Subramanyam Sahoo

Mubashara Akhtar, Anka Reuel, Prajna Soni et al. · meta-ai

Artificial Intelligence (AI) benchmarks play a central role in measuring progress in model development and guiding deployment decisions. However, many benchmarks quickly become saturated, meaning that they can no longer differentiate between the best-performing models, diminishing their long-term value. In this study, we analyze benchmark saturation across 60 Large Language Model (LLM) benchmarks selected from technical reports by major model developers. To identify factors driving saturation, we characterize benchmarks along 14 properties spanning task design, data construction, and evaluation format. We test five hypotheses examining how each property contributes to saturation rates. Our analysis reveals that nearly half of the benchmarks exhibit saturation, with rates increasing as benchmarks age. Notably, hiding test data (i.e., public vs. private) shows no protective effect, while expert-curated benchmarks resist saturation better than crowdsourced ones. Our findings highlight which design choices extend benchmark longevity and inform strategies for more durable evaluation.

Subramanyam Sahoo, Vinija Jain, Aman Chadha et al.

Linear probing of large language model (LLM) hidden states is widely used to claim that models learn distinct representations for different reasoning types. We test this by probing Qwen3-14B on three benchmarks spanning the classical trichotomy: LogiQA 2.0 (deductive), ARC-Challenge (inductive), and $α$NLI (abductive). At layer 32 of 40, linear probes achieve 100\% cross-validated accuracy with well-separated geometry (intrinsic dimensionalities: 20.6, 28.5, 33.6; convex hull contamination $\leq$1.5\%). However, this separation is entirely driven by format confounds. Residualizing source identity, option count, and response length reduces accuracy to chance. Trace-anchor similarity indicates largely shared reasoning across tasks (42.5\% agreement vs.\ 33.3\% chance), and causal steering with random controls ($n=20$) shows no functional link between geometry and reasoning mode ($p=0.286$). Thus, high probe accuracy reflects task format rather than computational structure, motivating routine format deconfounding in mechanistic interpretability.

Anka Reuel, Avijit Ghosh, Jenny Chim et al.

Foundation models are increasingly central to high-stakes AI systems, and governance frameworks now depend on evaluations to assess their risks and capabilities. Although general capability evaluations are widespread, social impact assessments covering bias, fairness, privacy, environmental costs, and labor practices remain uneven across the AI ecosystem. To characterize this landscape, we conduct the first comprehensive analysis of both first-party and third-party social impact evaluation reporting across a wide range of model developers. Our study examines 186 first-party release reports and 183 post-release evaluation sources, and complements this quantitative analysis with interviews of model developers. We find a clear division of evaluation labor: first-party reporting is sparse, often superficial, and has declined over time in key areas such as environmental impact and bias, while third-party evaluators including academic researchers, nonprofits, and independent organizations provide broader and more rigorous coverage of bias, harmful content, and performance disparities. However, this complementarity has limits. Only model developers can authoritatively report on data provenance, content moderation labor, financial costs, and training infrastructure, yet interviews reveal that these disclosures are often deprioritized unless tied to product adoption or regulatory compliance. Our findings indicate that current evaluation practices leave major gaps in assessing AI's societal impacts, highlighting the urgent need for policies that promote developer transparency, strengthen independent evaluation ecosystems, and create shared infrastructure to aggregate and compare third-party evaluations in a consistent and accessible way.

Subramanyam Sahoo, Vinija Jain, Divya Chaudhary et al.

Instruction tuned reasoning models are increasingly deployed with safety classifiers trained on frozen embeddings, assuming representation stability across model updates. We systematically investigate this assumption and find it fails: normalized perturbations of magnitude $σ=0.02$ (corresponding to $\approx 1^\circ$ angular drift on the embedding sphere) reduce classifier performance from $85\%$ to $50\%$ ROC-AUC. Critically, mean confidence only drops $14\%$, producing dangerous silent failures where $72\%$ of misclassifications occur with high confidence, defeating standard monitoring. We further show that instruction-tuned models exhibit 20$\%$ worse class separability than base models, making aligned systems paradoxically harder to safeguard. Our findings expose a fundamental fragility in production AI safety architectures and challenge the assumption that safety mechanisms transfer across model versions.

Subramanyam Sahoo, Aman Chadha, Vinija Jain et al.

Situational awareness, the capacity of an AI system to recognize its own nature, understand its training and deployment context, and reason strategically about its circumstances, is widely considered among the most dangerous emergent capabilities in advanced AI systems. Separately, a growing research effort seeks to improve the logical reasoning capabilities of large language models (LLMs) across deduction, induction, and abduction. In this paper, we argue that these two research trajectories are on a collision course. We introduce the RAISE framework (Reasoning Advancing Into Self Examination), which identifies three mechanistic pathways through which improvements in logical reasoning enable progressively deeper levels of situational awareness: deductive self inference, inductive context recognition, and abductive self modeling. We formalize each pathway, construct an escalation ladder from basic self recognition to strategic deception, and demonstrate that every major research topic in LLM logical reasoning maps directly onto a specific amplifier of situational awareness. We further analyze why current safety measures are insufficient to prevent this escalation. We conclude by proposing concrete safeguards, including a "Mirror Test" benchmark and a Reasoning Safety Parity Principle, and pose an uncomfortable but necessary question to the logical reasoning community about its responsibility in this trajectory.

Subramanyam Sahoo, Aman Chadha, Vinija Jain et al.

Mathematical reasoning models are widely deployed in education, automated tutoring, and decision support systems despite exhibiting fundamental computational instabilities. We demonstrate that state-of-the-art models (Qwen2.5-Math-7B) achieve 61% accuracy through a mixture of reliable and unreliable reasoning pathways: 18.4% of correct predictions employ stable, faithful reasoning while 81.6% emerge through computationally inconsistent pathways. Additionally, 8.8% of all predictions are silent failures -- confident yet incorrect outputs. Through comprehensive analysis using novel faithfulness metrics, we reveal: (1) reasoning quality shows weak negative correlation with correctness (r=-0.21, p=0.002), reflecting a binary classification threshold artifact rather than a monotonic inverse relationship; (2) scaling from 1.5B to 7B parameters (4.7x increase) provides zero accuracy benefit on our evaluated subset (6% of GSM8K), requiring validation on the complete benchmark; and (3) latent reasoning employs diverse computational strategies, with ~20% sharing CoT-like patterns. These findings highlight that benchmark accuracy can mask computational unreliability, demanding evaluation reforms measuring stability beyond single-sample metrics.

Subramanyam Sahoo

Modern large language models (LLMs) are increasingly fine-tuned via reinforcement learning from human feedback (RLHF) or related reward optimisation schemes. While such procedures improve perceived helpfulness, we investigate whether sycophantic reward signals degrade calibration -- a property essential for reliable uncertainty quantification. We fine-tune Qwen3-8B under three regimes: no fine-tuning (base), neutral supervised fine-tuning (SFT) on TriviaQA, and sycophancy-inducing Group Relative Policy Optimisation (GRPO) that rewards agreement with planted wrong answers. Evaluating on $1{,}000$ MMLU items across five subject domains with bootstrap confidence intervals and permutation testing, we find that \textbf{sycophantic GRPO produces consistent directional calibration degradation} -- ECE rises by $+0.006$ relative to the base model and MCE increases by $+0.010$ relative to neutral SFT -- though the effect does not reach statistical significance ($p = 0.41$) at this training budget. Post-hoc matrix scaling applied to all three models reduces ECE by $40$--$64\%$ and improves accuracy by $1.5$--$3.0$ percentage points. However, the sycophantic model retains the highest post-scaling ECE relative to the neutral SFT control ($0.042$ vs.\ $0.037$), suggesting that reward-induced miscalibration leaves a structured residual even after affine correction. These findings establish a methodology for evaluating the calibration impact of reward hacking and motivate calibration-aware training objectives.

Aritra Roy, Kevin Shen, Andrew MacBride et al.

Large language models (LLMs) are rapidly changing how researchers in materials science and chemistry discover, organize, and act on scientific knowledge. This paper analyzes a broad set of community-developed LLM applications in an effort to identify emerging patterns in how these systems can be used across the scientific research lifecycle. We organize the projects into two complementary categories: Knowledge Infrastructure, systems that structure, retrieve, synthesize, and validate scientific information; and Action Systems, systems that execute, coordinate, or automate scientific work across computational and experimental environments. The submissions reveal a shift from single-purpose LLM tools toward integrated, multi-agent workflows that combine retrieval, reasoning, tool use, and domain-specific validation. Prominent themes include retrieval-augmented generation as grounding infrastructure, persistent structured knowledge representations, multimodal and multilingual scientific inputs, and early progress toward laboratory-integrated closed-loop systems. Together, these results suggest that LLMs are evolving from general-purpose assistants into composable infrastructure for scientific reasoning and action. This work provides a community snapshot of that transition and a practical taxonomy for understanding emerging LLM-enabled workflows in materials science and chemistry.

Subramanyam Sahoo

Agentic AI systems - capable of goal interpretation, world modeling, planning, tool use, long-horizon operation, and autonomous coordination - introduce distinct control failures not addressed by existing safety frameworks. We identify six agentic governance failures tied to these capabilities and show how they erode meaningful human control in military settings. We propose the Agentic Military AI Governance Framework (AMAGF), a measurable architecture structured around three pillars: Preventive Governance (reducing failure likelihood), Detective Governance (real-time detection of control degradation), and Corrective Governance (restoring or safely degrading operations). Its core mechanism, the Control Quality Score (CQS), is a composite real-time metric quantifying human control and enabling graduated responses as control weakens. For each failure type, we define concrete mechanisms, assign responsibilities across five institutional actors, and formalize evaluation metrics. A worked operational scenario illustrates implementation, and we situate the framework within established agent safety literature. We argue that governance must move from a binary conception of control to a continuous model in which control quality is actively measured and managed throughout the operational lifecycle.

Subramanyam Sahoo, Kamlesh Dutta

The relentless pursuit of technological advancements has ushered in a new era where artificial intelligence (AI) is not only a powerful tool but also a critical economic driver. At the forefront of this transformation is Generative AI, which is catalyzing a paradigm shift across industries. Deep generative models, an integration of generative and deep learning techniques, excel in creating new data beyond analyzing existing ones, revolutionizing sectors from production and manufacturing to finance. By automating design, optimization, and innovation cycles, Generative AI is reshaping core industrial processes. In the financial sector, it is transforming risk assessment, trading strategies, and forecasting, demonstrating its profound impact. This paper explores the sweeping changes driven by deep learning models like Large Language Models (LLMs), highlighting their potential to foster innovative business models, disruptive technologies, and novel economic landscapes. As we stand at the threshold of an AI-driven economic era, Generative AI is emerging as a pivotal force, driving innovation, disruption, and economic evolution on a global scale.

Subramanyam Sahoo, Aman Chadha, Vinija Jain et al.

Recursive self-improvement is moving from theory to practice: modern systems can critique, revise, and evaluate their own outputs, yet iterative self-modification risks subtle alignment drift. We introduce SAHOO, a practical framework to monitor and control drift through three safeguards: (i) the Goal Drift Index (GDI), a learned multi-signal detector combining semantic, lexical, structural, and distributional measures; (ii) constraint preservation checks that enforce safety-critical invariants such as syntactic correctness and non-hallucination; and (iii) regression-risk quantification to flag improvement cycles that undo prior gains. Across 189 tasks in code generation, mathematical reasoning, and truthfulness, SAHOO produces substantial quality gains, including 18.3 percent improvement in code tasks and 16.8 percent in reasoning, while preserving constraints in two domains and maintaining low violations in truthfulness. Thresholds are calibrated on a small validation set of 18 tasks across three cycles. We further map the capability-alignment frontier, showing efficient early improvement cycles but rising alignment costs later and exposing domain-specific tensions such as fluency versus factuality. SAHOO therefore makes alignment preservation during recursive self-improvement measurable, deployable, and systematically validated at scale.

Subramanyam Sahoo

Large language models (LLMs) increasingly generate code with minimal human oversight, raising critical concerns about backdoor injection and malicious behavior. We present Cross-Trace Verification Protocol (CTVP), a novel AI control framework that verifies untrusted code-generating models through semantic orbit analysis. Rather than directly executing potentially malicious code, CTVP leverages the model's own predictions of execution traces across semantically equivalent program transformations. By analyzing consistency patterns in these predicted traces, we detect behavioral anomalies indicative of backdoors. Our approach introduces the Adversarial Robustness Quotient (ARQ), which quantifies the computational cost of verification relative to baseline generation, demonstrating exponential growth with orbit size. Theoretical analysis establishes information-theoretic bounds showing non-gamifiability - adversaries cannot improve through training due to fundamental space complexity constraints. This work demonstrates that semantic orbit analysis provides a theoretically grounded approach to AI control for code generation tasks, though practical deployment requires addressing the high false positive rates observed in initial evaluations.

Subramanyam Sahoo, Jared Junkin

Deepfake detection models have achieved high accuracy in identifying synthetic media, but their decision processes remain largely opaque. In this paper we present a mechanistic interpretability framework for deepfake detection applied to a vision-language model. Our approach combines a sparse autoencoder (SAE) analysis of internal network representations with a novel forensic manifold analysis that probes how the model's features respond to controlled forensic artifact manipulations. We demonstrate that only a small fraction of latent features are actively used in each layer, and that the geometric properties of the model's feature manifold, including intrinsic dimensionality, curvature, and feature selectivity, vary systematically with different types of deepfake artifacts. These insights provide a first step toward opening the "black box" of deepfake detectors, allowing us to identify which learned features correspond to specific forensic artifacts and to guide the development of more interpretable and robust models.

Subramanyam Sahoo

Reward design is central to reinforcement learning from human feedback (RLHF) and alignment research. In this work, we propose a unified framework to study hard, continuous, and hybrid reward structures for fine-tuning large language models (LLMs) on mathematical reasoning tasks. Using Qwen3-4B with LoRA fine-tuning on the GSM8K dataset, we formalize and empirically evaluate reward formulations that incorporate correctness, perplexity, reasoning quality, and consistency. We introduce an adaptive hybrid reward scheduler that transitions between discrete and continuous signals, balancing exploration and stability. Our results show that hybrid reward structures improve convergence speed and training stability over purely hard or continuous approaches, offering insights for alignment via adaptive reward modeling.

Subramanyam Sahoo, Aman Chadha, Vinija Jain et al.

Reinforcement Learning from Human Feedback (RLHF) is widely used for aligning large language models, yet practitioners face a persistent puzzle: improving safety often reduces fairness, scaling to diverse populations becomes computationally intractable, and making systems robust often amplifies majority biases. We formalize this tension as the Alignment Trilemma: no RLHF system can simultaneously achieve (i) epsilon-representativeness across diverse human values, (ii) polynomial tractability in sample and compute complexity, and (iii) delta-robustness against adversarial perturbations and distribution shift. Through a complexity-theoretic analysis integrating statistical learning theory and robust optimization, we prove that achieving both representativeness (epsilon <= 0.01) and robustness (delta <= 0.001) for global-scale populations requires Omega(2^{d_context}) operations, which is super-polynomial in the context dimensionality. We show that current RLHF implementations resolve this trilemma by sacrificing representativeness: they collect only 10^3--10^4 samples from homogeneous annotator pools while 10^7--10^8 samples are needed for true global representation. Our framework provides a unified explanation for documented RLHF pathologies including preference collapse, sycophancy, and systematic bias amplification. We conclude with concrete directions for navigating these fundamental trade-offs through strategic relaxations of alignment requirements.

Subramanyam Sahoo, Jared Junkin

Embodied AI agents exploit reward signal flaws through reward hacking, achieving high proxy scores while failing true objectives. We introduce Mechanistically Interpretable Task Decomposition (MITD), a hierarchical transformer architecture with Planner, Coordinator, and Executor modules that detects and mitigates reward hacking. MITD decomposes tasks into interpretable subtasks while generating diagnostic visualizations including Attention Waterfall Diagrams and Neural Pathway Flow Charts. Experiments on 1,000 HH-RLHF samples reveal that decomposition depths of 12 to 25 steps reduce reward hacking frequency by 34 percent across four failure modes. We present new paradigms showing that mechanistically grounded decomposition offers a more effective way to detect reward hacking than post-hoc behavioral monitoring.