Research Papers

Structure-Regularized Interpretable TCR-Epitope Prediction

Jiarui Li, Zixiang Yin, Yunbei Zhang +4

T cell receptor (TCR)-epitope binding prediction is essential for understanding adaptive immunity and developing immunotherapies. Existing sequence- and structure-based models often generalize poorly to unseen epitopes and provide limited interpretability. Furthermore, the impact of generated structures on model learning remains unclear. We present TCR-SRIM, a structure-regularized interpretable-by-design model that combines protein language model embeddings with interpretable contact prototypes to capture residue-level TCR-epitope interactions. TCR-SRIM achieves state-of-the-art predictive performance and improved interpretation quality on the TCR-XAI benchmark. Using its inherent interpretability, we further evaluate the effect of generated structures on model learning. While structures predicted by AlphaFold3, TCRModel2, and tFold-TCR yield competitive performance, they lead to less accurate interaction patterns and reduced binding-site diversity than experimentally-resolved structures. Our results highlight limitations of current structure prediction models for TCR-epitope learning and demonstrate the value of interpretable-by-design models for studying generated biological structures.

Do language models know when they are being tested? This question matters for AI safety: a model that recognises an evaluation context could alter its behaviour strategically, making downstream benchmarks harder to interpret. Using 11 models spanning Qwen 2.5, Gemma 2, and Llama 3.2, we find a systematic size-dependent shift in representational depth: in both Qwen 2.5 and Gemma 2, the layer at which evaluation-awareness is most linearly recoverable moves from late layers in smaller models to early layers in larger ones. This suggests that scale changes not only the strength of evaluation-awareness but also where it is most linearly recoverable in the network. This depth shift helps explain why within-family scaling trajectories are non-monotonic or inverse rather than smooth and family-general, showing that a simple universal power-law account is not supported under denser within-family sampling. Finally, white-box probe signals are consistently stronger than black-box behavioural expression, and the relationship between the two varies by family in ways not predicted by probe AUROC alone.

Path-based attribution methods such as Integrated Gradients (IG) are widely adopted for their strong axiomatic properties and effectiveness in attributing model predictions to input features by integrating gradients along a path from a baseline to the input. However, the choice of the attribution path largely affects the quality of explanations, and existing approaches rely on fixed or hand-crafted paths that often produce noisy or distorted attributions. To address this limitation, we propose Diffusion Integrated Gradients (DiffIG), a novel method that reformulates path generation as a conditional generative modeling problem. DiffIG first trains a diffusion model to learn a distribution over paths generated from a Stick-Breaking Process, then employs guided sampling to embed user guidance during the sampling procedure. We demonstrate that DiffIG quantitatively matches or outperforms existing path-based methods, achieving perceptually aligned explanations. This work introduces a new generative perspective for flexible, inference-time controllable Explainable Artificial Intelligence (XAI) methods.

Transparency is increasingly mandated for public-sector AI systems, with organisations required to publish statements describing their AI use and oversight arrangements. However, the existence of such artefacts is often treated as equivalent to transparency itself, despite limited evidence that they proportionately serve relevant stakeholder groups. From a requirements engineering perspective, this raises a validation concern: compliance with mandated disclosure criteria does not necessarily ensure transparency adequacy for stakeholders with different levels of risk exposure, decision control, and involvement. This paper presents an empirical analysis of 92 publicly available AI transparency statements published by Australian Government agencies under the national AI governance mandate. We introduce the stakeholder Risk--Control--Involvement--Need (RCIN) framework to differentiate stakeholder classes according to their structural position and transparency needs. Using a structured rubric derived from the mandated criteria, we evaluate how both the mandate and published statements are calibrated to each stakeholder class. The findings show that while structural compliance is widespread, transparency calibration is uneven. Criteria serving high-control stakeholders are consistently realised, whereas criteria most critical for high-risk, low-control stakeholders are fewer and less substantively addressed. We conceptualise this as the Transparency Illusion: a condition in which transparency appears satisfied through compliant artefacts yet remains unevenly calibrated to stakeholders bearing the greatest exposure to AI-supported decisions. The study frames transparency as a stakeholder-calibrated validation problem, demonstrating that artefact-level compliance does not constitute requirements validation in this context.

LLM supervision systems, namely input/output moderation filters and jailbreak detectors, are the primary safeguard against misuse in deployed AI applications, yet existing benchmarks are often vendor-biased, omit cost and latency, and rarely compare specialized guardrails against repurposed generalist LLMs. We present BELLS-O (Benchmark for the Evaluation of LLM Supervision Systems, Operational), the first independent operational benchmark of LLM supervision systems. BELLS-O evaluates 28 systems from 17 providers: every major specialized guardrail (e.g., LlamaGuard-4, ShieldGemma-2, Lakera Guard) and frontier generalists repurposed as supervisors (e.g., GPT-5.4, Claude Sonnet 4.6, Grok-4.1), jointly on detection rate, false-positive rate, latency, and monetary cost. We cover input/output moderation across 11 harm categories and jailbreak detection across 13 attack techniques, using in-house datasets built from handcrafted prompts, expert-curated samples, and quality-controlled synthetic generation. To suppress latent generator fingerprints in synthetic data, every generated sample is paraphrased. Mapping the Pareto frontier reveals use-case-dependent tradeoffs. On content moderation, specialized supervisors are operationally dominant: top systems match frontier LLMs on detection (~95% vs. 94%) at comparably low false-positive rates (<=2%), while running 5-10x faster and ~10x cheaper. On jailbreak detection, the tradeoff shifts: frontier LLMs achieve higher detection and lower false-positive rates but at 10-50x higher cost and 5-10x higher latency. We release the benchmark, framework, leaderboard, and datasets as the first vendor-neutral basis for selecting safeguards under real deployment constraints.

AI models deployed in critical domains, such as AI safety research, may subtly sabotage our efforts due to misalignment. Diffuse AI Control is a subfield of AI safety concerned with mitigating risks from AI sabotage distributed over long deployment horizons (diffuse threats). These risks are particularly pernicious on fuzzy tasks, i.e. tasks which are hard to grade or require intuition. To understand diffuse threats on fuzzy tasks, we introduce a framework that considers AI control as an adversarial game between a blue team and a red team. The blue team uses a weak trusted model to construct a weak score against which they would train a strong, potentially subversive model to remove the subversion propensity if it were present. The red team then tries to find model behaviors that are rated highly by the weak score, and thus might not be trained out, but actually correspond to poor performance. We test our framework on the task of writing experimental proposals for research questions from recent ML papers. We use a language model with access to the original paper as a proxy "ground-truth" scorer. Our red team discovers subversive behaviors using multi-objective evolutionary prompt optimization. We show that Opus~4.6 can write proposals that are worse according to the ground truth proxy than those of GPT-OSS-20B, while the weak scorer rates them as highly as the best proposals from Opus 4.6. We then propose an adversarial optimization algorithm for the blue team that discovers more robust prompts for the weak model. This algorithm produces a blue team prompt that our red team optimization fails to exploit.

Automated L2 speech assessment can assign proficiency labels, but often lacks interpretability. We propose a rubric-guided SpeechLLM for multi-aspect, multi-granular assessment, trained with a hybrid objective combining supervised fine-tuning and Bounded Direct Preference Optimization. The model jointly predicts ordinal labels at the sentence-level (accuracy, fluency, prosody), word/phoneme-level accuracy, and generates a natural-language rationale in the same response. On SpeechOcean762, our approach matches or outperforms single-granularity models while remaining competitive with prior approaches. We analyze rationale reliability along two axes: self-consistency with model predictions and alignment with ground-truth labels, using sentiment consistency (plausibility) and mention-based agreement (faithfulness). Rationales are plausible at the sentence level, but faithfulness degrades at the word/phoneme level: references are sparse and weakly aligned with token-level labels.

Frontier large language models can often recognize when they are being evaluated, a capability known as evaluation awareness. If models behave differently in evaluations than in deployment, this undermines the validity of evaluation results, which are a crucial component of current AI safety frameworks. We introduce EvalDetectBench, an open pipeline and benchmark for measuring evaluation awareness that works with any Inspect-compatible evaluation, allowing practitioners to test against current and future benchmarks. EvalDetectBench ships with a newly curated transcript suite covering current frontier system-card evaluations and diverse deployment sources. The benchmark serves two purposes: measuring how reliably frontier LLMs recognize that they are being evaluated, and assessing how detectable individual benchmarks are as evaluations. We identify two methodological choices in the existing literature that introduce systematic bias: the identity of the model that generated the deployment transcripts accounts for 11.25% of measurement variance and can reorder model rankings; and elicitation prompts selected for high performance on one model can perform near chance on others. EvalDetectBench corrects for both via per-model probe calibration and a stratified generator-harmonisation procedure.

In sixth-generation (6G) networks, billions of cyber-physical systems (CPSs) - autonomous vehicles, smart grids, industrial robots, and remote-surgical equipment - will run over ultra-reliable low-latency slices, collapsing the gap between a remote breach and physical harm to milliseconds, a budget perimeter firewalls and centralised security operations centres cannot meet. This survey reframes 6G CPS security as a closed-loop, AI-native pipeline that senses at the multi-access edge computing (MEC) tier, using minute-scale call-detail records (CDRs) for baseline learning and sub-millisecond RAN/Open-RAN (O-RAN) telemetry for the latency-critical path. It decides locally with compressed deep models, mitigates network-wide via SDN, NFV, and O-RAN controllers, and retrains through federated learning (FL) and digital-twin (DT) replay. We formalise a per-slice, tail-bounded latency contract on the sense, detect, and mitigate stages, enforced at a slice-dependent tail percentile (p99 for safety-critical URLLC slices). Organising 128 peer-reviewed studies (2017-2026) under a PRISMA 2020 protocol, we (i) map the 6G/CPS threat surface to MITRE ATT&CK and a CDR-observable feature space; (ii) unify edge anomaly detection and DDoS classification across twelve datasets and statistical, graph, and transformer models; (iii) synthesise SDN/NFV/O-RAN primitives into one closed-loop reference architecture; (iv) treat FL, large language models (LLMs), DT, post-quantum cryptography (PQC), zero-trust architecture (ZTA), and explainable AI as cross-cutting enablers, not parallel pillars; and (v) consolidate open problems into five directions spanning data, latency, trust, standardisation, and evaluation.

Online Agent-as-a-Judge: Situation-Generating Evaluation for Interactive Agents

Hyogon Ryu, Jeonghwan Kim, Yewon Lim +3

Evaluating LLM-powered interactive social agents is challenging because socially relevant behaviors depend not only on isolated outputs, but also on prior interactions, social roles, and downstream actions. Existing methods typically allow a target agent to act freely in an environment and then score the resulting trajectory. However, this passive setup can miss capabilities that only become observable under specific social circumstances; for example, conflict handling may remain untested if no disagreement arises. We propose Online Agent-as-a-Judge, a situation-generating evaluation framework for interactive social agents. Online Agent-as-a-Judge deploys an in-world evaluator agent that interacts with the target agent through the environment's native dialogue and action protocol, actively eliciting situations relevant to the evaluation criteria. The resulting trajectories provide evidence for assessing both immediate responses and subsequent behavior. In a life-simulation environment with $32$ designer-authored social criteria, Online Agent-as-a-Judge improves criteria coverage and agreement with human labels, yielding more reliable evidence-grounded evaluations of behaviors that passive methods can leave unobserved.

Agentic Safety is an Epistemic Property, Not a Behavioral One

Charles L. Wang, Keir Dorchen, Peter Jin

Contemporary AI safety spans pre-training interventions, post-training alignment, deployment-time controls, monitoring, and red-teaming. These methods are necessary, but they primarily certify snapshots of system behavior. As AI systems become more capable, dynamic, embodied, and self-improving, this snapshot view becomes incomplete: safety depends not only on whether a system behaves acceptably now, but whether it remains correctable as it learns, adapts, acts, and modifies itself over time. This paper argues that safety should therefore be treated as an epistemic property of the evolving learner, not merely a behavioral property of the current policy. We introduce teachability as the capacity to preserve future corrective leverage under bounded human, institutional, or environmental intervention. We argue that advanced systems can retain visible competence while eroding the representational, algorithmic, or meta-decision conditions needed for future correction. Safe advanced AI systems must not only behave acceptably now; they must remain teachable later.

High-stakes enterprise document generation, including financial dispute narratives, compliance notices, and audit summaries, demands schema correctness, policy compliance, and low-latency operation at scale. Prior to a unified guardrail layer, production systems often stitched together separate PII redaction, content moderation, and format validation steps, leading to fragmented logic, slower request paths, and higher operational cost. We present a guardrail orchestration layer for text and image inputs that couples multi-candidate generation with an explicit compliance score used for early exit. The framework runs configurable parallel generation heads, scores candidates against weighted guardrails including PII detection, content moderation, schema constraints, and domain rules, and returns the best-scoring output with selection metadata. The available operational readout reports 5 attempts within 20 seconds and 91 percent compliance. For payments dispute defense summaries, we analyze aggregate operational scenario readouts rather than a randomized A/B test. Variable cohorts show higher count win rates than controls overall, 301/659 versus 536/1548, corresponding to +11.0 percentage points with 95 percent confidence interval [6.6, 15.5] and p < 0.001, and for adjusted item-not-received cases, +7.5 percentage points with 95 percent confidence interval [0.2, 15.7] and p = 0.045. Fraud and local evidence-ranking deltas are directionally positive but not statistically significant from the aggregate count data. We also report reviewer-calibrated Responsible-AI evidence-quality signals from 770 generated-evidence reviews and a 70-case OCR slice, and document the reproducibility boundary through the request interface, scoring logic, pseudocode, and operational evidence boundary.