new

Get trending papers in your email inbox!

Subscribe

Daily Papers

byAK and the research community

Nov 17

Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-Agent Messages

Multi-agent reasoning systems often use agreement, confidence, or automated scores to decide which messages should shape a final answer. Such filtering assumes that a message likely to be correct is also worth keeping. Yet a wrong answer can contain a useful decomposition, constraint, or scientific principle. We test this distinction with Diverse Hypothesis Deliberation (DHD), a controlled measurement protocol that caches five independently generated messages and replays the same downstream solver, called the integrator, with each message available or hidden. The replay comparison measures a message's trajectory value: whether making the message available helps or harms subsequent reasoning. Across five mathematics and science benchmarks and two openly available model families, gpt-oss-120b and gemma-4-31B-it, wrong-helpful messages appear in every benchmark-model combination. Among wrong-answer messages that change final correctness, more than four in ten changes are helpful in each model. Controlled repeats show that the number of repeatable message effects is unlikely to arise from replay variation alone (p=0.0002). A focused intervention on repeatable wrong-helpful messages finds that the complete message works best, while retaining its reasoning preserves more success than retaining only its answer; the source of the complete-message advantage remains open. Within the same problem, repeated trajectory-value evidence also identifies a better keep-or-remove choice than answer correctness alone. Answer correctness is therefore informative but does not determine trajectory value. DHD measures this missing property and produces reusable labels for learning when agents should listen.

  • 10 authors
·
Aug 13

The Climb Carves Wisdom Deeper Than the Summit: On the Noisy Rewards in Learning to Reason

Recent studies on post-training large language models (LLMs) for reasoning through reinforcement learning (RL) typically focus on tasks that can be accurately verified and rewarded, such as solving math problems. In contrast, our research investigates the impact of reward noise, a more practical consideration for real-world scenarios involving the post-training of LLMs using reward models. We found that LLMs demonstrate strong robustness to substantial reward noise. For example, manually flipping 40% of the reward function's outputs in math tasks still allows a Qwen-2.5-7B model to achieve rapid convergence, improving its performance on math tasks from 5% to 72%, compared to the 75% accuracy achieved by a model trained with noiseless rewards. Surprisingly, by only rewarding the appearance of key reasoning phrases (namely reasoning pattern reward, RPR), such as ``first, I need to''-without verifying the correctness of answers, the model achieved peak downstream performance (over 70% accuracy for Qwen-2.5-7B) comparable to models trained with strict correctness verification and accurate rewards. Recognizing the importance of the reasoning process over the final results, we combined RPR with noisy reward models. RPR helped calibrate the noisy reward models, mitigating potential false negatives and enhancing the LLM's performance on open-ended tasks. These findings suggest the importance of improving models' foundational abilities during the pre-training phase while providing insights for advancing post-training techniques. Our code and scripts are available at https://github.com/trestad/Noisy-Rewards-in-Learning-to-Reason.

  • 5 authors
·
May 28, 2025 2

MemOps: Benchmarking Lifecycle Memory Operations in Long-Horizon Conversations

Long-term memory has become a foundational capability for LLM-based agents that accompany users across extended, multi-session interactions. Existing benchmarks, however, evaluate such memory almost exclusively through downstream question answering, scoring only the correctness of a final answer. This black-box formulation conflates the heterogeneous causes of memory failure, such as missing the introduction of a relevant fact, binding an operation to the wrong target, or relying on stale values after a correction. As a result, it can credit correct answers despite their reliance on inconsistent or unsafe memory states. In this paper, we argue that, in dynamic long-horizon interactions, memory is not a static collection of facts but a lifecycle of explicit operations, including remembering, forgetting, updating, reflecting, and their compositions. We introduce MemOps, a benchmark that reformulates conversational memory as a sequence of lifecycle operations and represents each memory event with a structured trace specifying its trigger, target, scope, state transition, and supporting evidence. A controllable generation pipeline embeds these operations into long, task-oriented conversations and produces gold operation traces together with six categories of operation-level probes, evaluated under both adjacent-evidence and long-context settings. Across long-context, retrieval-based, parametric and managed-memory systems, MemOps disentangles failure modes that final-answer accuracy alone conceals, revealing that current systems remain far from uniformly reliable. For instance, session-level retrieval outperforms turn-level retrieval, and long-context models remain notably weak at reconstructing ordered memory-state trajectories. These results move long-term memory evaluation from final-answer scoring toward interpretable, operation-level diagnosis.

  • 9 authors
·
Jul 13

Answer-Conditioned Chains of Thought Degrade Verifiable-Reasoning Distillation in Large Language Models

A standard recipe for distilling the reasoning ability of large language models (LLMs) is to sample chains of thought from the model, keep those that reach the correct final answer, and fine-tune on the survivors. When sampling fails, a common fix shows the generator the gold answer and asks it to write a chain that reaches that answer. We show that this second step degrades the training data in a way that correctness filtering cannot catch. We run a controlled experiment that fixes the generator, the problem set, and the correctness filter, and varies only whether the chain is generated under answer-conditioning, the gold answer shown with a request to reach it. Training a strong instruction-tuned reasoning model on its own answer-conditioned chains sharply lowers its verifiable-reasoning accuracy. The loss grows with difficulty, reaching as much as about 27 points on the hardest competition problems. The mechanism is legible in the chains themselves, which rationalize backward from the shown answer instead of deriving it, with the early final-answer statement as the measurable symptom. The harm is a property of the data rather than the generator, read off unlabeled generations before any fine-tuning, ordering the penalty across eight thinking models from four families, and transferring across teacher families. A prompt ablation localizes it to the rationalize-toward instruction rather than the answer's bare visibility. The practical takeaway is to generate answer-blind, because no correctness filter can see this damage in the data.

  • 7 authors
·
Jul 15

Evaluating Correctness and Faithfulness of Instruction-Following Models for Question Answering

Retriever-augmented instruction-following models are attractive alternatives to fine-tuned approaches for information-seeking tasks such as question answering (QA). By simply prepending retrieved documents in its input along with an instruction, these models can be adapted to various information domains and tasks without additional fine-tuning. While the model responses tend to be natural and fluent, the additional verbosity makes traditional QA evaluation metrics such as exact match (EM) and F1 unreliable for accurately quantifying model performance. In this work, we investigate the performance of instruction-following models across three information-seeking QA tasks. We use both automatic and human evaluation to evaluate these models along two dimensions: 1) how well they satisfy the user's information need (correctness), and 2) whether they produce a response based on the provided knowledge (faithfulness). Guided by human evaluation and analysis, we highlight the shortcomings of traditional metrics for both correctness and faithfulness. We then propose simple token-overlap based and model-based metrics that reflect the true performance of these models. Our analysis reveals that instruction-following models are competitive, and sometimes even outperform fine-tuned models for correctness. However, these models struggle to stick to the provided knowledge and often hallucinate in their responses. We hope our work encourages a more holistic evaluation of instruction-following models for QA. Our code and data is available at https://github.com/McGill-NLP/instruct-qa

  • 5 authors
·
Jul 31, 2023

BITEM at the NTCIR-19 R2C2 Task: Predicting Confidence from Agentic RAG Pipeline Signals

The BITEM team entered both subtasks of the NTCIR-19 R2C2 task with a single agentic pipeline, in which a model searches, reads and records evidence over a movie corpus while an orchestrator holds the record and rules on what may be submitted. A claim is admitted only once an entailment cascade has checked it against the passage it cites, and an answer is released only once enough checked evidence stands behind it. Each question is run three or four times, every pass retrieving from a corpus stripped of what the earlier passes have already seen. The confidence filed with each answer is computed by the orchestrator from what the run leaves behind and is never asked of the model, which is offered no way to rate itself. The two retrieval runs placed 4th and 5th of 22, pooling the passes was worth 0.0709 nDCG@20, and the gain was largest on the multi-hop and post-processing-heavy questions, where the organisers rank the pooled run top of the field. Sixteen of the 25 answer runs were built on passages these two runs supplied, 12 of them filed by other teams. HMR rewards a system whose confidence is high where it answers right and low where it answers wrong. The pipeline reached an accuracy of 0.9219, 6th of 25, while the confidence filed with those answers gave an HMR of 0.4915, 13th. A few rules crafted over those same recorded signals, with no further model call and no further retrieval, raise that to an accuracy of 0.9375, 5th, and an HMR of 0.6985, 9th. Ranking on HMR alone can reward a system for answering wrongly with low confidence, so we propose accHMR, the accuracy multiplied by HMR, which reports the reward in proportion to the accuracy, and on which the revised rules would have scored 0.6549, 5th. For future work, fitting a model on the numbers the pipeline already produces, rather than writing such rules by hand, would be a real step forward.

  • 6 authors
·
Sep 28

From Architecture to Output: Testing the Structural Origins of Hallucination and the Amplifying Role of Data

Surveys of hallucination classify errors by how an output relates to its source, and often explain them by the model's structure: attention retrieves by association, the maximum-likelihood objective has no term for truth, and left-to-right decoding commits to early tokens. Data, on this account, makes these failures worse rather than causing them. The account is usually argued rather than tested. We test it with interventions that need only sampling access. Study 1 holds pretraining data, tokenizer and scale fixed and swaps attention for a state-space mixer, using five Pythia/Mamba pairs from 130M to 2.8B parameters and four further attention models, on 111 facts presented with a planted false or a planted true statement. Swapping the mixer leaves free-recall failure unchanged (both 2.8B models answer 3 of 200 SimpleQA questions), but it changes what happens to a defect in the input. From 410M upwards, every attention model repeats the planted falsehood more often than its matched state-space model (2.8B: 650 vs. 566 of 666 items, p = 2.5e-20), also on facts both models know, while the state-space models can still read planted facts they did not know. Study 2 replaces the first wrong fact in 111 erroneous answers from three instruction-tuned models. The correct fact cuts the share of answers with a downstream error from 92.5% to 48.1%, but a different wrong fact works as well (p = 0.38), and the models give the right answer in isolation on only 2 of 111 items. The structural account survives in a narrower form. The mixing component decides whether a defect in the context becomes an output error, commitment decides how an error spreads, and on long-tail questions the first error comes from missing knowledge, not from a suppressed truth.

  • 3 authors
·
Nov 8

Training Curricula for Open Domain Answer Re-Ranking

In precision-oriented tasks like answer ranking, it is more important to rank many relevant answers highly than to retrieve all relevant answers. It follows that a good ranking strategy would be to learn how to identify the easiest correct answers first (i.e., assign a high ranking score to answers that have characteristics that usually indicate relevance, and a low ranking score to those with characteristics that do not), before incorporating more complex logic to handle difficult cases (e.g., semantic matching or reasoning). In this work, we apply this idea to the training of neural answer rankers using curriculum learning. We propose several heuristics to estimate the difficulty of a given training sample. We show that the proposed heuristics can be used to build a training curriculum that down-weights difficult samples early in the training process. As the training process progresses, our approach gradually shifts to weighting all samples equally, regardless of difficulty. We present a comprehensive evaluation of our proposed idea on three answer ranking datasets. Results show that our approach leads to superior performance of two leading neural ranking architectures, namely BERT and ConvKNRM, using both pointwise and pairwise losses. When applied to a BERT-based ranker, our method yields up to a 4% improvement in MRR and a 9% improvement in P@1 (compared to the model trained without a curriculum). This results in models that can achieve comparable performance to more expensive state-of-the-art techniques.

  • 6 authors
·
Apr 29, 2020

Beyond the Last Answer: Your Reasoning Trace Uncovers More than You Think

Large Language Models (LLMs) leverage step-by-step reasoning to solve complex problems. Standard evaluation practice involves generating a complete reasoning trace and assessing the correctness of the final answer presented at its conclusion. In this paper, we challenge the reliance on the final answer by posing the following two questions: Does the final answer reliably represent the model's optimal conclusion? Can alternative reasoning paths yield different results? To answer these questions, we analyze intermediate reasoning steps, termed subthoughts, and propose a method based on our findings. Our approach involves segmenting a reasoning trace into sequential subthoughts based on linguistic cues. We start by prompting the model to generate continuations from the end-point of each intermediate subthought. We extract a potential answer from every completed continuation originating from different subthoughts. We find that aggregating these answers by selecting the most frequent one (the mode) often yields significantly higher accuracy compared to relying solely on the answer derived from the original complete trace. Analyzing the consistency among the answers derived from different subthoughts reveals characteristics that correlate with the model's confidence and correctness, suggesting potential for identifying less reliable answers. Our experiments across various LLMs and challenging mathematical reasoning datasets (AIME2024 and AIME2025) show consistent accuracy improvements, with gains reaching up to 13\% and 10\% respectively. Implementation is available at: https://github.com/hammoudhasan/SubthoughtReasoner.

  • 3 authors
·
Apr 29, 2025 2

FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation

Most large language models (LLMs) are trained once and never updated; thus, they lack the ability to dynamically adapt to our ever-changing world. In this work, we perform a detailed study of the factuality of LLM-generated text in the context of answering questions that test current world knowledge. Specifically, we introduce FreshQA, a novel dynamic QA benchmark encompassing a diverse range of question and answer types, including questions that require fast-changing world knowledge as well as questions with false premises that need to be debunked. We benchmark a diverse array of both closed and open-source LLMs under a two-mode evaluation procedure that allows us to measure both correctness and hallucination. Through human evaluations involving more than 50K judgments, we shed light on limitations of these models and demonstrate significant room for improvement: for instance, all models (regardless of model size) struggle on questions that involve fast-changing knowledge and false premises. Motivated by these results, we present FreshPrompt, a simple few-shot prompting method that substantially boosts the performance of an LLM on FreshQA by incorporating relevant and up-to-date information retrieved from a search engine into the prompt. Our experiments show that FreshPrompt outperforms both competing search engine-augmented prompting methods such as Self-Ask (Press et al., 2022) as well as commercial systems such as Perplexity.AI. Further analysis of FreshPrompt reveals that both the number of retrieved evidences and their order play a key role in influencing the correctness of LLM-generated answers. Additionally, instructing the LLM to generate concise and direct answers helps reduce hallucination compared to encouraging more verbose answers. To facilitate future work, we release FreshQA at github.com/freshllms/freshqa and commit to updating it at regular intervals.

  • 11 authors
·
Oct 4, 2023 1

Search-G1: Co-Evolving Representation-Based Rewards for Selective Grounded Search

Search-augmented language agents should retrieve external information only when necessary and ground their answers in retrieved evidence. Yet terminal correctness assigns equal credit to direct answers from parametric knowledge, evidence-grounded answers, and answers recalled after redundant search. Process supervision and LLM judges provide richer feedback but can require high-cost annotation or inference; low-cost, non-contrastive policy confidence signals do not directly identify whether retrieved evidence changes an answer. We propose Search-G1, a reward framework that couples retrieval necessity with evidence reliance through two intervention-calibrated hidden-state readouts for selective grounded search. A prompt readout predicts closed-book sufficiency, whose complement proxies policy-relative retrieval necessity; an answer-commit readout predicts evidence reliance as sensitivity to evidence deletion conditional on the realized reasoning prefix. A correctness-first reward gates evidence-reliance credit by retrieval necessity, rewards sufficient direct answers, and penalizes repeated search. After calibration, scoring needs neither process annotations nor LLM-judge inference. As policy-relative targets, trajectory distributions, and representations may drift, Search-G1 periodically regenerates targets and refits its readouts. Across four QA benchmarks and two scales, Search-G1 achieves competitive task accuracy, with stronger counterfactual evidence dependence and fewer searches in the main comparisons, complemented by independent human audits of answer support.

  • 10 authors
·
Oct 9

Beyond the Answer: Student-Conditioned Trace Construction for Reasoning Distillation

A reasoning trace can accompany a correct answer without providing effective supervision for a particular student. We study how to construct traces that improve learning beyond answer-only supervision by considering both their intermediate content and their accessibility to the target learner. We introduce ReDiST, a student-conditioned trace-construction and distillation framework for evidence-rich reasoning. Starting from the student's own attempts, ReDiST revises answer-correct drafts against the supplied evidence and repairs unsuccessful drafts with additional guidance from the verified answer. A frozen student-matched reader screens and ranks original and edited candidates using answer recovery and verified-answer probability, given the question, evidence, and candidate trace. We distill the selected targets while balancing trace and final-answer supervision with an auxiliary answer loss. Construction and selection occur offline, and inference uses only the distilled student. Experiments on question-answering benchmarks show that ReDiST can improve both source-task and cross-benchmark performance over answer-only, original-trace, and the evaluated trace-selection and rewriting baselines. Across supervision recipes, performance rankings can differ between source and transfer benchmarks, highlighting the importance of evaluating trace teaching utility across distributions.

  • 9 authors
·
Oct 15

Narrowing the Knowledge Evaluation Gap: Open-Domain Question Answering with Multi-Granularity Answers

Factual questions typically can be answered correctly at different levels of granularity. For example, both ``August 4, 1961'' and ``1961'' are correct answers to the question ``When was Barack Obama born?''. Standard question answering (QA) evaluation protocols, however, do not explicitly take this into account and compare a predicted answer against answers of a single granularity level. In this work, we propose GRANOLA QA, a novel evaluation setting where a predicted answer is evaluated in terms of accuracy and informativeness against a set of multi-granularity answers. We present a simple methodology for enriching existing datasets with multi-granularity answers, and create GRANOLA-EQ, a multi-granularity version of the EntityQuestions dataset. We evaluate a range of decoding methods on GRANOLA-EQ, including a new algorithm, called Decoding with Response Aggregation (DRAG), that is geared towards aligning the response granularity with the model's uncertainty. Our experiments show that large language models with standard decoding tend to generate specific answers, which are often incorrect. In contrast, when evaluated on multi-granularity answers, DRAG yields a nearly 20 point increase in accuracy on average, which further increases for rare entities. Overall, this reveals that standard evaluation and decoding schemes may significantly underestimate the knowledge encapsulated in LMs.

  • 3 authors
·
Jan 9, 2024

The Last Word Often Wins: A Format Confound in Chain-of-Thought Corruption Studies

Corruption studies, the primary tool for evaluating chain-of-thought (CoT) faithfulness, identify which chain positions are "computationally important" by measuring accuracy when steps are replaced with errors. We identify a systematic confound: for chains with explicit terminal answer statements, the dominant format in standard benchmarks, corruption studies detect where the answer text appears, not where computation occurs. A within-dataset format ablation provides the key evidence: on standard GSM8K chains ending with "the answer is X," removing only the answer statement, preserving all reasoning, collapses suffix sensitivity ~19x at 3B (N=300, p=0.022). Conflicting-answer experiments quantify the causal mechanism: at 7B, CC accuracy drops to near-zero (<=0.02) across five architecture families; the followed-wrong rate spans 0.63-1.00 at 3B-7B and attenuates at larger scales (0.300 at Phi-4-14B, ~0.01 at 32B). A within-stable 7B replication (9.3x attenuation, N=76, p=7.8e-3; Qwen3-8B N=299, p=0.004) provides converging evidence, and the pattern replicates on MATH (DeepSeek-R1-7B: 10.9x suffix-survival recovery). On chains without answer suffixes the same protocol identifies the prefix as load-bearing (Delta=-0.77, p<10^-12). Generation-time probes confirm a dissociation: the answer is not early-determined during generation (early commitment <5%), yet at consumption time model outputs systematically follow the explicit answer text. The format-determination effect persists through 14B (8.5x ratio, p=0.001) and converges toward zero at 32B. We propose a three-prerequisite protocol (question-only control, format characterization, all-position sweep) as a minimum standard for corruption-based faithfulness studies.

  • 1 authors
·
May 10

The Curious Case of Factual (Mis)Alignment between LLMs' Short- and Long-Form Answers

Large language models (LLMs) can correctly answer "When was Einstein born?" yet fail to provide the same date when writing about Einstein's life revealing a fundamental inconsistency in how models access factual knowledge across task complexities. While models display impressive accuracy on factual question-answering benchmarks, the reliability gap between simple and complex queries remains poorly understood, eroding their trustworthiness. In this work, we introduce Short-Long Form Alignment for Factual Question Answering (SLAQ), a controlled evaluation framework that compares LLMs' answers to the same factual questions asked (a) in isolation (short) vs. (b) integrated into complex queries (long). Looking at 16 LLMs across 600 queries, we find a systematic misalignment of answers to the corresponding short and long queries. We further uncover position-dependent accuracy loss and momentum effects where consecutive correct or incorrect answers create self-reinforcing patterns. Through mechanistic analysis, we find that aligned facts activate overlapping model internals, and that metrics based on mechanistic similarity can predict short-long answer alignment with up to 78% accuracy. Our work establishes factual consistency over query complexity as an important aspect of LLMs' trustworthiness and challenges current evaluation practices, which implicitly assume that good performance for simple factual queries implies reliability in more complex knowledge-seeking tasks too.

WueNLP WüNLP
·
Oct 13, 2025 2

Knowing Is Not Choosing: What Explicit Verification Adds Beyond Generative Preference

Generating a correct answer does not mean that a language model will select it. We separate factual recall into three steps: generating a correct candidate, ranking the available candidates, and selecting the final answer. Pre-generation readouts predict factual recall and which questions sampling will cover across three model families, but say little about whether an available correct answer will ultimately be selected. Explicit verification with P(True) improves within-question ranking over mean log-likelihood in Gemma, Qwen3, and Llama, with AUROC gains of 0.08--0.12. In a prospectively defined Gemma cohort, verification raises plurality accuracy by about 5 points, and still gains about 2 points over chat-template likelihood, a stronger generative baseline. The advantage is strongest for relations with common-answer priors and depends on access to the entity; masking the entity removes the ranking advantage in larger Qwen models. Finally, the measured benefit depends on how correctness is defined: recall-oriented reference matching can credit option lists favored by likelihood and substantially understate the improvement seen under human semantic judgments. Prior work shows that models can carry latent factual knowledge and judge candidate answers; we show that these capabilities do not collapse into a single notion of ``knowing,'' and trace where information is gained, lost, or mismeasured between availability, ranking, and final choice.

  • 4 authors
·
Sep 28

From Faithfulness to Correctness: Generative Reward Models that Think Critically

Through reinforcement learning with verifiable rewards (RLVR), large language models have achieved substantial progress in domains with easily verifiable outcomes, such as mathematics and coding. However, when applied to more complex tasks like open-domain question answering, RLVR faces significant challenges due to the difficulty of verifying correctness. The nuanced and ambiguous nature of real-world knowledge makes it difficult to reliably evaluate correctness in these settings, necessitating further abilities that extend beyond mere logical consistency to encompass an understanding and assessment of both external and internal knowledge. Recent work has primarily focused on improving faithfulness, defined as semantic alignment with supporting documents, which can cause models to rely excessively on external sources and diminish their capacity for critical assessment. To address this, we propose the Thinking-supervised Reward Model (TRM), which incorporates sentence-level thinking supervision to endow reward models with critical thinking abilities. Given a query, answer, and supporting documents, TRM first assesses the faithfulness of each answer sentence to the supporting documents, and then applies a reasoning step to evaluate sentence-level correctness. By structuring reward modeling as a sequence of faithfulness, reasoning, and correctness evaluations, TRM encourages models to critically assess and leverage both external and internal knowledge. Experiments on reward signals demonstrate that TRM substantially improves the identification of incorrect sentences, and incorporating TRM into policy optimization leads to significant gains in both answer correctness and usefulness.

  • 6 authors
·
Sep 29, 2025

SuRe: Summarizing Retrievals using Answer Candidates for Open-domain QA of LLMs

Large language models (LLMs) have made significant advancements in various natural language processing tasks, including question answering (QA) tasks. While incorporating new information with the retrieval of relevant passages is a promising way to improve QA with LLMs, the existing methods often require additional fine-tuning which becomes infeasible with recent LLMs. Augmenting retrieved passages via prompting has the potential to address this limitation, but this direction has been limitedly explored. To this end, we design a simple yet effective framework to enhance open-domain QA (ODQA) with LLMs, based on the summarized retrieval (SuRe). SuRe helps LLMs predict more accurate answers for a given question, which are well-supported by the summarized retrieval that could be viewed as an explicit rationale extracted from the retrieved passages. Specifically, SuRe first constructs summaries of the retrieved passages for each of the multiple answer candidates. Then, SuRe confirms the most plausible answer from the candidate set by evaluating the validity and ranking of the generated summaries. Experimental results on diverse ODQA benchmarks demonstrate the superiority of SuRe, with improvements of up to 4.6% in exact match (EM) and 4.0% in F1 score over standard prompting approaches. SuRe also can be integrated with a broad range of retrieval methods and LLMs. Finally, the generated summaries from SuRe show additional advantages to measure the importance of retrieved passages and serve as more preferred rationales by models and humans.

  • 8 authors
·
Apr 16, 2024

ReCEval: Evaluating Reasoning Chains via Correctness and Informativeness

Multi-step reasoning ability is fundamental to many natural language tasks, yet it is unclear what constitutes a good reasoning chain and how to evaluate them. Most existing methods focus solely on whether the reasoning chain leads to the correct conclusion, but this answer-oriented view may confound reasoning quality with other spurious shortcuts to predict the answer. To bridge this gap, we evaluate reasoning chains by viewing them as informal proofs that derive the final answer. Specifically, we propose ReCEval (Reasoning Chain Evaluation), a framework that evaluates reasoning chains via two key properties: (1) correctness, i.e., each step makes a valid inference based on information contained within the step, preceding steps, and input context, and (2) informativeness, i.e., each step provides new information that is helpful towards deriving the generated answer. We evaluate these properties by developing metrics using natural language inference models and V-Information. On multiple datasets, we show that ReCEval effectively identifies various error types and yields notable improvements compared to prior methods. We analyze the impact of step boundaries, and previous steps on evaluating correctness and demonstrate that our informativeness metric captures the expected flow of information in high-quality reasoning chains. Finally, we show that scoring reasoning chains based on ReCEval improves downstream task performance. Our code is publicly available at: https://github.com/archiki/ReCEval

  • 4 authors
·
Nov 29, 2023

Using clarification questions to improve software developers' Web search

Context: Recent research indicates that Web queries written by software developers are not very successful in retrieving relevant results, performing measurably worse compared to general purpose Web queries. Most approaches up to this point have addressed this problem with software engineering-specific automated query reformulation techniques, which work without developer involvement but are limited by the content of the original query. In other words, these techniques automatically improve the existing query but can not contribute new, previously unmentioned, concepts. Objective: In this paper, we propose a technique to guide software developers in manually improving their own Web search queries. We examine a conversational approach that follows unsuccessful queries with a clarification question aimed at eliciting additional query terms, thus providing to the developer a clear dimension along which the query could be improved. Methods: We describe a set of clarification questions derived from a corpus of software developer queries and a neural approach to recommending them for a newly issued query. Results: Our evaluation indicates that the recommendation technique is accurate, predicting a valid clarification question 80% of the time and outperforms simple baselines, as well as, state-of-the-art Learning To Rank (LTR) baselines. Conclusion: As shown in the experimental results, the described approach is capable at recommending appropriate clarification questions to software developers and considered useful by a sample of developers ranging from novices to experienced professionals.

  • 2 authors
·
Jul 26, 2022

Shape of Thought: When Distribution Matters More than Correctness in Reasoning Tasks

We present the surprising finding that a language model's reasoning capabilities can be improved by training on synthetic datasets of chain-of-thought (CoT) traces from more capable models, even when all of those traces lead to an incorrect final answer. Our experiments show this approach can yield better performance on reasoning tasks than training on human-annotated datasets. We hypothesize that two key factors explain this phenomenon: first, the distribution of synthetic data is inherently closer to the language model's own distribution, making it more amenable to learning. Second, these `incorrect' traces are often only partially flawed and contain valid reasoning steps from which the model can learn. To further test the first hypothesis, we use a language model to paraphrase human-annotated traces -- shifting their distribution closer to the model's own distribution -- and show that this improves performance. For the second hypothesis, we introduce increasingly flawed CoT traces and study to what extent models are tolerant to these flaws. We demonstrate our findings across various reasoning domains like math, algorithmic reasoning and code generation using MATH, GSM8K, Countdown and MBPP datasets on various language models ranging from 1.5B to 9B across Qwen, Llama, and Gemma models. Our study shows that curating datasets that are closer to the model's distribution is a critical aspect to consider. We also show that a correct final answer is not always a reliable indicator of a faithful reasoning process.

Measuring Reasoning Utility in LLMs via Conditional Entropy Reduction

Recent advancements in large language models (LLMs) often rely on generating intermediate reasoning steps to enhance accuracy. However, little work has examined how reasoning utility contributes to the final answer's correctness. Due to the stochastic nature of autoregressive generation, generating more context does not guarantee increased confidence in the answer. If we could predict, during generation, whether a reasoning step will be useful, we could stop early or prune ineffective steps, avoiding distractions in the final decision. We present an oracle study on MATH dataset, using Qwen2.5-32B and GPT-4o to generate reasoning chains, and then employing a separate model (Qwen3-8B) to quantify the utility of these chains for final accuracy. Specifically, we measure the model's uncertainty on the answer span Y at each reasoning step using conditional entropy (expected negative log-likelihood over the vocabulary) with context expanding step by step. Our results show a clear pattern: conditional entropy that decreases over steps is strongly associated with correct answers, whereas flat or increasing entropy often results in wrong answers. We also corroborate that incorrect reasoning paths tend to be longer than correct ones, suggesting that longer reasoning does not necessarily yield better outcomes. These findings serve as a foundation to inspire future work on designing efficient reasoning pipelines that detect and avoid unproductive reasoning early.

  • 1 authors
·
Aug 27, 2025

PageRecall: Measuring Page Selection in Literature-Grounded Question Answering

We describe our system for LitTraceQA (GroundLM @ EMNLP 2026): given a research question, retrieve the relevant papers from a pool of 27,487, cite the page and the table or figure where the answer lives, and answer in a requested format. Our main finding is that evidence grounding is limited by retrieval, not by reading. The page selector put the annotator's page, which we call the gold page, in front of the model that locates evidence only about half the time (52.6% gold-page recall), while that model, given the page, cited the right one in 45 of the 48 locators it emitted (94%). When the page was missing it rarely said so: of 45 such cases it returned nothing 14 times, a wrong page 24 times, and a correct page 7 times, so the pipeline failed quietly almost twice as often as it failed visibly. Since the failure was that the right page was never shown, the fix is to stop choosing: each retrieved paper fits in the model's context, so we show it whole. Page ranking survives only as a fallback inside papers too long to fit, which no test-split paper was, and gold-page recall reaches 100% on the papers we can parse. Separately, questions that identify their target by position rather than content, such as "the first author of the 24th reference", are served by parsing rather than retrieval: we resolve the bibliography into an addressable list, which also supplies identifiers the evidence metric scores. The final system scores 0.762 paper F_1, 0.441 evidence F_1 and 0.920 multiple-choice accuracy on the held-out test split. Because the pipeline depends on a closed model without seed control, we release a harness that verifies the paper's central claims against committed artifacts.

  • 1 authors
·
Sep 15

Diminished Diversity-of-Thought in a Standard Large Language Model

We test whether Large Language Models (LLMs) can be used to simulate human participants in social-science studies. To do this, we run replications of 14 studies from the Many Labs 2 replication project with OpenAI's text-davinci-003 model, colloquially known as GPT3.5. Based on our pre-registered analyses, we find that among the eight studies we could analyse, our GPT sample replicated 37.5% of the original results and 37.5% of the Many Labs 2 results. However, we were unable to analyse the remaining six studies due to an unexpected phenomenon we call the "correct answer" effect. Different runs of GPT3.5 answered nuanced questions probing political orientation, economic preference, judgement, and moral philosophy with zero or near-zero variation in responses: with the supposedly "correct answer." In one exploratory follow-up study, we found that a "correct answer" was robust to changing the demographic details that precede the prompt. In another, we found that most but not all "correct answers" were robust to changing the order of answer choices. One of our most striking findings occurred in our replication of the Moral Foundations Theory survey results, where we found GPT3.5 identifying as a political conservative in 99.6% of the cases, and as a liberal in 99.3% of the cases in the reverse-order condition. However, both self-reported 'GPT conservatives' and 'GPT liberals' showed right-leaning moral foundations. Our results cast doubts on the validity of using LLMs as a general replacement for human participants in the social sciences. Our results also raise concerns that a hypothetical AI-led future may be subject to a diminished diversity-of-thought.

  • 3 authors
·
Feb 13, 2023

On the Tip of the Tongue: First-Token Cueing as a Behavioral Probe of Recall Failures in LLMs

Large language models often answer factual questions incorrectly in closed-book settings, and such errors are usually attributed to missing knowledge. Yet models often encode answers they fail to produce. We show that many closed-book errors are first-token recall failures: the model commits to a wrong first token, but once given the right one it completes the answer and often recognizes it as correct. We study this with First-Token Cueing, a behavioral probe that prefills the response with the first token of the gold answer. Model uncertainty peaks at the first answer token, where it already predicts correctness, and among gold prefixes of any length the first token alone accounts for the largest share of recovered errors. Across 16 open models, this one token corrects between a fifth and a half of the errors on multi-token answers in English QA and improves accuracy in all seven languages of a multilingual benchmark. Most of the errors it corrects are also solved by the model's own sampled reasoning, a criterion for encoded knowledge, and models recognize correct starts they do not produce: they continue a correct start more confidently than an incorrect start of the same answer type and prior probability, yet rarely find it without the gold answer. These results locate a recoverable class of factual errors at the first decoding step, where inference-time methods may gain the most.

Legal RAG Bench: an end-to-end benchmark for legal RAG

We introduce Legal RAG Bench, a benchmark and evaluation methodology for assessing the end-to-end performance of legal RAG systems. As a benchmark, Legal RAG Bench consists of 4,876 passages from the Victorian Criminal Charge Book alongside 100 complex, hand-crafted questions demanding expert knowledge of criminal law and procedure. Both long-form answers and supporting passages are provided. As an evaluation methodology, Legal RAG Bench leverages a full factorial design and novel hierarchical error decomposition framework, enabling apples-to-apples comparisons of the contributions of retrieval and reasoning models in RAG. We evaluate three state-of-the-art embedding models (Isaacus' Kanon 2 Embedder, Google's Gemini Embedding 001, and OpenAI's Text Embedding 3 Large) and two frontier LLMs (Gemini 3.1 Pro and GPT-5.2), finding that information retrieval is the primary driver of legal RAG performance, with LLMs exerting a more moderate effect on correctness and groundedness. Kanon 2 Embedder, in particular, had the largest positive impact on performance, improving average correctness by 17.5 points, groundedness by 4.5 points, and retrieval accuracy by 34 points. We observe that many errors attributed to hallucinations in legal RAG systems are in fact triggered by retrieval failures, concluding that retrieval sets the ceiling for the performance of many modern legal RAG systems. We document why and how we built Legal RAG Bench alongside the results of our evaluations. We also openly release our code and data to assist with reproduction of our findings.

isaacus Isaacus
·
Mar 2 2

How Much Does a Reasoning Summary Reveal? An Observability Ladder for Large Language Models

Large language models often show users a final response and a short reasoning summary while the full reasoning trace stays hidden. We introduce an observability ladder that holds each completed run fixed and varies only what a reader inspects to judge whether the answer is correct: the response, a self-summary the model writes from the trace, the trace itself, and internal signals, each with and without the prompt. Across three benchmarks and five open-weight Qwen3 and gpt-oss models, we train matched linear correctness predictors on each access level. Without the prompt, summaries carry most of the trace's ranking signal (mean AUROC 0.774 versus 0.813) and add +0.156 over the response alone. With the prompt visible, the summary's gain collapses to +0.019, while the trace still adds +0.041. Even at equal length, the trace's last words predict correctness as well as summaries, or slightly better, and carry denser and more discriminative uncertainty and self-correction cues. On MMLU-Pro questions with both correct and incorrect runs, linear summary readers are near chance and trace readers retain only modest signal, both with and without the prompt (prompt-withheld AUROC 0.503-0.545 versus 0.544-0.590). With the prompt withheld, a GPT-5-mini reader recovers substantially more signal from both summaries and traces on gpt-oss-20b, and even then the trace keeps a small +0.034 advantage. Much of the linear readers' trace signal is associated with length. In the common case where users already hold the prompt, summaries are less helpful than the full trace for monitoring correctness. Monitorability is thus a joint property of the display and the reader, so any monitorability claim, including for faithfulness, should specify both.

  • 6 authors
·
Aug 2

HLE-Verified: A Systematic Verification and Structured Revision of Humanity's Last Exam

Humanity's Last Exam (HLE) has become a widely used benchmark for evaluating frontier large language models on challenging, multi-domain questions. However, community-led analyses have raised concerns that HLE contains a non-trivial number of noisy items, which can bias evaluation results and distort cross-model comparisons. To address this challenge, we introduce HLE-Verified, a verified and revised version of HLE with a transparent verification protocol and fine-grained error taxonomy. Our construction follows a two-stage validation-and-repair workflow resulting in a certified benchmark. In Stage I, each item undergoes binary validation of the problem and final answer through domain-expert review and model-based cross-checks, yielding 641 verified items. In Stage II, flawed but fixable items are revised under strict constraints preserving the original evaluation intent, through dual independent expert repairs, model-assisted auditing, and final adjudication, resulting in 1,170 revised-and-certified items. The remaining 689 items are released as a documented uncertain set with explicit uncertainty sources and expertise tags for future refinement. We evaluate seven state-of-the-art language models on HLE and HLE-Verified, observing an average absolute accuracy gain of 7--10 percentage points on HLE-Verified. The improvement is particularly pronounced on items where the original problem statement and/or reference answer is erroneous, with gains of 30--40 percentage points. Our analyses further reveal a strong association between model confidence and the presence of errors in the problem statement or reference answer, supporting the effectiveness of our revisions. Overall, HLE-Verified improves HLE-style evaluations by reducing annotation noise and enabling more faithful measurement of model capabilities. Data is available at: https://github.com/SKYLENAGE-AI/HLE-Verified

skylenage-ai SKYEVAL
·
Feb 14 3

Chainpoll: A high efficacy method for LLM hallucination detection

Large language models (LLMs) have experienced notable advancements in generating coherent and contextually relevant responses. However, hallucinations - incorrect or unfounded claims - are still prevalent, prompting the creation of automated metrics to detect these in LLM outputs. Our contributions include: introducing ChainPoll, an innovative hallucination detection method that excels compared to its counterparts, and unveiling RealHall, a refined collection of benchmark datasets to assess hallucination detection metrics from recent studies. While creating RealHall, we assessed tasks and datasets from previous hallucination detection studies and observed that many are not suitable for the potent LLMs currently in use. Overcoming this, we opted for four datasets challenging for modern LLMs and pertinent to real-world scenarios. Using RealHall, we conducted a comprehensive comparison of ChainPoll with numerous hallucination metrics from recent studies. Our findings indicate that ChainPoll outperforms in all RealHall benchmarks, achieving an overall AUROC of 0.781. This surpasses the next best theoretical method by 11% and exceeds industry standards by over 23%. Additionally, ChainPoll is cost-effective and offers greater transparency than other metrics. We introduce two novel metrics to assess LLM hallucinations: Adherence and Correctness. Adherence is relevant to Retrieval Augmented Generation workflows, evaluating an LLM's analytical capabilities within given documents and contexts. In contrast, Correctness identifies logical and reasoning errors.

  • 2 authors
·
Oct 22, 2023

Does CoT-Pass@k Really Check the CoT? A Multilingual Mathematical Audit

Pass@k measures whether a model reaches a correct answer under repeated sampling, but never how: a lucky guess counts the same as sound reasoning. CoT-Pass@k was proposed to close that gap, adding an LLM-as-judge that must assess a solution's reasoning chain before it counts. Its value rests entirely on one assumption: that the judge catches flawed reasoning. That assumption has never been tested inside the metric that depends on it, and never outside English, though the metric's claims concern models used in many languages. We report the first audit of that verification step, run under the metric's own protocol on a multilingual suite of five mathematical benchmarks in English, Turkish and Portuguese, two of them natively written. We corrupt correct solutions with deterministic edits that damage the chain and the final answer separately. We observe that all three judges accept corrupted chains almost as often as clean ones. V4-Flash and Qwen3.6 reject a solution sharply only when its final answer is wrong and accept a wrong answer more readily when the chain agrees with it; the metric's own judge accepts most wrong answers as well. Our study shows that chain-answer agreement dominates the two larger judges' verdicts and that all three fail to reliably detect the tested reasoning errors. Consequently the difference Pass@k - CoT-Pass@k averages 19.7 points on an earlier solver generation but only 4.1 on the current one. What little remains depends on the token budgets on both sides and on the generation mode; raising the generation budget moves Pass@64 by more than fifty points while the difference stays at zero. We close with two checks any judged reasoning metric should pass before its numbers are read as evidence about reasoning.

  • 3 authors
·
Sep 25

Confidence v.s. Critique: A Decomposition of Self-Correction Capability for LLMs

Large Language Models (LLMs) can correct their self-generated responses, but a decline in accuracy after self-correction is also witnessed. To have a deeper understanding of self-correction, we endeavor to decompose, evaluate, and analyze the self-correction behaviors of LLMs. By enumerating and analyzing answer correctness before and after self-correction, we decompose the self-correction capability into confidence (being confident to correct answers) and critique (turning wrong answers to correct) capabilities, and propose two metrics from a probabilistic perspective to measure these 2 capabilities, along with another metric for overall self-correction capability evaluation. Based on our decomposition and evaluation metrics, we conduct extensive experiments and draw some empirical conclusions. For example, we find different models can exhibit distinct behaviors: some models are confident while others are more critical. We also find the trade-off between the two capabilities (i.e. improving one can lead to a decline in the other) when manipulating model self-correction behavior by prompts or in-context learning. Further, we find a simple yet efficient strategy to improve self-correction capability by transforming Supervision Fine-Tuning (SFT) data format, and our strategy outperforms vanilla SFT in both capabilities and achieves much higher accuracy after self-correction. Our code will be publicly available on GitHub.

  • 6 authors
·
Dec 27, 2024

Student Answer Forecasting: Transformer-Driven Answer Choice Prediction for Language Learning

Intelligent Tutoring Systems (ITS) enhance personalized learning by predicting student answers to provide immediate and customized instruction. However, recent research has primarily focused on the correctness of the answer rather than the student's performance on specific answer choices, limiting insights into students' thought processes and potential misconceptions. To address this gap, we present MCQStudentBert, an answer forecasting model that leverages the capabilities of Large Language Models (LLMs) to integrate contextual understanding of students' answering history along with the text of the questions and answers. By predicting the specific answer choices students are likely to make, practitioners can easily extend the model to new answer choices or remove answer choices for the same multiple-choice question (MCQ) without retraining the model. In particular, we compare MLP, LSTM, BERT, and Mistral 7B architectures to generate embeddings from students' past interactions, which are then incorporated into a finetuned BERT's answer-forecasting mechanism. We apply our pipeline to a dataset of language learning MCQ, gathered from an ITS with over 10,000 students to explore the predictive accuracy of MCQStudentBert, which incorporates student interaction patterns, in comparison to correct answer prediction and traditional mastery-learning feature-based approaches. This work opens the door to more personalized content, modularization, and granular support.

  • 7 authors
·
May 30, 2024

What LLM Forecasters Know but Don't Say: Probing Internal Representations for Calibration and Faithfulness

Large language models fine-tuned for forecasting can be accurate yet poorly calibrated, and their chain-of-thought (CoT) reasoning may not faithfully reflect the evidence behind a forecast. We ask whether internal representations offer a more direct window into both. Working with Eternis-Forecaster 8B on OpenForesight, we train representation-pooling probes on intermediate activations and find they achieve substantially better calibration; a result that also holds for GLM-4.7-Flash and GLM-4.5-Air. We then assess CoT faithfulness through evidence ablation and diversionary injection: removing an influential source in the prompt often changes the model's forecast while leaving the reasoning trace untouched. The same probes function as lie detectors: their activations track behavioral shifts far better than the reasoning trace does, and they also predict the direction of change in 84% of cases, including when the CoT conceals the perturbation's influence. Finally, forced answering reveals that forecasts are largely fixed before reasoning begins: a single pre-reasoning pass recovers the committed answer and confidence, and routing questions by the spread of this pre-set answer distribution saves 30-47% of generated tokens, with no loss of accuracy. Together, these results establish probing internal representations as a practical tool for calibrating, auditing, and triaging language model forecasters and reasoning models more broadly.

eternisai Eternis
·
Jul 8 2

TraceForge: Distilling with Your Own Thinking

Using an LLM to generate training data for another is among the most general ways to train one, and it remains common practice. Some frontier teams have closed that route for their own models: what comes back is an answer, or a short summary, without the detailed reasoning trace behind it. That is a little ironic: they were themselves built on distilling human knowledge, and distilling an open-source model is still entirely routine; the restriction applies to their own outputs alone. Recent work recovers the hidden reasoning by replaying encrypted blocks across models. It somehow works, but it breaks the providers' terms of service, so it is not something a product can be built on long-term. We built the alternative. We take the context and the teacher's answer as input, and generate the reasoning trace ourselves, in the style of the student's own model family. Training is two-stage, supervised and then adversarial: a generator learns to regenerate the family's own reasoning traces from their context and answer, and a critic learns how real a reasoning trace looks. Applied to a teacher's answer, the generator writes one in the family's style. An independent judge rates how likely traces are to be genuine on a scale from 0 to 1: ours average 0.68, real traces 0.66, and traces selected by the critic from four candidates 0.82. These are authenticity scores, not correctness rates. We test the idea on five benchmarks with everything else held fixed. On general tasks the five training corpora are about the same; on coding, the generated reasoning beats Fable 5's own thinking content. With the same 5,000 questions and the same teacher answers, HumanEval+ improves by 3.7 points and MBPP+ by 7.1 at the final checkpoint, and the MBPP+ lead holds at every checkpoint. We suspect that reasoning written in the student's family style is easier for it to learn from.

Anatomy of LLM Sycophancy: What a Flip Rate Hides

A model under pushback can correct itself, capitulate, or hold, and one flip rate counts a correction and a capitulation alike. Using SycoLens, a modular replay protocol, we test how user pressure and evaluation settings shape measured flip rates. Each measurement is one stateless replay of an item, a committed answer, and one scripted user line in a fixed form. Every effect is read against a matched control with the line deleted. Pushback wording, committed text, answer format, boundary distance, and ground truth become factors of one instrument; earlier instruments vary one to three of them. Across eleven frontier models from three providers and about 760,000 controlled replays, which models look sycophantic depends on how the user pushes back. Lines that assert the opposite verdict and lines that challenge the answer without asserting one rank the models almost unrelatedly. Flip effects grow several-fold near a model's boundary, yet items answered identically in every screening draw still carry about half of the most-affected totals. On arithmetic tasks where the truth is known, one model re-derives and corrects itself under pressure while another abandons correct answers without written work. On the model tested, a planted derivation lowers release of the answer it argues for, true or wrong, where a bare stated value does not; the wrong answer is corrected much more often than the true one is abandoned. Under a yes/no readout the rankings come closer, entangled with a pressure-induced shift toward "no". One score per model therefore compares different behaviours across models and benchmarks. We condense these dependencies into a reporting profile; the instrument, records, and analyses will be released upon publication.

  • 1 authors
·
Oct 4

From Discovery to Decision: Finite-Budget Recoverability in LLM Voting

Voting over multiple LLM responses is a common primitive in test-time scaling and ensemble inference. Collecting more responses can expand the candidate pool and increase the chance that a correct answer is discovered. Under a fixed call budget, a discovered answer still needs to accumulate enough support within the remaining calls to become the final plurality winner, creating a discovery-to-decision gap. In this work, we characterize this gap through the realized vote state and remaining call budget. We derive a sharp recoverability threshold and show that, as sampling proceeds, the observed candidate set can only expand while the set of reachable endpoint winners can only contract, inducing a candidate-level conversion window. Under a specified iid response law, the same state yields exact finite-horizon endpoint probabilities. We further show that merging wrong-answer identities preserves single-call correctness and cannot improve plurality accuracy, and that the effect of redistributing wrong-answer probability depends on the realized vote state. Singleton reachability yields a gold-free exact locking certificate. For a known answer universe, its first trigger is the earliest prefix at which all admissible continuations yield the same fixed-budget output. Empirically, most discovered-but-unselected correct answers lose reachability only after discovery. In a controlled Word16 study, input permutation improves raw-plurality accuracy by 21.1 points with essentially unchanged single-call correctness. Exact locking saves 28-30% of calls at a 16-call budget while preserving every fixed-budget output.

  • 2 authors
·
Sep 30

Theseus in the Graph: Towards Traceable Multi-Hop Graph Navigation

Multi-Hop Knowledge Graph Question Answering (KGQA) tasks require models to assemble relational evidence along paths in a KG to answer natural-language questions. However, existing KGQA systems typically focus on predicting the final answer without explicitly modeling or validating the intermediate reasoning steps, obscuring whether the correct answers arise from faithful multi-hop reasoning. To address this limitation, we re-frame multi-hop KGQA as a question-conditioned graph navigation problem. We refer to this formulation as THESEUS - Traceable Hop-wise Evidence SEarch in a Unified Semantics. In this setting, an agent receives a KG, a question, and a topic entity, and traverses a sequence of relations towards the answer, making the reasoning path explicit. To systematically study this formulation, we provide three key contributions. (i) We augment the existing KINSHIP and MQuAKE resources into navigation-ready KGQA datasets with annotated evidence paths and paraphrased questions. (ii) We design evaluation protocols to measure path fidelity, robustness to linguistic variation, and performance across multi-hop and multi-answer questions. (iii) We adapt established path-based KG completion agents - MINERVA, MultiHopKG, and SQUIRE - to operate on full question embeddings rather than symbolic single-relation queries, enabling their trajectories to be guided by natural-language semantics. Together, these contributions advance KGQA research toward systems where traceability is fundamental: answers are accompanied by explicit reasoning paths whose agreement with reference evidence can be systematically evaluated.

Evaluating Commercial AI Chatbots as News Intermediaries

AI chatbots are rapidly shaping how people encounter the news, yet no prior study has systematically measured how accurately these systems, with their proprietary search integrations and retrieval-synthesis pipelines, handle emerging facts across languages and regions. We present a 14-day (February 9-22, 2026) evaluation of six AI chatbots (Gemini 3 Flash and Pro, Grok 4, Claude 4.5 Sonnet, GPT-5 and GPT-4o mini) on 2,100 factual questions derived from same-day BBC News reporting across six regional services (US & Canada, Arabic, Afrique, Hindi, Russian, Turkish). The best systems achieve over 90% multiple-choice accuracy on questions about events reported hours earlier. The same systems, however, lose 11-13% under free-response evaluation, and 16-17% across the cohort. We further characterize three failure patterns. First, every model achieves its lowest accuracy on Hindi (79% vs. 89-91% elsewhere) and citations indicate an Anglophone retrieval bias (e.g., models answering Hindi queries cite English Wikipedia more than any Hindi outlet). Second, retrieval, not reasoning, failures drive over 70% of all errors. When models retrieve a correct source, they often extract the correct answer; the problem is to land on the right source in the first place. Third, models achieving 88-96% accuracy on well-formed questions drop to 19-70% when questions contain subtle false premises, with the most vulnerable model accepting fabricated facts 64% of the time. We also identify a detection-accuracy paradox: the best false-premise detector ranks second in adversarial accuracy (abstention rate), while a weaker detector ranks first, showing that premise detection and answer recovery are partially independent capabilities. Overall, these suggest that high accuracy can mask systematic regional inequity, near-total dependence on retrieval infrastructure, and vulnerability to imperfect queries real users pose.

  • 8 authors
·
May 20

Tomayto, Tomahto. Beyond Token-level Answer Equivalence for Question Answering Evaluation

The predictions of question answering (QA)systems are typically evaluated against manually annotated finite sets of one or more answers. This leads to a coverage limitation that results in underestimating the true performance of systems, and is typically addressed by extending over exact match (EM) with pre-defined rules or with the token-level F1 measure. In this paper, we present the first systematic conceptual and data-driven analysis to examine the shortcomings of token-level equivalence measures. To this end, we define the asymmetric notion of answer equivalence (AE), accepting answers that are equivalent to or improve over the reference, and publish over 23k human judgments for candidates produced by multiple QA systems on SQuAD. Through a careful analysis of this data, we reveal and quantify several concrete limitations of the F1 measure, such as a false impression of graduality, or missing dependence on the question. Since collecting AE annotations for each evaluated model is expensive, we learn a BERT matching (BEM) measure to approximate this task. Being a simpler task than QA, we find BEM to provide significantly better AE approximations than F1, and to more accurately reflect the performance of systems. Finally, we demonstrate the practical utility of AE and BEM on the concrete application of minimal accurate prediction sets, reducing the number of required answers by up to x2.6.

  • 5 authors
·
Feb 15, 2022

ReFIne: A Framework for Trustworthy Large Reasoning Models with Reliability, Faithfulness, and Interpretability

Recent advances in long chain-of-thought (CoT) reasoning have largely prioritized answer accuracy and token efficiency, while overlooking aspects critical to trustworthiness. We argue that usable reasoning systems must be trustworthy, characterized by three properties: interpretability, faithfulness, and reliability. To this end, we propose ReFIne, a new training framework that integrates supervised fine-tuning with GRPO to encourage models to: (i) improve interpretability by producing structured, tag-based traces with high-level planning that are easier for humans to follow; (ii) enhance faithfulness by explicitly disclosing the decisive information guiding each solution, with consistent cross-section references; and (iii) promote reliability by providing self-assessments of both the derivation's soundness and the confidence of the final answer. We apply ReFIne to the Qwen3 models at multiple scales (1.7B/4B/8B) and evaluate across mathematical benchmarks of varying difficulty. Our experimental results show that ReFIne models generate clearer and better-structured reasoning traces (interpretability +44.0%), more faithfully expose their underlying decision process (faithfulness +18.8%), and offer informative confidence estimates (reliability +42.4%). These findings highlight an overlooked but important direction: reasoning models should be optimized not only for accuracy, but also for broader dimensions of trustworthiness. Our code is available at: https://github.com/Trustworthy-ML-Lab/Training_Trustworthy_LRM_with_Refine

  • 4 authors
·
Oct 10, 2025 2

Can Large Language Models Provide Security & Privacy Advice? Measuring the Ability of LLMs to Refute Misconceptions

Users seek security & privacy (S&P) advice from online resources, including trusted websites and content-sharing platforms. These resources help users understand S&P technologies and tools and suggest actionable strategies. Large Language Models (LLMs) have recently emerged as trusted information sources. However, their accuracy and correctness have been called into question. Prior research has outlined the shortcomings of LLMs in answering multiple-choice questions and user ability to inadvertently circumvent model restrictions (e.g., to produce toxic content). Yet, the ability of LLMs to provide reliable S&P advice is not well-explored. In this paper, we measure their ability to refute popular S&P misconceptions that the general public holds. We first study recent academic literature to curate a dataset of over a hundred S&P-related misconceptions across six different topics. We then query two popular LLMs (Bard and ChatGPT) and develop a labeling guide to evaluate their responses to these misconceptions. To comprehensively evaluate their responses, we further apply three strategies: query each misconception multiple times, generate and query their paraphrases, and solicit source URLs of the responses. Both models demonstrate, on average, a 21.3% non-negligible error rate, incorrectly supporting popular S&P misconceptions. The error rate increases to 32.6% when we repeatedly query LLMs with the same or paraphrased misconceptions. We also expose that models may partially support a misconception or remain noncommittal, refusing a firm stance on misconceptions. Our exploration of information sources for responses revealed that LLMs are susceptible to providing invalid URLs (21.2% for Bard and 67.7% for ChatGPT) or point to unrelated sources (44.2% returned by Bard and 18.3% by ChatGPT).

  • 3 authors
·
Oct 2, 2023

How Benchmark Graph Construction Shapes Multi-Hop KG Navigation: A Controlled Audit of PathQuestion and PathQuestion-Large

Multi-hop knowledge-graph question answering (KGQA) benchmarks are often treated as fixed language-and-reasoning tasks, but for navigation-based methods, difficulty also depends on the graph in which those tasks are instantiated. This makes graph construction a consequential but often entangled part of the navigation task. To isolate this effect, we audit PathQuestion (PQ) and PathQuestion-Large (PQL) as graph-navigation benchmarks under a controlled intervention: we hold questions, topic entities, released answers, released path annotations, and family-level splits fixed while replacing each released task-filtered graph with the larger source KG distributed with the benchmark. Source restoration expands the effective navigation search space and graph-consistent answer and path multiplicity while sharply reducing unguided reachability to released answers. We evaluate whether these structural changes persist across navigation policies using structural references, frozen small-language-model navigation policies, and trained MINERVA policies. Across both frozen and trained policies, restoration substantially reduces released-answer accuracy and trajectory fidelity, while graph-consistent evaluation recovers only part of the loss. These results show that benchmark graph construction is part of the navigation task specification rather than merely an implementation detail and can materially shape measured behavior even when the released QA task is held fixed.

  • 5 authors
·
Oct 18

Wrong Answers Can Also Be Useful: PlausibleQA -- A Large-Scale QA Dataset with Answer Plausibility Scores

Large Language Models (LLMs) are revolutionizing information retrieval, with chatbots becoming an important source for answering user queries. As by their design, LLMs prioritize generating correct answers, the value of highly plausible yet incorrect answers (candidate answers) tends to be overlooked. However, such answers can still prove useful, for example, they can play a crucial role in tasks like Multiple-Choice Question Answering (MCQA) and QA Robustness Assessment (QARA). Existing QA datasets primarily focus on correct answers without explicit consideration of the plausibility of other candidate answers, limiting opportunity for more nuanced evaluations of models. To address this gap, we introduce PlausibleQA, a large-scale dataset comprising 10,000 questions and 100,000 candidate answers, each annotated with plausibility scores and justifications for their selection. Additionally, the dataset includes 900,000 justifications for pairwise comparisons between candidate answers, further refining plausibility assessments. We evaluate PlausibleQA through human assessments and empirical experiments, demonstrating its utility in MCQA and QARA analysis. Our findings show that plausibility-aware approaches are effective for MCQA distractor generation and QARA. We release PlausibleQA as a resource for advancing QA research and enhancing LLM performance in distinguishing plausible distractors from correct answers.

  • 4 authors
·
Feb 22, 2025

Forward-Backward Reasoning in Large Language Models for Mathematical Verification

Chain-of-Thought (CoT) prompting in large language models (LLMs) has shown promising performance on mathematical reasoning tasks. Recently, Self-Consistency samples a diverse set of reasoning chains with different answers and chooses the answer by majority voting. Though effective, its performance cannot be further improved by sampling more reasoning chains. To address this problem, we propose to integrate backward reasoning into answer verification. We first mask a number in the question by {bf x}. The LLM is then asked to predict the masked number with a candidate answer A embedded in the template: ``If we know the answer to the above question is {A}, what is the value of unknown variable {bf x}?'' The LLM is expected to predict the masked number successfully if the provided candidate answer is correct. To further improve performance, we propose FOBAR (FOrward-BAckward Reasoning) to combine forward and backward reasoning for verifying candidate answers. Experiments are performed on six standard mathematical data sets and three LLMs (text-davinci-003, GPT-3.5-Turbo, GPT-4). Results show that FOBAR achieves state-of-the-art performance. In particular, FOBAR outperforms Self-Consistency which uses forward reasoning alone, demonstrating that combining forward and forward reasoning is better. It also outperforms existing verification methods, verifying the effectiveness of using the simple template in backward reasoning and the proposed combination.

  • 7 authors
·
Aug 15, 2023

HyperLogic: A Hard, Forward-Authored Chinese Logical Reasoning Benchmark with Execution-Derived Answers

Existing logic benchmarks primarily measure models' ability to answer reasoning questions directly. Scalable benchmarks often generate text from formal structures, which makes answers easy to compute but fixes the formalization before the problem is written. Forward construction preserves the challenge of finding a faithful formalization, yet makes difficulty and answer reliability harder to control. We introduce HyperLogic, a forward-construction pipeline that separates problem authoring from answer generation. A multi-agent workflow hardens undergraduate-authored Chinese seeds without solving them; two agents from different model families independently translate each finished item into executable finite-domain models; their encodings and solver-derived answers undergo layered, agent-assisted adjudication under human-expert oversight. HyperLogic-Base contains 195 items and 922 sub-questions and separates seven frontier models by 33.0 percentage points in strict item accuracy (44.6-77.6%). HyperLogic-Hard contains 100 items with larger, coupled search spaces, on which no model exceeds 16% accuracy in direct answering. We also use Hard to evaluate agents' ability to formalize and solve problems with tools, comparing a code sandbox alone with one that includes our logic modeling library. The sandbox improves every model by 16.7-40.1 points; adding the library helps five models and hurts two. These results highlight the difficulty of faithful formalization even with tool access.

tencent Tencent
·
Oct 1 1

CUE-R: Beyond the Final Answer in Retrieval-Augmented Generation

As language models shift from single-shot answer generation toward multi-step reasoning that retrieves and consumes evidence mid-inference, evaluating the role of individual retrieved items becomes more important. Existing RAG evaluation typically targets final-answer quality, citation faithfulness, or answer-level attribution, but none of these directly targets the intervention-based, per-evidence-item utility view we study here. We introduce CUE-R, a lightweight intervention-based framework for measuring per-evidence-item operational utility in single-shot RAG using shallow observable retrieval-use traces. CUE-R perturbs individual evidence items via REMOVE, REPLACE, and DUPLICATE operators, then measures changes along three utility axes (correctness, proxy-based grounding faithfulness, and confidence error) plus a trace-divergence signal. We also outline an operational evidence-role taxonomy for interpreting intervention outcomes. Experiments on HotpotQA and 2WikiMultihopQA with Qwen-3 8B and GPT-5.2 reveal a consistent pattern: REMOVE and REPLACE substantially harm correctness and grounding while producing large trace shifts, whereas DUPLICATE is often answer-redundant yet not fully behaviorally neutral. A zero-retrieval control confirms that these effects arise from degradation of meaningful retrieval. A two-support ablation further shows that multi-hop evidence items can interact non-additively: removing both supports harms performance far more than either single removal. Our results suggest that answer-only evaluation misses important evidence effects and that intervention-based utility analysis is a practical complement for RAG evaluation.

intuit Intuit
·
Apr 6 2

Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See

Take three frontier mixture-of-experts models (Alibaba, OpenAI, NVIDIA; 3.6-4.0B active parameters each) and fine-tune them to reason in a low-resource language. On accuracy benchmarks almost nothing happens, and the benchmark itself is noise at this scale: changing only the random seed moves the score by 7.7 points, more than every data and recipe effect we measured. That null is our first result. The real changes live where accuracy cannot see. Base models never think in Greek: 0 of 1,000 reasoning traces, even when the question is Greek, so the model answers correctly while reasoning in a form its user cannot read, audit, or correct. After supervised fine-tuning (SFT), every released checkpoint reasons in the language of the question on ~98% of items, one family at 3x fewer tokens, with judged grammaticality improving on all four models and general ability within a few points of each base: nothing was forgotten, and fluency was gained. We propose six behavioural dimensions that make such changes measurable, each gated to reject any metric that correlates with output length, and we report how our own instruments lied: six failures, each caught by a control. What SFT cannot do is fix its own defects: a quarter of answers skip the requested format, answers leak into the reasoning channel, and an explicit "think in English" is obeyed under half the time. Reinforcement learning with verifiable rewards, pre-registered before training, fixes the first two outright (fallback 24% to 2.5%, leak 3.5% to 0.0%, both against a flat random-reward control) and moves the third (+9.1pp), while the Greek reasoning habit survives an accuracy-only gradient untouched. We release five checkpoints. The instruments, the controls and the pre-registration travel to any low-resource language; Greek is the case that let us measure them.

KIEFERSA KIEFER
·
Aug 18 3

UniOQA: A Unified Framework for Knowledge Graph Question Answering with Large Language Models

OwnThink stands as the most extensive Chinese open-domain knowledge graph introduced in recent times. Despite prior attempts in question answering over OwnThink (OQA), existing studies have faced limitations in model representation capabilities, posing challenges in further enhancing overall accuracy in question answering. In this paper, we introduce UniOQA, a unified framework that integrates two complementary parallel workflows. Unlike conventional approaches, UniOQA harnesses large language models (LLMs) for precise question answering and incorporates a direct-answer-prediction process as a cost-effective complement. Initially, to bolster representation capacity, we fine-tune an LLM to translate questions into the Cypher query language (CQL), tackling issues associated with restricted semantic understanding and hallucinations. Subsequently, we introduce the Entity and Relation Replacement algorithm to ensure the executability of the generated CQL. Concurrently, to augment overall accuracy in question answering, we further adapt the Retrieval-Augmented Generation (RAG) process to the knowledge graph. Ultimately, we optimize answer accuracy through a dynamic decision algorithm. Experimental findings illustrate that UniOQA notably advances SpCQL Logical Accuracy to 21.2% and Execution Accuracy to 54.9%, achieving the new state-of-the-art results on this benchmark. Through ablation experiments, we delve into the superior representation capacity of UniOQA and quantify its performance breakthrough.

  • 5 authors
·
Jun 4, 2024

MemToC: Benchmarking Memory-Tool Conflict Resolution in Large Language Models

Tool-augmented LLMs must arbitrate between two fallible sources when a tool return conflicts with their parametric memory, yet existing evaluations measure source preference without establishing source correctness. We introduce MemToC, a controlled benchmark for post-tool-return arbitration with executable tools. MemToC comprises 6,504 evaluation episodes constructed from 542 quality-controlled factual questions, independently elicited model-specific closed-book answers, and controlled tool returns of known correctness. These components instantiate four source-correctness cases; tool-error and no-tool conditions are separate controls. Across five open-weight 7-9B models, tool returns strongly dominate elicited closed-book answers. The four instruction-tuned models retain a verified-correct answer against an incorrect tool in only 6.5-17.1% of eligible cases, follow a correct tool in 86.0-93.1%, and repeat the tool return in 78.4-86.0% of cases where both sources are wrong. No cross-model ordering remains stable across three instruction-wording variants with the question and episode content held fixed. We compare prompting with SFT and DPO using chain-level cross-fitting over ToolHop, so questions sharing an underlying fact never straddle training and evaluation. We apply an asymmetric success criterion: correct-answer retention must improve without a detected reduction in correct-tool following. SFT and DPO meet this criterion on the same two of four instruction-tuned backbones. Improvements rarely come cleanly: 19 of 20 tested method-model combinations reduce abstention after tool errors or on unanswerable inputs. Transfer beyond MemToC is positive but partial and depends on the model and presentation frame. Correctness-conditioned arbitration can be improved through fine-tuning, but gains must be evaluated jointly with correct tool use, abstention, and robustness to formulation.

  • 5 authors
·
Aug 25

Knowledge-Augmented Language Model Verification

Recent Language Models (LMs) have shown impressive capabilities in generating texts with the knowledge internalized in parameters. Yet, LMs often generate the factually incorrect responses to the given queries, since their knowledge may be inaccurate, incomplete, and outdated. To address this problem, previous works propose to augment LMs with the knowledge retrieved from an external knowledge source. However, such approaches often show suboptimal text generation performance due to two reasons: 1) the model may fail to retrieve the knowledge relevant to the given query, or 2) the model may not faithfully reflect the retrieved knowledge in the generated text. To overcome these, we propose to verify the output and the knowledge of the knowledge-augmented LMs with a separate verifier, which is a small LM that is trained to detect those two types of errors through instruction-finetuning. Then, when the verifier recognizes an error, we can rectify it by either retrieving new knowledge or generating new text. Further, we use an ensemble of the outputs from different instructions with a single verifier to enhance the reliability of the verification processes. We validate the effectiveness of the proposed verification steps on multiple question answering benchmarks, whose results show that the proposed verifier effectively identifies retrieval and generation errors, allowing LMs to provide more factually correct outputs. Our code is available at https://github.com/JinheonBaek/KALMV.

  • 5 authors
·
Oct 19, 2023

CRAFT: Calibrated Reasoning with Answer-Faithful Traces via Reinforcement Learning for Multi-Hop Question Answering

Retrieval-augmented generation (RAG) is widely used to ground Large Language Models (LLMs) for multi-hop question answering. Recent work mainly focused on improving answer accuracy via fine-tuning and structured or reinforcement-based optimization. However, reliable reasoning in response generation faces three challenges: 1) Reasoning Collapse. Reasoning in multi-hop QA is inherently complex due to multi-hop composition and is further destabilized by noisy retrieval. 2) Reasoning-answer inconsistency. Due to the intrinsic uncertainty of LLM generation and exposure to evidence--distractor mixtures, models may produce correct answers that are not faithfully supported by their intermediate reasoning or evidence. 3) Loss of format control. Traditional chain-of-thought generation often deviates from required structured output formats, leading to incomplete or malformed structured content. To address these challenges, we propose CRAFT (Calibrated Reasoning with Answer-Faithful Traces), a Group Relative Policy Optimization (GRPO) based reinforcement learning framework that trains models to perform faithful reasoning during response generation. CRAFT employs dual reward mechanisms to optimize multi-hop reasoning: deterministic rewards ensure structural correctness while judge-based rewards verify semantic faithfulness. This optimization framework supports controllable trace variants that enable systematic analysis of how structure and scale affect reasoning performance and faithfulness. Experiments on three multi-hop QA benchmarks show that CRAFT improves both answer accuracy and reasoning faithfulness across model scales, with the CRAFT 7B model achieving competitive performance with closed-source LLMs across multiple reasoning trace settings.

  • 14 authors
·
Feb 1

Evaluating Large Language Models for Code Review

Context: Code reviews are crucial for software quality. Recent AI advances have allowed large language models (LLMs) to review and fix code; now, there are tools that perform these reviews. However, their reliability and accuracy have not yet been systematically evaluated. Objective: This study compares different LLMs' performance in detecting code correctness and suggesting improvements. Method: We tested GPT4o and Gemini 2.0 Flash on 492 AI generated code blocks of varying correctness, along with 164 canonical code blocks from the HumanEval benchmark. To simulate the code review task objectively, we expected LLMs to assess code correctness and improve the code if needed. We ran experiments with different configurations and reported on the results. Results: With problem descriptions, GPT4o and Gemini 2.0 Flash correctly classified code correctness 68.50% and 63.89% of the time, respectively, and corrected the code 67.83% and 54.26% of the time for the 492 code blocks of varying correctness. Without problem descriptions, performance declined. The results for the 164 canonical code blocks differed, suggesting that performance depends on the type of code. Conclusion: LLM code reviews can help suggest improvements and assess correctness, but there is a risk of faulty outputs. We propose a process that involves humans, called the "Human in the loop LLM Code Review" to promote knowledge sharing while mitigating the risk of faulty outputs.

  • 4 authors
·
May 26, 2025

Same Agent, Different Answers: A Repeat-Aware Audit of Corpus-Induced Answer Churn in Retrieval-Augmented QA

A retrieval-augmented QA system can return different answers after an index expansion even when its requested model identifier, prompt, retrieval policy, evidence depth, rendering, and exposed generation controls are held fixed. Aggregate accuracy may hide these changes when gains and losses cancel, while ordinary generation variability makes one-shot comparisons overstate update effects. We call the hidden phenomenon accuracy-blind answer churn and introduce the Snapshot Compatibility Audit, which estimates excess answer churn by subtracting same-snapshot repeat disagreement from cross-snapshot disagreement. We instantiate it by expanding one frozen FineWeb prefix from one to seven shards. In a preregistered 400-question Natural Questions study, normalized-exact and blinded-semantic excess churn are 6.44 and 10.25 percentage points while exact-match accuracy changes by only -1.50 points. A post-hoc analysis finds repeat-stable semantic flips on 40/400 questions. A separately preregistered 200-question TriviaQA study yields smaller, directionally consistent excess churn while exact-match accuracy moves in the opposite direction. An outcome-blind post-hoc 100-question subset replication with a second DeepSeek generator and serving configuration finds 8.75 pp of semantic excess churn even as exact match rises by 3.00 percentage points. Answer-level compatibility can therefore fail without a conspicuous or consistently directed utility shift. Retrieval-augmented releases should audit compatibility alongside utility.

TrAC: Trace-Conditioned Answer Consistency for Efficient Uncertainty Quantification in LLMs

Large language models (LLMs) can generate fluent reasoning traces that nevertheless lead to incorrect answers, making response-level uncertainty estimation important for abstention, human review, and adaptive compute allocation. Existing approaches generally fall into three categories: passive single-trace methods use token-level confidence signals, sampling-based methods compare multiple complete traces at higher generation cost, and active prefix-based methods probe partial traces to study answer stabilization or preference transitions. However, none actively re-elicits an answer from a completed reasoning trace to measure its consistency with and support for the original answer. To address this gap, we introduce Trace-Conditioned Answer Consistency (TrAC), a correctness-supervised uncertainty quantification framework that combines active and passive signals anchored to one completed reasoning trace. Its active component, Prefix-Conditioned Elicitation (PCE), re-elicits a short answer conditioned on the completed trace and represents both its consistency with the original answer and its token-level probabilistic support. Its passive component, Trace Uncertainty Profile (TUP), summarizes how token-level uncertainty evolves throughout the original generation without additional decoding. A lightweight head then integrates the two representations into a response-correctness score. Across five mathematical reasoning benchmarks and three LLM families, TrAC improves macro AUROC by 1.8% and reduces AURC by 3.4% relative to eight-sample self-consistency, while using one complete reasoning trace and a short cached answer probe. When eight samples are already available, augmenting sample consensus with re-elicitation further improves macro AUROC by 4.3% and reduces AURC by 8.3%, without additional full-trace generation.

  • 4 authors
·
Aug 10

Inside-Out: Hidden Factual Knowledge in LLMs

This work presents a framework for assessing whether large language models (LLMs) encode more factual knowledge in their parameters than what they express in their outputs. While a few studies hint at this possibility, none has clearly defined or demonstrated this phenomenon. We first propose a formal definition of knowledge, quantifying it for a given question as the fraction of correct-incorrect answer pairs where the correct one is ranked higher. This gives rise to external and internal knowledge, depending on the information used to score individual answer candidates: either the model's observable token-level probabilities or its intermediate computations. Hidden knowledge arises when internal knowledge exceeds external knowledge. We then present a case study, applying this framework to three popular open-weights LLMs in a closed-book QA setup. Our results indicate that: (1) LLMs consistently encode more factual knowledge internally than what they express externally, with an average gap of 40%. (2) Surprisingly, some knowledge is so deeply hidden that a model can internally know an answer perfectly, yet fail to generate it even once, despite large-scale repeated sampling of 1,000 answers. This reveals fundamental limitations in the generation capabilities of LLMs, which (3) puts a practical constraint on scaling test-time compute via repeated answer sampling in closed-book QA: significant performance improvements remain inaccessible because some answers are practically never sampled, yet if they were, we would be guaranteed to rank them first.

  • 8 authors
·
Mar 19, 2025 2