arXiv subjects

Dekun Yang

Publications and source records attributed to Dekun Yang.

5 recordsLinked to original sources

arXiv2026

A Calibrated Test of Internal Action Maps: State Signals Without Global Affine Closure

A hidden state signal can be decodable or causally usable without supporting a reusable action map. We test whether action maps fitted without a source reach its natural post-action activation and compose. We organize the tests as an evidence lattice and validate the geometric branch on a known affine S_5 carrier: all held-source folds pass one-step, composition, inverse, decoding, and commutativity gates. Structured curvature and held-domain conjugacy raise error monotonically, but only 23/30 strongest cells flip a closure gate, bounding rather than universalizing calibration. In post-trained Qwen/Qwen3-4B, frozen final-token h28 affine maps have mean held-entity error .519, versus .398 for within-test-domain cross-fit. Seven randomized entity splits and map geometry do not support a purely entity-specific account. Earlier h4/h16 layers fit one-step transitions better, but h4 conflict-state decoding is weak and lexical controls remain unresolved. Three matched intervention datasets regenerated from one frozen checkpoint show causal effects only at h28/h36. Outcome-aware refitting improves h28 one-step error to .474 (.469 with weighting), yet no refit passes composition. Learned finite worlds likewise preserve relative algebraic signals or shared charts without held-source affine closure. Within the tested carriers, state availability, causal use, local geometry, and reusable closure are separable. The result is limited to one pretrained model, sampled final-token layers, two finite worlds, and the tested affine or diagnostic function classes.

cs.AI↗

arXiv2026

Prompt Framing Distorts Count Based Evaluation of LLM Error Detection: Evidence from Numeric Anchoring

Count-based F1 is widely used as a proxy for LLM error-detection quality, but this paper shows that it can rise dramatically without a corresponding improvement in span localization, a gap termed F1 Inflation. The paper introduces ErrorBench, a controlled stress-test protocol for prompt-induced count distortion. ErrorBench evaluates six contemporary LLMs under five prompt conditions over 4,290 responses from 143 CoNLL-2014 passages. Under CoNLL-2014 M2-style scoring, anchored prompts produce up to 0.79 points of F1 Inflation, and up to 0.96 under strict matching. A 100-passage replication using the official ERRANT 3.0.0 pipeline and multi-reference scoring reproduces the pattern: averaged over six models, the Blind-to-Anchored prompt shift raises Count-F1 by +0.21 while raising multi-reference ERRANT F0.5 by only +0.04. The study finds larger count responses in highly instruction-compliant GPT/Claude systems and smaller responses in the Gemini family under this stress-test protocol. The findings suggest that LLM proofreading and document-review evaluations should avoid pre-populated error counts and should report span-aware metrics alongside count-based metrics.

cs.CL↗

arXiv2026

Calibrated Selective Fact-Checking via Evidence Chain Evaluation

Large language models (LLMs) can achieve strong fact-checking accuracy, yet forced binary decisions conceal a critical reliability problem: systems may issue confident verdicts even when supporting evidence is weak, sparse, or internally inconsistent. We address this issue through Evidence Chain Evaluation (ECE), a selective fact-checking framework that permits abstention via an uncertain verdict instead of requiring a true/false decision for every claim. The evaluated system is a tool-using verification agent that gathers evidence through web search, scholarly search, and executable checks, and then returns a structured verdict with confidence and source-level metadata. On ECE-Bench, ECE achieves 91.6% standard accuracy, 93.7% coverage, and 97.8% selective accuracy on answered claims. Although ECE does not outperform the strongest retrieval baseline on aggregate calibration metrics such as Expected Calibration Error, Brier score, or AURC, it delivers a clear selective-prediction trade-off: the system maintains very high accuracy on answered claims while deferring 6 of 95 cases. These deferred cases are concentrated in lower-reliability evidence settings (5/6 at source level L4), supporting the view that abstention functions as a safety-oriented mechanism for handling epistemically weak evidence. Code is available at https://github.com/ cheshireyang/ECE.git

cs.AI↗

arXiv2024

Tuning electronic and optical properties of 2D polymeric C$_{60}$ by stacking two layers

Benefiting from improved stability due to stronger interlayer van der Waals interactions, few-layer fullerene networks are experimentally more accessible compared to monolayer polymeric C$_{60}$. However, there is a lack of systematic theoretical studies on the material properties of few-layer C$_{60}$ networks. Here, we compare the structural, electronic and optical properties of bilayer and monolayer fullerene networks. The band gap and band-edge positions remain mostly unchanged after stacking two layers into a bilayer, enabling the bilayer to be almost as efficient a photocatalyst as the monolayer. The effective mass ratio along different directions is varied for conduction band states due to interlayer interactions,leading to enhanced anisotropy in carrier transport. Additionally, stronger exciton absorption is found in the bilayer than that in the monolayer over the entire visible light range, rendering the bilayer a more promising candidate for photovoltaics. Moreoever, the polarisation dependence of optical absorption in the bilayer is increased in the red-yellow light range, offering unique opportunities in photonics and display technologies with tailored optical properties over specific directions. Our study provides strategies to tune electronic and optical properties of 2D polymeric C$_{60}$ via the introduction of stacking degrees of freedom.

cond-mat.mtrl-sci↗

arXiv2020

MRIF: Multi-resolution Interest Fusion for Recommendation

The main task of personalized recommendation is capturing users' interests based on their historical behaviors. Most of recent advances in recommender systems mainly focus on modeling users' preferences accurately using deep learning based approaches. There are two important properties of users' interests, one is that users' interests are dynamic and evolve over time, the other is that users' interests have different resolutions, or temporal-ranges to be precise, such as long-term and short-term preferences. Existing approaches either use Recurrent Neural Networks (RNNs) to address the drifts in users' interests without considering different temporal-ranges, or design two different networks to model long-term and short-term preferences separately. This paper presents a multi-resolution interest fusion model (MRIF) that takes both properties of users' interests into consideration. The proposed model is capable to capture the dynamic changes in users' interests at different temporal-ranges, and provides an effective way to combine a group of multi-resolution user interests to make predictions. Experiments show that our method outperforms state-of-the-art recommendation methods consistently.

cs.IR↗