further increases the importance of validating the proxy. The paper also overclaims a 'large gap with humans' without a human baseline for most tasks (only HotpotQA human accuracy is cited, in a footnote), which supports the reader's suggestion to temper that phrase. My concrete test targets the proxy directly: matching on confounds (proof length, premise count, hop count) and/or human annotation would determine whether the 56% vs 78% gap reflects non-sequential structure or simply harder/longer proofs. Thus, CONDITIONAL is the appropriate verdict: the central claim is defensible, but the non-sequential claim needs validation and the human-gap phrasing needs softening.","tokens_in":23856,"tokens_out":2139,"duration_ms":19902,"concrete_test":"Construct two matched subsets from EntailmentBank with equal distributions of proof length, number of premises, and number of hops, one selected by the depth-length mismatch criterion and one with depth equal to length, then rerun the Section 5 experiment on these matched subsets. If the accuracy gap between the two subsets disappears or shrinks dramatically, the depth-length criterion is not isolating non-sequential structure and the Section 5 conclusion should be weakened. A complementary check: have human annotators independently label a random sample of EntailmentBank proofs as sequential vs non-sequential, and measure agreement between human labels and the depth-length criterion.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central qualitative claim — that LLMs 'still suffer from severe drawbacks' in multi-hop reasoning — is broadly consistent with prior work and is reasonably supported by the overall accuracy declines in the distractor and hop-count experiments, but the section explicitly dedicated to non-sequential reasoning (Section 5) rests on a fragile operationalization. The paper selects 'non-sequential' EntailmentBank samples using the criterion that 'depth and length of the entailment tree are not the same' (Section 5). This conflates structural non-sequentiality with other properties that correlate with tree depth, such as total proof length, number of premises, or answer complexity. The reported comparison is 56% accuracy on the selected samples versus 78% on the remaining (sequential) samples; if the selected samples simply contain more premises or longer proofs, the accuracy gap could reflect quantity of reasoning steps rather than non-sequential structure. The paper provides no validation of the proxy (e.g., against human judgments, against an independent structural taxonomy, or against a matched control set), and the few illustrative prompts in Appendix A.2 do show branching/combining proofs, but anecdotes do not establish that the depth-length mismatch reliably separates sequential from non-sequential reasoning across the dataset. Because the Section 5 claim that 'CoT fails on non-sequential reasoning' is one of the three headline contributions, this unvalidated proxy is the most load-bearing weakness. The other claims (external knowledge helps, distractors hurt, counterfactual knowledge confuses, limited hop generalization) are each supported by direct comparisons and do not depend on this proxy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports experiments with text-davinci-002 on four multi-hop reasoning benchmarks (HotpotQA, EntailmentBank, QASC, bAbI15) under Chain-of-Thought prompting and its variants. The experiments manipulate the availability of external knowledge, the proportion of distractor sentences, the factual consistency of the provided knowledge, the structural type of reasoning (sequential/non-sequential), and the number of hops in the prompt exemplars. The central empirical claims are that external knowledge helps but distractors and counterfactual knowledge confuse the model, that accuracy drops on selected non-sequential entailment samples, and that least-to-most prompting does not improve hop-count generalization.","tokens_in":24072,"tokens_out":6769,"duration_ms":71403,"significance":"If the results hold, this is a useful negative result: it documents that knowledge selection, counterfactual robustness, and non-sequential reasoning remain weak points for CoT-prompted LLMs, consistent with but more comprehensive than existing findings. The study's strengths are its four-benchmark coverage, the transparent prompt templates in the appendix, and the manual error analysis that identifies specific failure modes such as word bias and hallucination. However, the paper's quantitative support is currently limited by small single-model samples and by an unvalidated structural proxy, so the contribution is exploratory and would need robustness work to serve as a definitive claim.","major_comments":[{"comment":"The paper identifies 'non-sequential' samples by the criterion that 'depth and length of the entailment tree are not the same.' This criterion is not validated, and the two groups may differ in proof length, number of premises, or answer complexity rather than in structural non-sequentiality alone. The reported accuracy gap (56% vs. 78%) is therefore not unambiguously attributable to non-sequential reasoning. Please provide a validation of the proxy (e.g., human-annotated structural labels, matched control sets, or an analysis showing that proof length is not a confound); the examples in Appendix A.2 illustrate the intended phenomenon but do not establish that the criterion separates the classes reliably across the dataset.","section":"Section 5, 'Reasoning in Non-Sequential Cases'"},{"comment":"Every experiment uses 100 test samples and a single model (text-davinci-002), with no confidence intervals, bootstrap estimates, or significance tests, and the sampling procedure for the 100 samples is not described. Differences such as 83% vs. 74% in Table 2 and 64% vs. 46% in Table 6 could be within sampling noise, especially because decoding is stochastic and correctness is checked manually. The abstract's plural 'LLMs' claim also goes beyond the single evaluated model. Please add uncertainty quantification and either replicate on at least one other model or restrict the claims to GPT-3.5 as evaluated here.","section":"Section 3 and Tables 2-6"},{"comment":"The manual evaluation is not sufficiently specified to be reproducible. The paper reports that 'we check each sample manually' and that 'none of the generated reasoning paths is correct' in the non-sequential experiment, but no annotation protocol, error taxonomy, second annotator, inter-annotator agreement, or released annotations are provided. This matters because the qualitative conclusions about error types and about the absence of coherent reasoning paths carry as much weight as the accuracy numbers. Please make the rubric and the per-sample annotations available.","section":"Section 5 and Footnote 1"}],"minor_comments":[{"comment":"Clarify whether text-davinci-002 is the GPT-3.5 model referenced in the abstract, since the naming is potentially confusing.","section":"Section 3"},{"comment":"Define 'depth' and 'length' of an entailment tree explicitly; these terms are used without formal definitions.","section":"Section 5"},{"comment":"The second value in each cell is a raw count (out of 50) rather than a percentage, while the first value is a percentage; this mixed format makes the table hard to read.","section":"Table 5"},{"comment":"There are typos in the prompts ('posive', 'lind', 'invisble') that should be corrected.","section":"Appendix A.2"},{"comment":"The figure caption does not state the number of generated reasoning paths per setting, which makes the length distributions hard to interpret.","section":"Section 6 and Figure 2"},{"comment":"The paper would benefit from a brief statement on data and code availability; the appendices provide prompts but not the test subsets or evaluation code.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The authors should be encouraged to release test subsets, annotations, and evaluation code. The Section 5 proxy should be validated or the corresponding conclusion softened. The journal should also decide whether single-model, 100-sample experiments meet its empirical standards for claims about 'LLMs' as a class."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the take: this is a small, honest empirical study showing that text-davinci-002 still struggles with multi-hop reasoning under CoT prompting, and that distractor noise and counterfactual context hurt. The central qualitative claim is defensible and consistent with prior work. The load-bearing weakness is Section 5: the non-sequential reasoning conclusion rests on an unvalidated proxy (tree depth != path length), which could just be picking out longer or more premise-heavy proofs.\n\nWhat the paper does well: the setup is straightforward, the prompts are in the appendix, and the error analysis (word bias, hallucination, over-decomposition) is genuinely useful. The distractor experiments on HotpotQA and EntailmentBank with EM/OneHit and precision/recall are a nice way to separate selection ability from answer correctness. The counterfactual bAbI15 design, where prompt and test factualness are crossed, is a clean 2x2 that shows prompt knowledge bleeding into reasoning. These are solid contributions for a replication-focused study.\n\nThe main problem is the proxy. The paper defines non-sequential as depth != length with no validation. The reported 56% vs 78% could be driven by proof length or number of premises. I'd want to see a matched control set, or at least a human judgment on a sample of items. Second, every experiment uses 100 samples, one model, manual evaluation, and no significance tests. The paper acknowledges this in Limitations, so it's not hidden, but it does mean the exact numbers shouldn't be over-read. Third, the 'large gap with humans' line cites published human accuracy on HotpotQA (83.6%) but doesn't run humans on the same samples, so it's a rough comparison.\n\nWho is this for? Practitioners building retrieval-augmented QA systems who want another data point on GPT-3.5-era limitations, and methodologists working on evaluating reasoning in LLMs. It's not going to change any research direction, but it's a fair, clear-eyed empirical note.\n\nI'd send it to review rather than desk-reject. The central claim is sound, and a good referee could push the author to validate the non-sequential proxy, add a second model, and report confidence intervals. As is, it's a conditional accept or a workshop-level paper.","headline":"A small, honest empirical study confirming GPT-3.5's fragility in multi-hop reasoning; the non-sequential reasoning claim needs better support than the depth-length proxy.","tokens_in":24622,"tokens_out":2615,"would_cite":false,"duration_ms":27341,"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":"Even with chain-of-thought prompting, GPT-3.5 still fails to reliably select and combine external knowledge, handle non-sequential reasoning structures, or generalise across numbers of hops.","keywords":["multi-hop reasoning","Chain-of-Thought prompting","external knowledge","counterfactual reasoning","non-sequential reasoning","least-to-most prompting","GPT-3.5","reasoning generalization"],"falsifier":"Take the EntailmentBank samples the paper labels non-sequential (depth differs from length), have independent human annotators judge whether the reasoning is genuinely non-sequential, and compare GPT-3.5's accuracy on those against a matched set of sequential samples of equal proof length; if most labelled samples are judged sequential, or if a newer model with explicit non-sequential prompting shows no accuracy gap, the central claim is refuted.","tokens_in":23602,"feed_emoji":"🧩","tokens_out":6233,"duration_ms":50823,"temperature":0.7,"pith_summary":"This paper argues that even the best-known prompting technique, chain-of-thought, does not give large language models reliable multi-hop reasoning when the knowledge must come from outside the model. Using GPT-3.5 on four reasoning benchmarks, it shows that the model often cannot pick out the right knowledge pieces, gets confused when that knowledge contradicts what it learned during pretraining, fails to build correct reasoning chains when the reasoning structure is non-sequential, and cannot generalise from seeing few-hop examples to many-hop questions. If true, the result matters because it locates the bottleneck not in the language model's factual memory but in the composition step: combining separate pieces of evidence into a sound argument. The paper concludes that current LLMs still show a large gap with human reasoning on these tasks.","feed_headline":"CoT prompting still fails on multi-hop reasoning with extra knowledge.","feed_subtitle":"GPT-3.5 with chain-of-thought cannot select and chain external facts, handle non-sequential proofs, or scale hop counts.","key_machinery":"The central mechanism under test is chain-of-thought (CoT) prompting, the few-shot instruction that asks the model to write out intermediate steps before the final answer, together with its variant least-to-most prompting, which inserts an explicit decomposition stage. The paper probes this mechanism with four benchmarks that isolate different sub-skills: HotpotQA for selecting two disjoint evidence pieces, EntailmentBank for composing many steps including non-sequential entailment trees, QASC for two-hop answer composition with distractors, and bAbI15 for unanswerable (non-deductive) cases. The key diagnostic object is the 'entailment tree' from EntailmentBank, whose depth versus path length is used to label a reasoning problem as non-sequential, and the paper also uses counterfactual knowledge swaps in QASC and bAbI15 to test how the model separates external from internal knowledge.","core_discovery":"The paper's central claim is that LLMs 'still suffer from severe drawbacks' in multi-hop reasoning with external knowledge even when prompted with chain-of-thought (CoT) and its variations. Concretely, four sets of experiments on HotpotQA, EntailmentBank, QASC and bAbI15 with text-davinci-002 show: internal knowledge alone gives low accuracy; adding external knowledge helps but the model rapidly loses the ability to select all gold knowledge as distractors increase; counterfactual knowledge causes the model to retreat to its pretraining knowledge or to the prompt's format, producing answers like 'not mentioned' far too rarely; on samples the paper classifies as non-sequential (where entailment-tree depth and path length differ), accuracy drops to 56% versus 78% on sequential samples and none of the generated reasoning paths is correct; and least-to-most prompting does not transfer across hop counts, with overall accuracy falling as the exemplars' hop count rises.","pith_inferences":["If the same pattern holds on later and larger models, the bottleneck would look like a property of CoT-style prompting itself rather than of one checkpoint; an easy test is to rerun the non-sequential and counterfactual settings on newer models and on open-weight models.","The depth-versus-length criterion could be validated against human judgments of non-sequential structure; if the proxy over- or under-selects examples, the 56% versus 78% gap may be partly a labelling artefact rather than a genuine reasoning deficit.","The counterfactual results suggest a practical recipe for making models more robust: teach them when to trust the context over parametric memory by mixing counterfactual exemplars or by adding explicit instructions to follow the context even when it contradicts pretraining, an extension the paper does not test.","The shortcut analysis (accuracy above exact-match in the 20% setting) implies that benchmark scores on HotpotQA-style tasks can overstate true multi-hop ability, so future benchmarks should report per-sample evidence that both hops were actually used."],"forward_implications":["With only internal knowledge, GPT-3.5 scores 40% on HotpotQA and 35% on EntailmentBank, so multi-hop benchmarks cannot be treated as tests of parametric knowledge alone.","Adding distractor-free external knowledge raises accuracy substantially, but the selection step degrades quickly: exact-match selection of both gold paragraphs in HotpotQA falls from 100% to 66% when only 20% of the context is gold.","Counterfactual context does not make the model reason counterfactually: it tends to hallucinate the factual knowledge it pretrained on, or to copy the prompt's format, and it fails most unanswerable (non-deductive) bAbI15 items.","On non-sequential entailment samples (depth differs from path length), accuracy is 56% and no generated proof is correct, whereas sequential samples reach 78%.","Least-to-most prompting does not generalise across hop counts: overall accuracy drops as the exemplars' hop count rises, and the model over-decomposes or mis-decomposes questions even when the final answer is right."],"supporting_citations":[{"why":"Supplies the chain-of-thought prompting method that all experiments are built on.","marker":"[9]"},{"why":"Supplies the least-to-most prompting variant used in the generalisation experiment.","marker":"[10]"},{"why":"Provides the HotpotQA two-hop question-answering dataset used for knowledge-selection experiments.","marker":"[34]"},{"why":"Provides the EntailmentBank dataset and its entailment trees, which are the basis for the non-sequential classification.","marker":"[35]"},{"why":"Provides the QASC two-hop science question-answering dataset used for the counterfactual-knowledge tests.","marker":"[36]"},{"why":"Provides the bAbI15 deduction dataset, which the paper modifies to include non-deductive (unanswerable) questions.","marker":"[37]"},{"why":"Supplies the Selection-Inference prompting baseline that the paper compares against in the knowledge-selection setting.","marker":"[32]"},{"why":"Supports the claim that LLMs rarely answer 'I don't know', which the paper uses to explain failures on non-deductive bAbI15 samples.","marker":"[39]"}],"fun_headline_variants":["CoT fails to chain external knowledge in multi-hop reasoning","GPT-3.5 still struggles with multi-hop external knowledge tasks","Multi-hop reasoning with external knowledge eludes GPT-3.5","External knowledge trips up LLMs in multi-hop reasoning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's classification of a reasoning problem as 'non-sequential' rests entirely on the criterion that the depth and the length of its entailment tree differ; if that proxy does not actually capture non-sequential structure, the claim that chain-of-thought fails on non-sequential reasoning is unsupported.","fun_headline_variants_meta":{"raw":{"variants":["CoT fails to chain external knowledge in multi-hop reasoning","GPT-3.5 still struggles with multi-hop external knowledge tasks","Multi-hop reasoning with external knowledge eludes GPT-3.5","External knowledge trips up LLMs in multi-hop reasoning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000148,"raw_usage":{"total_tokens":1127,"prompt_tokens":822,"completion_tokens":305,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":438,"completion_tokens_details":{"reasoning_tokens":235}},"tokens_in":438,"tokens_out":305,"duration_ms":3305,"temperature":1.0,"reasoning_tokens":235,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:56:31.841687+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the EntailmentBank samples the paper labels non-sequential (depth differs from length), have independent human annotators judge whether the reasoning is genuinely non-sequential, and compare GPT-3.5's accuracy on those against a matched set of sequential samples of equal proof length; if most labelled samples are judged sequential, or if a newer model with explicit non-sequential prompting shows no accuracy gap, the central claim is refuted.","supporting_citations":[{"cited_title":"Chain-of-thought prompting elicits reasoning in large language models","cited_arxiv_id":null,"evidence_quote":"Supplies the chain-of-thought prompting method that all experiments are built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the least-to-most prompting variant used in the generalisation experiment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the HotpotQA two-hop question-answering dataset used for knowledge-selection experiments."},{"cited_title":"Explaining answers with entailment trees","cited_arxiv_id":null,"evidence_quote":"Provides the EntailmentBank dataset and its entailment trees, which are the basis for the non-sequential classification."},{"cited_title":"Qasc: A dataset for question answering via sentence composition","cited_arxiv_id":null,"evidence_quote":"Provides the QASC two-hop science question-answering dataset used for the counterfactual-knowledge tests."},{"cited_title":"Do large language models know what they don’t know? In Findings of the Association for Computational Linguistics: ACL 2023, pages 8653–8665, 2023","cited_arxiv_id":null,"evidence_quote":"Supports the claim that LLMs rarely answer 'I don't know', which the paper uses to explain failures on non-deductive bAbI15 samples."}],"review_version":1}