{"paper":{"title":"Do Latent Tokens Think? A Causal and Adversarial Analysis of Chain-of-Continuous-Thought","arxiv_id":"2512.21711","primary_cat":"cs.CL","submitted_at":"2025-12-25T15:14:53+00:00","version_compiled":"v2"},"notice":"Every quoted sentence is the authors' text from the arXiv listing abstract. Labels and probabilities are the outputs of classifier-1.13.0 on the question bank abstract-p1-20260923, exactly as returned. No sentence in this record was written by a language model.","schema":"pith.compiled-record/v0","status":"compiled","abstract":{"source":"arXiv listing abstract v2","sentences":[{"i":0,"p":0.77,"id":"s0","own":false,"role":"background","text":"Latent tokens are gaining attention for enhancing reasoning in large language models (LLMs), yet their internal mechanisms remain unclear.","hedge":{"p":0.98,"label":"asserted"},"own_p":0.04,"main_claim_p":0.02,"quantitative":false,"quantitative_p":0.05,"role_probabilities":{"gap":0.23,"method":0.0,"result":0.0,"background":0.77,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":1,"p":0.81,"id":"s1","own":true,"role":"result","text":"This paper examines the problem from a reliability perspective, uncovering fundamental weaknesses: latent tokens function as uninterpretable placeholders rather than encoding faithful reasoning.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.89,"main_claim_p":0.81,"quantitative":false,"quantitative_p":0.06,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.81,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.1,"interpretation":0.08}},{"i":2,"p":0.83,"id":"s2","own":true,"role":"result","text":"While resistant to perturbation, they promote shortcut usage over genuine reasoning.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.71,"main_claim_p":0.17,"quantitative":false,"quantitative_p":0.32,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.83,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.17}},{"i":3,"p":0.68,"id":"s3","own":false,"role":"background","text":"We focus on Chain-of-Continuous-Thought (COCONUT), which claims better efficiency and stability than explicit Chain-of-Thought (CoT) while maintaining performance.","hedge":{"p":0.83,"label":"asserted"},"own_p":0.1,"main_claim_p":0.03,"quantitative":true,"quantitative_p":0.72,"role_probabilities":{"gap":0.01,"method":0.02,"result":0.0,"background":0.68,"limitation":0.0,"future_work":0.0,"contribution":0.29,"interpretation":0.0}},{"i":4,"p":0.73,"id":"s4","own":true,"role":"method","text":"We investigate this through two complementary approaches.","hedge":{"p":0.58,"label":"unclear"},"own_p":0.52,"main_claim_p":0.02,"quantitative":false,"quantitative_p":0.2,"role_probabilities":{"gap":0.0,"method":0.73,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.27,"interpretation":0.0}},{"i":5,"p":0.99,"id":"s5","own":true,"role":"method","text":"First, steering experiments perturb specific token subsets, namely COCONUT and explicit CoT.","hedge":{"p":0.91,"label":"asserted"},"own_p":0.88,"main_claim_p":0.02,"quantitative":false,"quantitative_p":0.15,"role_probabilities":{"gap":0.0,"method":0.99,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.01,"interpretation":0.0}},{"i":6,"p":1.0,"id":"s6","own":true,"role":"result","text":"Unlike CoT tokens, COCONUT tokens show minimal sensitivity to steering and lack reasoning-critical information.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.91,"main_claim_p":0.15,"quantitative":true,"quantitative_p":0.81,"role_probabilities":{"gap":0.0,"method":0.0,"result":1.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":7,"p":0.99,"id":"s7","own":true,"role":"method","text":"Second, shortcut experiments evaluate models under biased and out-of-distribution settings.","hedge":{"p":0.75,"label":"asserted"},"own_p":0.83,"main_claim_p":0.02,"quantitative":false,"quantitative_p":0.05,"role_probabilities":{"gap":0.0,"method":0.99,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.01,"interpretation":0.0}},{"i":8,"p":1.0,"id":"s8","own":true,"role":"result","text":"Results on MMLU and HotpotQA demonstrate that COCONUT consistently exploits dataset artifacts, inflating benchmark performance without true reasoning.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.94,"main_claim_p":0.21,"quantitative":false,"quantitative_p":0.19,"role_probabilities":{"gap":0.0,"method":0.0,"result":1.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":9,"p":1.0,"id":"s9","own":true,"role":"interpretation","text":"These findings reposition COCONUT as a pseudo-reasoning mechanism: it generates plausible traces that conceal shortcut dependence rather than faithfully representing reasoning processes.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.89,"main_claim_p":0.88,"quantitative":false,"quantitative_p":0.06,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":1.0}}]},"compiled":{"main_claim":{"p":0.75,"text":"These findings reposition COCONUT as a pseudo-reasoning mechanism: it generates plausible traces that conceal shortcut dependence rather than faithfully representing reasoning processes.","location":"abstract s9","hedge_strength":{"p":1.0,"label":"asserted"}},"provenance":{"route":{"via":"TypeSafe API","model":"classifier-1.13.0","endpoint":"https://api.typesafe.ai/v1/systemone","provider":"TypeSafe"},"packing":"one request per paper","requests":1,"input_tokens":7380,"cost_usd_reported":0.0},"compiled_at":"2026-10-27","bank_version":"abstract-p1-20260923","model_version":"classifier-1.13.0","abstract_states":{"claims_first":0.04,"states_limitation":0.15,"human_participants":0.03,"controlled_comparison":0.93,"randomized_assignment":0.02,"states_quantitative_result":0.05,"compares_with_prior_results":0.92,"mentions_code_or_data_availability":0.02},"evidential_basis":{"p":1.0,"value":"observation or experiment","probabilities":{"argument":0.0,"literature":0.0,"proof_or_derivation":0.0,"analysis_of_existing_data":0.0,"observation_or_experiment":1.0,"simulation_or_computation":0.0}},"contribution_type":{"p":1.0,"value":"empirical study","probabilities":{"review":0.0,"method_or_tool":0.0,"empirical_study":1.0,"simulation_study":0.0,"theoretical_result":0.0,"dataset_or_benchmark":0.0,"position_or_commentary":0.0}},"contribution_statement":{"p":0.48,"text":"We investigate this through two complementary approaches.","location":"abstract s4"}},"coverage":{"tier":"abstract","levels_run":["P abstract"],"not_examined":["full text","figures and tables","references and citing sentences","relations to other papers"],"sources_read":["arXiv listing abstract v2"],"levels_not_run":["S sections","T body sentences","C citations","R cross-paper relations"],"sentences_examined":10,"sentences_in_abstract":10},"measured_accuracy":{"url":"https://pith.science/accuracy","measured_on":"2026-10-02","method":"two independent model families labelled a fresh random sample blind; the figure is how often the published label matches their consensus","sentence_role_agreement":0.92,"main_claim_agreement":0.908,"typed_citation_agreement":0.933,"typed_citation_agreement_95_interval":[0.917,0.945],"review_comments_accurate":0.758,"review_comments_accurate_95_interval":[0.684,0.82],"review_caught_admitted_error_found_or_related":0.448,"review_comments_judged":{"judged_by":["gpt-5.6-luna","grok-4.7"],"judged_against":"Pith's own text copy of each paper: text extracted from the PDF, the first 90,000 characters","measured_on":"2026-10-04","accurate":113,"decided":149,"sampled_reviews":200,"split_by_lane":false,"by_lane":"https://pith.science/accuracy#by-lane"}}}