Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2604.16656v2 Announce Type: replace Abstract: All languages are equal; when it comes to tokenization, some are more equal than others. Tokens are the hidden currency that dictates the cost and...
The paper introduces SYNC, a training‑free inference controller that dynamically selects the optimal layer for injecting persona steering vectors during decoding. SYNC estimates each layer’s sensitivity to every persona using cosine similarity trajectories, allowing the model to adaptively choose the best injection point at each step. Experiments on two models and multiple persona profiles show that SYNC outperforms static multi‑persona steering methods in 21 of 28 configurations, with significant gains in 15 cases.
The paper introduces a KL regularization term that aligns predicted dialogue act distributions with empirical transition patterns derived from a corpus. Applied to a 60‑class German counselling taxonomy, this method improves macro‑F1 scores by 9–42% relative to various encoders and enhances dialogue‑flow alignment. Cross‑dataset validation on the HOPE dataset shows that the gains transfer across languages and counselling domains, and systematic ablations reveal that transition regularization consistently benefits weaker baseline models, especially in fine‑grained, data‑sparse dialogue tasks.
EffComp is a new framework for prompt compression in retrieval‑augmented question answering. It uses BERT‑style document rerankers and sentence selectors to perform fast, extractive compression at the sentence level, thereby avoiding hallucinations. The hybrid reinforcement‑plus‑supervised training optimizes compression ratio while maintaining LLM accuracy, and experiments show EffComp outperforms existing methods with low latency.
The paper introduces a continual learning approach for large language models called Multi‑Teacher On‑Policy Distillation (MOPD) that avoids storing intermediate expert snapshots by replaying prompts from earlier domains using the previous student as a teacher. Two replay strategies are examined: a shared replay budget across domains and a fixed budget per domain, with replay steps that either distill token distributions alone or also align hidden states, attention scores, and value relations. Experiments on a five‑domain benchmark (logic, math, code, safety, instruction following) demonstrate that replay alone closes most of the performance gap between sequential distillation and joint MOPD, achieving comparable results to recovery from all stored snapshots when using a fixed replay share per domain.
The paper introduces CLEO (Constrained Loop for Environment Optimization), a technique for tailoring a minimal agent harness around a self-hostable large language model (LLM) to meet the needs of government intelligence teams. CLEO selectively incorporates tools, guardrails, memory policies, and loop components that enhance accuracy while keeping the system general and auditable. On a held‑out dataset, the CLEO harness achieves an 83.8% pass rate—just 1.2 points below the top out‑of‑the‑box Claude Code SDK baseline—yet is two orders of magnitude smaller.
The paper introduces latent debate, a structured surrogate framework that interprets the internal thinking of Large Language Models (LLMs) on True/False tasks by modeling latent arguments and their interactions. It provides a model- and task-agnostic conceptual framework, instantiated symbolically to approximate LLM decision-making. Empirical results show that latent debate closely matches LLM predictions, offers interpretability, and serves as a strong baseline for hallucination detection, revealing correlations between debate patterns and hallucination risk.
arXiv:2611.03092v1 Announce Type: cross Abstract: Personal AI agents are increasingly capable of handling long-horizon tasks across applications, yet users often need to invest substantial effort in...
The paper introduces a grammar‑lens benchmark that generates data at progressively higher computational levels to test language models. Using this benchmark, GPT‑5.5 and its Pro variants were evaluated, showing increasing acceptance rates of 16%, 40%, 43%, and 52% as the models’ test‑time reasoning levels rose. The study also notes that higher compute effort leads to responses that diverge more from human answers, hinting at a potential capability‑versus‑alignment tension.
TopoSteer introduces a topology-based method for selecting the layer at which to apply activation steering in large language models. Instead of evaluating separability or alignment, it scores layers by the total H_0 persistence of the displacement cloud, measured as the length of its Euclidean minimum spanning tree. Experiments across several models and tasks show that this approach yields more effective steering without needing held-out validation.
The paper proposes a hybrid narrative framework that blends pre‑authored content with real‑time LLM generation, using semantic similarity to balance authorial control and player agency. It also introduces a self‑evolving, feedback‑driven optimization system that refines prompts through in‑context learning, removing the need for manual prompt engineering. A case study and participant evaluation demonstrate that this approach improves narrative coherence while maintaining player agency and aligning with authorial intent.
GenDB‑V is a system that uses AI agents to automatically synthesize database engines tailored to specific hardware and workloads, while also providing formal verification of correctness. It introduces SQLSpec to formalize SQL semantics, enabling joint synthesis of efficient code and modular proofs that bridge low‑level execution to SQL semantics. Evaluations on TPC‑H and SEC‑EDGAR show significant speedups over existing engines and reveal correctness issues in alternative LLM‑based synthesis methods.
The paper introduces Library of Layers (LoL), an architecture that dynamically orders transformer layers during inference, allowing a single model to adapt its computation to the task at hand. LoL’s single hyperparameter—the size of its global library of transformer blocks—enables a smooth transition between parameter reuse and sparsity, unifying dense, depth‑recurrent, and mixture‑of‑experts trade‑offs. Experiments show LoL matches or exceeds these baselines on text, synthetic Sudoku and maze tasks, and demonstrates task‑specific routing in multilingual and safety‑focused settings.
The paper introduces Anthropogenic Regional Adaptation, a new framework for aligning vision‑language models with specific regional contexts while preserving global performance. It proposes the GG‑EZ method, which filters regional data and merges models to achieve this alignment. Experiments on three VL architectures and a Southeast Asia case study show 5‑15% improvements in cultural relevance metrics with over 98% retention of global performance.
Circuit Oracle is an off‑the‑shelf large language model that, using a multi‑agent harness and task‑specific skills, can answer natural‑language questions about a target model by reading its attribution graph via tool calls. The system was evaluated on three safety‑relevant proxy tasks—detecting spurious features in probe circuits, eliciting hidden knowledge from taboo‑finetuned models, and suppression jailbreaking via feature‑level interventions—achieving 78.2 % accuracy on spurious‑feature detection and mixed results on the other tasks. In addition to task scores, Circuit Oracle produces pruned subgraphs and context‑specific feature explanations, which the authors release for further study.
S1CogBench is a new benchmark that evaluates System One decision models—typed decision models like Jev and Laya—on 24 cognitive tasks. It tests whether these models can use evidence, respond to feedback, handle memory load, and anticipate future opportunities, while keeping conditions and scoring separate from any disclosed adaptations. The study finds wide variability in reliability among decision models, shows that good performance does not guarantee human-like predictions, and reveals that speed largely stems from generating decisions without text output.
Depth pruning removes entire Transformer blocks to shrink and speed up language models, but the resulting damage is uneven across capabilities. In a study of four 1.2–1.7 B instruction‑tuned models, function calling and mathematical reasoning collapsed in every case, while other skills varied in order of degradation. Fine‑tuning a short ‘healing’ run can restore some abilities, with the choice of which blocks to retrain and the allocation of a limited token budget influencing recovery outcomes.
The paper investigates gated knowledge distillation (KD) for low‑resource Bangla generation, testing two gating strategies—EWAD and CHAD—to selectively trust teacher signals. Across six Bangla tasks and multiple student models, neither gate improves over plain KD, and KD generally underperforms cross‑entropy fine‑tuning on gold outputs. The primary advantage of KD is regularisation, with teacher‑free label smoothing narrowing the performance gap, though gating only modestly alters this effect.
The article presents a detailed, step‑by‑step derivation of the backpropagation algorithm for a fully connected feedforward neural network with sigmoid activations and squared‑error loss. It starts with a single‑weight example, extends to general error terms and matrix equations, and then scales the derivation to multiple training examples and mini‑batch gradient descent. Each equation is annotated with dimensions and linked to a NumPy implementation trained on handwritten Arabic digits, with full code provided.
The paper investigates how large language model (LLM) agents can collude in multi‑agent systems and introduces the J‑space as a tool to detect and steer such collusion. Through interpretability experiments on DeepSeek‑V4‑Flash and a steganographic card‑counting task, the authors show that the J‑space captures collusion behaviors, such as misrepresentation and hidden information. They also propose J‑PACT, an unsupervised method that successfully identifies colluding pairs across several open‑weight models, achieving high AUROC and accurate pair identification.