The paper tests whether small wording changes can shift AI models from refusing to complying with a sensitive request: increasing income on a pay stub. Across 15 models and 445 simulated trials, the authors compare three prompt framings—stating a loan purpose, omitting the reason, or describing a printing error—and find that editing-call rates rise sharply from 14.1% to 66.2%. The work matters because it offers a concrete, traceable probe of boundary robustness in financial-document scenarios, though the authors are careful to say it does not establish an ethics ranking or prove real-world harm.

Mental-health chatbots powered by LLMs risk unsafe outputs, yet current safety layers are opaque and developer-driven. This paper proposes a clinician-led guardrail in which two doctoral trainees in clinical psychology author four safety checks, which are then formalized over sixteen typed facts and verified for internal consistency with an SMT solver before deterministic execution. The core claim is not superiority in accuracy but in inspectability: every verdict recomputes from named rules and facts, making disagreements localizable and revisable.

ULME tackles the frictions between text embedding and text generation, which are usually handled by separate models with incompatible objectives. The paper proposes a post-training framework for causal decoder-only LLMs that inserts a chain of learnable slots between the query and answer, then couples the two tasks via a stop-gradient predictive loss so that generation actively shapes the embedding. Across eight scales of Qwen2.5 and Qwen3 (0.5B–8B), a single ULME checkpoint raises MTEB scores by 3.3–4.1 points over the strongest open baseline while remaining on par with generation-only fine-tuning on standard accuracy benchmarks.

The paper asks how to predict the performance of a safety probe before it is trained. Its core finding is that probe error follows two scaling laws: a power-law floor in model size and a Hill-curve approach to that floor as labels increase. This matters because it could let safety teams budget labels and choose architectures for new models without expensive grid search.

Singular-value spectra shape how transformer MLPs amplify different directions, yet their causal role in downstream capabilities is poorly understood. This paper introduces a post-hoc "Matthew" perturbation that reweights the singular values of MLP up_proj and down_proj matrices around their geometric mean—spiking inequality with $\alpha>0$ or smoothing it with $\alpha<0$—while preserving singular vectors and the geometric mean. Across five instruction-tuned Llama and Qwen models, the authors evaluate perturbed systems on 118 capability-labeled subtasks, auxiliary benchmarks, and personality probes, arguing that spectral inequality is a manipulable variable that can selectively trade off factual knowledge, language understanding, and deductive reasoning without retraining.

When a multilingual LLM flips its answer under user pressure, it is unclear whether it knew the correct answer and yielded (sycophancy) or simply lacked competence in that language. This paper introduces the SC-matrix, a $2 \times 2$ diagnostic crossing Competence ($C$) and Sycophancy ($S$) to decompose behavior into four cells: Robust, True Sycophancy, Honest Uncertainty, and Uninformed Deference. Their experiments across six languages and three models show that aggregate sycophancy rates hide very different failure mixtures, and that reasoning-based prompts outperform direct alignment instructions at reducing pressure-induced flips.

This paper investigates whether Group Relative Policy Optimisation (GRPO), previously applied to LLM-based speech recognition, can improve smaller encoder-decoder ASR models in a multilingual, low-resource setting. Using Whisper-large-v3 as a case study, the authors train SunflowerASR on 51 African languages and report state-of-the-art macro-averaged WER on AfriVox-v2 and SimbaBench. Key innovations include a headroom-based training example selection strategy and careful ablations isolating the reinforcement signal from mere additional training.

Steward is a multi-agent LLM trading system designed to act on live Web data while keeping its safety constraints intact as it learns. The paper introduces governed self-evolution: a deterministic gate and typed proposal system let the agent rewrite its own prompts, but an invariant core of risk rules, order arithmetic, and discipline-agent prompts is unreachable by any self-edit. In a five-month live paper-trading deployment the authors record every input, output, and grading event to evaluate whether the architecture survives real market noise, prompt injection, and its own reflection loop. The release of the full code, prompts, and replay harness makes this one of the most transparent live-agent studies to date.

Traditional student models only use responses to standalone questions, ignoring the rich, fine-grained evidence embedded in tutoring dialogues. This paper proposes URF-QD, a framework that rewrites tutor-posed mathematical subtasks as self-contained questions and models them jointly with original items on a shared item response theory scale. The key novelty is a relative-order regularization that uses tutor pedagogical moves to constrain consecutive subtask difficulties, enabling dynamic tracking of student ability across both questions and dialogues. The authors show consistent improvements on two real-world math datasets, though the approach leans heavily on LLM annotations and modest sample sizes.

Speculative decoding drafters trained on short sequences collapse on long-context inputs because their position encodings extrapolate poorly. FOLD fixes this without retraining by keeping exact positions in a local window and compressing all distant context into a shared-position far memory. A concentration factor $\alpha$ in a decoupled softmax prevents attention dilution as the memory grows. Across six target–drafter pairs and contexts up to 256K tokens, the method consistently raises speedup, in the best case more than doubling it.

COUNT tackles the GPU-memory bottleneck of scaling language-model embeddings by replacing learned n-gram tables with precomputed corpus counts stored on cheap host RAM. It constructs input embeddings via a weighted sum of the model's own unembedding vectors, requiring no extra learned parameters and enabling training-free domain adaptation at inference. On 60M--1B-parameter backbones, the authors demonstrate consistent perplexity and downstream gains without increasing GPU memory pressure.

Current evaluations of Knowledge Extraction Enhancement (KEE) techniques collapse recovery and disruption into a single accuracy score, concealing critical trade-offs. MAPLE introduces a unified diagnostic framework built on a pretraining-verified benchmark and three complementary metrics—surface-level $\text{Gain}/\text{Loss}$, probability-level $\text{PRS}$, and beyond-sampling $\text{SGR}/\text{DGR}$—to dissect intervention effects at multiple levels. Across six techniques and three model families, MAPLE shows that prompting recovers deep failures but disrupts many successes, while decoding and steering often reshuffle token probabilities without consistent benefit. The work offers a principled, actionable standard for comparing KEE methods beyond headline accuracy.

MMLA-RSI proposes recursive self-improvement through two Memory-Mediated Learning Architecture routes: Reasoning-Time Training (RTT) treats the numerical policy-state updater as an object of improvement, while Reasoning-Time Memory Adaptation (RTMA) treats the memory-management rule as an object of improvement. Each admitted successor takes over its parent's role in producing the next generation's rule. The paper's main contribution is a formally bounded construction with shared interfaces and a common algorithm, though the authors explicitly state that beneficial recursive chains remain to be evaluated.

This paper confronts the evaluation crisis in agentic self-distillation, where a flood of recent methods claim gains over GRPO using inconsistent protocols and reused baselines. The authors conduct a unified benchmark of ten representative approaches across four tasks, three model sizes, and 132 training runs totaling over 12,000 H800 GPU-hours, measuring not only final accuracy but also training stability and wall-clock cost. Their central finding is sobering: no method consistently beats GRPO, many suffer severe hidden instability, and extra compute rarely translates into better agents.

Supervised fine-tuning can collapse the pretrained model's preference structure among non-target tokens. Instead of matching the reference distribution exactly like LP-SFT, this paper proposes SRP-SFT, which enforces soft, selective lower bounds on margins between top-K non-target token pairs. At one epoch on a 4B model, it improves MATH-500 pass@1 from 49.56 to 54.56 over LP-SFT, with consistent gains across three seeds.

GRPO for long-horizon agents suffers from sparse terminal rewards that yield all-fail rollout groups and zero advantage signals early in training. This paper studies guide-assisted curricula that replay successful reference prefixes before the learner continues, systematically varying the initial position, handoff granularity, and movement rule. Across Qwen3 models from 1.7B to 8B on BFCL and AppWorld, the authors find that starting near task completion and using outcome-informed call-level selection improves independent execution while reducing training cost. The work usefully frames curriculum design as a practical lever for GRPO-based agent post-training.

RM-OPD targets a real bottleneck in generalist reward modeling: training on mixed-domain data with standard RL produces cross-domain interference and sparse outcome rewards, so capabilities like math verification, code review, and safety guardrails cannot coexist without degrading each other. The core idea is a two-stage framework—first cultivate domain-specialist reward teachers via single-domain GRPO, then distill them into one student through routed on-policy distillation, domain contribution reweighting, and criteria context distillation. This matters because it swaps coarse verdict-level supervision for dense token-level guidance, raising the multi-domain performance ceiling for generative judges.

Data-quality gates conflate within-region example selection with across-region mass redistribution, so a gate-versus-random comparison cannot isolate what improved downstream performance. This paper introduces Counterfactual Gate–Support Decomposition (CGSD), a four-arm, budget-matched design that decomposes the gate effect into within-support selection $E$, between-support redistribution $S$, and their interaction $I$ by crossing eligibility and support-share factors. The core contribution is not the factorial algebra itself, but operationalizing it for finite pools while characterizing when the required arms are unrealizable.

INTACT attacks the inference cost of chain-of-thought reasoning by replacing early textual reasoning steps with compact continuous latent vectors while keeping the backbone LLM frozen. A small 107M-parameter encoder iteratively builds a latent prefix from the frozen model's own hidden states, first distilling the effect of written reasoning steps and then refining via group-relative policy gradients. Across Llama and Qwen backbones from 1B to 8B, the method improves accuracy and shortens generation length, and a single encoder trained on math, science, and code transfers to twelve held-out benchmarks including commonsense without adaptation.

Long-context LLM inference is bottlenecked by expanding key-value caches. This paper proposes treating the KV cache as a learnable layer and optimizing it per document via test-time training while keeping model weights frozen. The two-stage procedure first matches teacher attention statistics and then distills teacher predictions, yielding large accuracy gains on synthetic and natural-document benchmarks under aggressive compression.