Abstract

Experience from repeated analogous situations can bias subsequent decision-making, yet it remains unclear whether such effects arise from prior experience itself or the valence of that experience.
We investigated whether prior gain or loss modulates
subsequent uncertain decisions compared to a no-experience control. Participants completed a manipulated Iowa Gambling Task (M-IGT), inducing predetermined gain or loss, followed by the standard IGT. Prior experience group showed altered exploratory behavior in the early phase, with the loss group exhibiting early risk aversion. Behavioral differences across control and prior experience groups converged later in the task. The 𝑃𝑉𝐿 − Δ model revealed that prior experience initially reduced outcome sensitivity. In the later blocks, the gain group showed reduced learning rate and higher outcome sensitivity, indicating strategy stabilization. Prior loss was associated with higher
choice consistency across all blocks. These findings suggest that prior experience transiently biases early exploration, while valence shapes decision consistency and learning flexibility.

Abstract

Semantic banks, WordNets, and other sense inventories play a crucial role in the training
and evaluation of many NLP applications, including machine translation and semantic
understanding tasks. However, for low-resource languages such as Sanskrit, comprehensive and well-structured sense inventories are often unavailable. This not only poses a
significant challenge for traditional NLP pipelines but also for leveraging state-of-the-art
large language models (LLMs), which still require a foundational, machine usable sense
inventory to support tasks involving meaning disambiguation and semantic consistency.
Our work is motivated by this gap. We focus on creating a machine usable sense inventory
for Sanskrit by using the Monier-Williams dictionary as a primary lexical resource. Our
goal in building such an inventory, is to minimize ambiguity to the maximum possible
extent so that the resulting senses are clear and suitable for computational models. To
achieve this, we make use of clustering algorithms to group semantically similar senses
into synsets and to reorganize the existing dictionary structure in a way that it better
reflects semantic relationships such as polysemy and synonymy.

Abstract

How should we evaluate generation systems that combine autoregressive (AR) and diffusion decoding? Westudythis question through Speculative Refinement (SpecRef), a training-free hybrid method that warm-starts a masked diffusion language model from an AR draft using entropy-guided selective masking. Evaluating SpecRef across six benchmarks (HumanEval, MBPP, GSM8K, BBH, ARC-Challenge, HellaSwag) with three distinct evaluation protocols (execution-based pass@1, exact-match, loglikelihood scoring), we surface several findings relevant beyond our specific system: (1) code benchmarks conflate structural discovery with logical correctness: providing a syntactic scaffold lifts accuracy from near zero to over 20% without changing the model, indicating that much of the baseline failure is structural; (2) a refinement tension phenomenon where multistage correction degrades already-correct tokens, exposing benchmark saturation ceilings invisible to single-model evaluation; (3) loglikelihood and generative evaluation produce different model rankings for the same model pair, suggesting they measure different capabilities; (4) standard Python post-processing silently breaks code evaluation for non-AR generators. These observations apply to any multistage or non-autoregressive generation pipeline and point toward more diagnostic evaluation practices. Codes are here.

Abstract

Verbal humor involves reasoning through complex conversational contexts. Although LLMs have achieved strong performance on English humor datasets, their ability to interpret humor in Hindi remains unexplored. In this paper, we evaluate Hindi humor for which we extract dialogues from humorous video clips. We use a pipeline that transforms video content into detailed textual streams, including dialogue transcripts and scene descriptions, allowing reasoning over inputs exceeding 2,000 words. We test various LLMs, from efficient edge models (Qwen-2.5-7B, Qwen-3-7B, Gemma-3-27B) to Indic-focused models (Sarvam-M-24B) and large frontier models (Llama-3.1-70B, Gemini-2.0-Flash). Our findings show a concave performance pattern in long-context understanding, with reasoning quality peaking at moderate lengths (250-750 words) and declining at higher context lengths. We also show that standard metrics overstate pragmatic competence. While increasing model size generally improves performance, we also observe distinct failures in smaller LLMs due to instructional and linguistic issues, necessitating diversity metrics to capture hallucinations. Smaller, Hindi-focused models can compete with much larger generalist models. Importantly, our evaluation reveals that conversational humor is a challenge for even specialized models, making HinS a valuable benchmark for advancing research in Hindi Long-Context Humor Reasoning.

Abstract

Text-to-speech systems are improving fast, but measuring how natural they sound still requires expensive human listening tests; existing automatic methods struggle to generalise well across different datasets, so we present QMOS, a MOS prediction framework that extracts hierarchical speech quality features from a frozen Qwen2-Audio large audio language model and combines them with layer-weighted WavLM SSL representations through a learned cross-attention fusion, capturing both high-level semantic naturalness and low-level acoustic distortions in a unified model, and on two standard benchmarks, SOMOS and BVCC, QMOS achieves competitive system-level SRCC of 0.923 and 0.921 respectively while using no system-ID conditioning or listener embeddings, with cross-domain evaluation yielding a competitive SRCC of 0.702, suggesting the learned representations generalise across acoustic domains.

Abstract

Fine-tuning Large Language Models (LLMs) on
benign narrow data can sometimes induce broad
harmful behaviors, a vulnerability termed emer-
gent misalignment (EM).
While prior work links these failures to specific
directions in the activation space, their relation-
ship to the model's broader persona remains un-
explored. We map the latent personality space
of LLMs through established psychometric pro-
files like the Big Five, Dark Triad, and LLM-
specific behaviors (e.g. evil, sycophancy), and
show that the semantic geometry is highly sta-
ble across aligned models and their corrupted
fine-tunes. Through causal interventions, we find
that directions isolating social valence, such as
the 'Evil' persona vector, and a Semantic Va-
lence Vector (SVV) that we introduce, function
as intrinsic guardrails: ablating them drives the
misalignment rates above 40%, while amplifying
them suppresses the failure mode to less than 3%.
Leveraging the structural stability of the personal-
ity space, we show that vectors extracted a priori
from an instruct-tuned model transfer zero-shot to
successfully regulate EM in corrupted fine-tunes.
Overall, our findings suggest that harmful fine-
tuning does not overwrite a model's internal rep-
resentation of personality, allowing conserved rep-
resentations to serve as robust, cross-distribution
guardrails.

Abstract

Ensuring the freshness of drinking water at the point of use remains a challenge in building-scale storage systems, where intermittent demand can lead to prolonged storage. This paper presents a data-driven framework for estimating and managing water age using IoT-enabled flow monitoring, eliminating the need for tracer studies or detailed hydraulic models. An eight-month dataset collected from a cumulative flow meter installed on a system of three 80 L drinking water coolers is used to reconstruct continuous consumption behavior from discrete refilling events through interpolation and event aggregation. The reconstructed demand profile enables continuous estimation of water age, revealing persistent staleness under baseline operation, with water age exceeding 14 days for a significant portion of the monitoring period and reaching peaks of approximately 30 days. To address this, two operational strategies are investigated: demand-aware refilling based on observed hourly consumption patterns, and reduction of storage capacity to promote faster turnover. Both approaches demonstrate substantial reductions in water age; specifically, demand-aware refilling reduces the mean water age from 13.6 days to 5.3 days and the peak from approximately 30 days to 17 days, while capacity reduction to 120 L reduces the mean to 6.9 days. The results highlight the potential of IoT-based monitoring combined with simple control strategies to enhance water freshness in decentralized storage systems, offering a scalable and low-cost solution for smart water management.

Abstract

We present CadLLM, a training-free method
to accelerate the inference throughput of
diffusion-based LLMs (dLLMs). We first in
vestigate on the dynamic nature of the token
unmasking confidence across blocks and steps.
Based on this observation, we then present a
lightweight adaptive approach that can control
the generation block size, step size, and thresh
old based on the average confidence score of
the unmasked tokens. We further reduce the
softmaxing overhead of token probability gen
eration by dynamically leveraging a subset of
vocabulary size to regulate sampling breadth.
CadLLM is a plug-and-play model-agnostic
with KV caching based dLLMs. Extensive ex
periments on four popular tasks demonstrate
the efficacy of CadLLM to yield throughput
improvements ranging from 1.1-2.28× over
the state-of-the-art baseline with competitive
accuracy.

Abstract

The growing scale, complexity, interconnectivity, and autonomy of
modern software ecosystems introduce unprecedented levels of un
certainty, challenging the foundations of traditional self-adaptation.
Existing self-adaptive techniques, often based on rule-driven con
trollers or isolated learning components, struggle to generalize
across novel contexts and coordinate responses across distributed
subsystems, leaving them ill-equipped to handle emergent "un
known unknowns" in complex, dynamic environments. Recent dis
cussions on Self Adaptation 2.0 established an equal partnership
between AI and adaptive systems, merging learning-driven intelli
gence with adaptive control to enable predictive, data-driven, and
proactive adaptation. Building on this foundation, we introduce a
new wave of self-adaptation through POLARIS a three-layer multi
agentic self-adaptation framework that moves beyond reactiveadap
tation by combining (1) a low-latency Adapter layer for monitoring
and safe execution, (2) a transparent Reasoning layer that gener
ates and verifies plans using tool-aware, explainable agents, and (3)
a Meta layer that records experiences and meta-learns improved
adaptation policies over time. By using shared knowledge and pre
dictive models, POLARIS can handle uncertainty, learn from its
past actions, and evolve its strategies, enabling the engineering
of autonomous systems that can anticipate change to ensure re
silient and goal-directed behavior under uncertainty. Preliminary
evaluation across two distinct self-adaptive exemplars, SWIM and
SWITCH, demonstrates that POLARIS consistently outperforms
existing state-of-the-art baselines. With this, we motivate a shift
towards Self-Adaptation 3.0 akin to Software 3.0, a new paradigm
where systems move beyond merely learning from their environ
ment to reasoning about and evolving their own adaptation. In this
vision, adaptation contributes to a self-learning process that enables
systems to continuously improve and respond to novel challenges.

Abstract

A fundamental challenge in reasoning is navigating hypothetical, counterfactual worlds where logic may conflict with ingrained knowledge. We investigate this frontier for Large Language Models (LLMs) by asking: Can LLMs reason logically when the context contradicts their parametric knowledge? To facilitate a systematic analysis, we first introduce CounterLogic, a benchmark specifically designed to disentangle logical validity from knowledge alignment. Evaluation of 11 LLMs across six diverse reasoning datasets reveals a consistent failure: model accuracy plummets by an average of 14% in counterfactual scenarios compared to knowledge-aligned ones. We hypothesize that this gap stems not from a flaw in logical processing, but from an inability to manage the cognitive conflict between context and knowledge. Inspired by human metacognition, we propose a simple yet powerful intervention: Flag & Reason (FaR), where models are first prompted to flag potential knowledge conflicts before they reason. This metacognitive step is highly effective, narrowing the performance gap to just 7% and increasing overall accuracy by 4%. Our findings diagnose and study a critical limitation in modern LLMs' reasoning and demonstrate how metacognitive awareness can make them more robust and reliable thinkers.