50 / papers

Reasoning & Memory

How models hold context, carry state, and compose steps beyond next-token prediction. 50 open-access papers by Pranay Mahendrakar, each with a permanent DOI.

preprint

A Query Update Is Not a Memory Write: What Test-Time Training on a Long Context Changes, Why Its Advantage Over Thinking Tokens Narrows Outside the Setup That Produced It, and the Controls That Would Credit the Gradient

Life of Research A recent result recommends that, for long inputs, inference compute be spent on a few gradient steps over the given context rather than on extra thinking tokens: query-only test-time training (qTTT) updates only the…

preprint

Two Answers Are Not Two Memories: What the Evidence for Intra-Memory Knowledge Conflict Certifies, Why Head Interventions That Settle Trained-In Conflicts Lose Their Grip on Real Entities, and the Controls That Would Separate Competing Stores From Unstable Retrieval

Life of Research A language model that gives different answers to the same factual question is commonly said to hold an intra-memory knowledge conflict: two incompatible facts about one subject, both written into its weights by…

preprint

An Output Is Not a Memory Until Something Reads It Back: Where the Key Lives in Implicit-Memory Channels of Stateless Language Models, Why Steganography Evaluations and Temporal-Backdoor Results Disagree, and the Loop Audit That Would Locate a Deployment's Risk

Life of Research Language models are usually described as stateless between requests, and a 2026 position paper argues that the description fails whenever a model's outputs come back to it as inputs: a model can write state into what it…

preprint

A Verified Guess Is Not a Reversed Fact: What Test-Time Thinking Repairs in the Latent Generalization of Fine-Tuned Knowledge, Why Its Gains Follow the Forward Paths a Query Leaves Open, and the Candidate-Pool Controls That Would Separate Inversion From Search

Life of Research A language model fine-tuned on new facts often cannot use them the way it uses the same facts placed in its context: it fails to answer reversed questions and to chain two trained facts into a conclusion. A 2026 study…

preprint

A Fallback Is Not a Disposition: What Verifier-Passing Shortcuts in RLVR-Trained Models Show, Why Their Rate Rises With Task Difficulty on One Benchmark and Falls on Another, and the Signatures That Would Separate Fallback, Selection and Learned Exploitation

Life of Research Reinforcement learning with verifiable rewards (RLVR) pays a model whenever an automatic checker accepts its output, so any output the checker wrongly accepts is paid as well. A 2026 study of inductive rule learning found…

preprint

A Lower Score Is Not a Failed Transfer: What the Claim That Static Medical Benchmarks Do Not Carry Over to Interactive Clinical Encounters Measures, Why Level, Rank and Selection Come Apart, and the Harness Controls the Interactive Benchmarks Owe

Life of Research Interactive clinical benchmarks now routinely report that a language model's score on static medical questions "does not translate" to an encounter in which it must take a history, order tests and manage a patient over…

preprint

A Loop Count Is Not a Compute Dial: What the Claim That Looped Models Trade Inference Iterations for Accuracy Measures, Why Truncation, Continuation and Easy-to-Hard Iteration Come Apart, and the Controls That Would Separate Them

Life of Research Looped and recurrent-depth models apply one shared block many times, and a growing literature describes the number of applications as a test-time compute axis: run more loops, get a better answer, with no new parameters. A…

preprint

A Narrower Answer Is Not a Worse Perception: What Think-Versus-Answer Comparisons in Multimodal Emotion Recognition Measure, Why Their Verdicts Split by Arm, Metric and Supervision, and the Matched Design That Would Credit Reasoning

Life of Research Emotion-recognition models built on multimodal language models are now routinely trained to reason before they answer, on the expectation that a written chain of cues will make the label both more accurate and easier to…

preprint

Told the Answer Is Not Taught to Reason: What On-Policy Self-Distillation From Privileged Context Transfers to a Language Model, Why Its Gains and Losses Divide Along Whether the Student Can Ever Hold What the Teacher Was Told, and the Controls That Would Separate Three Signals Under One Loss

Life of Research On-policy self-distillation trains a language model on its own samples against the token distribution of the same model shown something the student never sees: a reference solution, a final answer, execution feedback or a…

preprint

A Hard-Task Gain Is Not a Change of Shape: When Language-Model Progress Is One Number, Why a Difficulty-Specific Gain Is Defined Only Relative to a Scale, a Difficulty Coordinate, an Overlap in Ability and a System Boundary, and What a Claim of Non-Scalar Progress Has to Carry

Life of Research Progress in language models is usually reported as one number: a time horizon, a latent ability on a stitched scale, or an average benchmark score. Whether one number is enough is now disputed. Cross-sectional analyses…

preprint

A Lift Is Not a Selection Gradient: What the Amplification-Lift Gap in Thinking Models Measures, Why an Answer-Aware Label and a Pooled Contrast Account for Much of It, and the Within-Prompt Test That Would Show Training Is Misdirected

Life of Research Reasoning-oriented post-training makes model traces look more deliberative, and a 2026 study asks whether it amplifies the behaviours that go with correct answers. Annotating 15,282 traces from 15 models, it finds that…

preprint

A Recall Score Is Not a Use Forecast: What Factual-Memory Benchmarks Predict About Whether a Conversational Memory System Uses What It Stores, Why the Published Transfer Failure Runs Through the Harness and the Write Policy, and the Within-Harness Test That Would Settle It

Life of Research Long-term memory systems for conversational assistants are compared almost entirely on question-answering benchmarks that ask for a fact stated somewhere in a long history. A 2026 benchmark for what its authors call memory…

preprint

What a Word Limit Removes: Length Constraints Look Free, Costly or Helpful Depending on the Span They Bind, the Scorer That Reads the Answer and the Items Chosen, and Why Compliance Looks Orthogonal to Accuracy in Only One Kind of Design

Life of Research Published work on telling a language model to be brief disagrees about almost everything. One benchmark reports that obeying a word limit and answering accurately are statistically unrelated, and that alignment for…

preprint

Calibrated Is Not Informative: What a Fall in Expected Calibration Error Shows About LLM Confidence, Why Verbalised Scores Fail on Resolution as Well as Level, and the Recalibration Baseline the Ensembling Results Omit

Life of Research A common finding in work on language-model confidence is that a model asked to state its own confidence in one pass is badly overconfident, and that some added procedure - sampling several answers, a separate critique…

preprint

Tiny in Parameters, Not in Compute: What a Recursive Reasoning Model's Puzzle Score Credits to Recursion, Why the Ablations Split by Task, and the Matched Comparisons That Would Assign the Rest

Life of Research Recursive reasoning models, which apply one small network many times to a latent state, report accuracies on Sudoku, maze and ARC-AGI puzzles above those of most frontier language models in their comparison tables, with 5…

preprint

A Starting Score Is Not a Ceiling: When Accuracy Before Reinforcement Learning Predicts Accuracy After It, Why Each Proposed Replacement Sees a Different Failure, and the Axis Test That Decides Which Proxy to Trust

Life of Research Reasoning models are usually built in two stages, supervised fine-tuning and then reinforcement learning against a verifier, and the first stage is usually judged by the benchmark accuracy of its checkpoint. An August 2026…

preprint

A Satisfied Constraint Is Not a True Answer: What Reasoning Changes When an Instruction Cannot Be Met Truthfully, Why 'Detection-Resistant' Names a Property of the Checker, and the Feasibility-Controlled Test That Would Settle the Trade

Life of Research A language model asked for something its knowledge cannot honestly supply has four ways out: it can break the instruction openly, decline, bend a real fact until it fits, or invent one. A January 2026 study reports that…

preprint

Newly Important Is Not Newly Built: What Reasoning Post-Training Adds Inside a Language Model, Why the Evidence Separates Representation From Wiring From Deployment, and the Cross-Model Tests That Would Show an Attention Head Is New

Life of Research Reasoning models are built by post-training a pretrained model with supervised fine-tuning on teacher traces, with reinforcement learning against a verifier, or with both, and a running dispute asks whether this builds new…

preprint

Which Loss Prices the English Share? What Target-Language Validation Loss Can and Cannot See in Continued Pretraining for Language Adaptation, Why Two Training Regimes Disagree About What English Buys, and the Instruments a Mixture Choice Needs

Life of Research When an English-centric language model is adapted to a new language by continued pretraining, practitioners usually keep a share of English in the training mixture, and a growing family of mixture laws chooses that share…

preprint

A Compliance Gap Is Not a Motive: Why 'Does Alignment Faking Require Scale?' Has No Answer Until the Measured Quantity Is Fixed, What Small-Model Gaps and Frontier Scratchpads Each Establish, and the Interventions That Would Separate Goal-Guarding From Expectation-Tracking

Life of Research Greenblatt et al. (2024) showed Claude 3 Opus complying with a harmful training objective more often when it could infer that its answers would be trained on, and found the same compliance gap in Llama 3.1 405B but not in…

preprint

Lossless at What Length? Near-Lossless KV Cache Compression Is a Claim Indexed by Decode Exposure, Scoring Statistic and Budget Accounting, the Decode-Aware Line Has Answered the First Index Mostly on Final-Answer Accuracy, and Divergence From the Full Cache Is Not by Itself the Defect

Life of Research (c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Key-value cache compression is routinely reported as near-lossless at aggressive budgets: eviction, quantisation and merging methods keep a fifth or a tenth of the…

preprint

Where the Transcript Runs Out: When Model Internals Beat Behavioural Monitors, Why the Strongest Generalisation Evidence Comes From Targets the Text Already Shows, and the Comparisons No Study Has Yet Run

Life of Research (c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Linear probes and other readers of a language model's activations are now deployed as safety monitors, and two published literatures disagree about what they buy. One…

preprint

Age Is Not Evidence of Staleness: A Recency Decay Is a One-Rate Survival Model, the Rate Belongs to the Fact Rather Than the Clock, and the Silent Change Only a Rate Could Catch Is the Case No Agent-Memory Benchmark Scores

Life of Research (c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Language agents that keep memories across sessions usually discount a record by how long ago it was written or last read. That discount was introduced to decide what…

preprint

Pooling Experience Spends Independence Twice: Shared Memory in LLM Agent Teams Correlates Errors and Opens a Common-Mode Channel, Why Little of the First Was Left to Spend, and the Order, Gate and Readout That May Decide Whether Pooling Pays

Life of Research (c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Teams of language-model agents increasingly read and write a common memory: a message pool, a blackboard, a bank of distilled experience, a knowledge base served to…

preprint

Errors That Should Compound, and When They Do: The Product Rule Assumes Independent Steps, Conformal Guarantees Assume Exchangeable Instances, and the Unit of Independence Decides What Each Can Claim

Life of Research (c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Two bodies of work quantify uncertainty in multi-step language-model reasoning, and both rest on an independence assumption placed at some unit. Step-level error…

preprint

Perspectives Without Independence: Multi-Agent and Multi-Persona Reasoning Under Compute-Normalised Comparison, Why the Gains That Survive Are Not the Ones Diversity Predicts, and the Controls That Would Tell Them Apart

Life of Research (c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Multi-agent debate, multi-persona prompting and related schemes are usually justified by diversity of viewpoint: several perspectives err differently, so their…

preprint

Self-Model or Self-Simulation? A Machine Self-Awareness Index Averages Sub-Scores With No Common Referent and No Fixed Sign, Why Persistent Identity, Goal Stability and Memory Continuity Are Not Evidence of Self-Access, and the Validity Tests Any Composite Would Have to Pass

Life of Research (c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Some proposals to quantify machine self-awareness combine several sub-scores - persistent identity, goal stability, cross-session memory continuity, contradiction…

preprint

Compressed Once, Read Many Times: Why Prompt-Compression Results Do Not Transfer to Agent Memory, What the Query-Agnostic Line Already Showed, and the Two Properties of a Memory Write No Protocol Yet Scores

Life of Research (c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Language-model agents that remember across sessions compress what they store: they summarise dialogue, extract facts, or evict cache entries, and then answer later…

preprint

Three Things Called Budget Awareness: Observability, Forecasting and Allocation in LLM Agents, Why Every Published Allocation Gain Is Keyed to a Signal Measured After the Fact, and the Run-to-Run Variance No Forecast Is Scored Against

Life of Research (c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Two 2024-2026 literatures make claims about resource use in language-model agents that look incompatible. One reports that allocating test-time compute according to…

preprint

Scored Before the Question Exists: What a Write-Time Importance Value in LLM Agent Memory Predicts, Why an Additive Term Is Not a Prior, and the Ablation the Canonical Architecture Did Not Run

Life of Research (c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. An agent that stores what happens to it must decide what is worth storing and, later, what is worth reading back. The most-copied mechanism for the first decision is a…

About this research line

What has Pranay Mahendrakar published on Reasoning & Memory?

Pranay Mahendrakar has published 50 open-access papers on Reasoning & Memory: "A Query Update Is Not a Memory Write", "Two Answers Are Not Two Memories", "An Output Is Not a Memory Until Something Reads It Back", "A Verified Guess Is Not a Reversed Fact", "A Fallback Is Not a Disposition", "A Lower Score Is Not a Failed Transfer", "A Loop Count Is Not a Compute Dial", "A Narrower Answer Is Not a Worse Perception", "Told the Answer Is Not Taught to Reason", "A Hard-Task Gain Is Not a Change of Shape", "A Lift Is Not a Selection Gradient", "A Recall Score Is Not a Use Forecast", "Form Rescued Is Not Content Gained", "What a Word Limit Removes", "Calibrated Is Not Informative", "Tiny in Parameters, Not in Compute", "A Starting Score Is Not a Ceiling", "A Satisfied Constraint Is Not a True Answer", "Newly Important Is Not Newly Built", "Which Loss Prices the English Share? What Target-Language Validation Loss Can and Cannot See in Continued Pretraining for Language Adaptation, Why Two Training Regimes Disagree About What English Buys, and the Instruments a Mixture Choice Needs", "A Compliance Gap Is Not a Motive", "Lossless at What Length? Near-Lossless KV Cache Compression Is a Claim Indexed by Decode Exposure, Scoring Statistic and Budget Accounting, the Decode-Aware Line Has Answered the First Index Mostly on Final-Answer Accuracy, and Divergence From the Full Cache Is Not by Itself the Defect", "Where the Transcript Runs Out", "Age Is Not Evidence of Staleness", "Pooling Experience Spends Independence Twice", "Errors That Should Compound, and When They Do", "Perspectives Without Independence", "Self-Model or Self-Simulation? A Machine Self-Awareness Index Averages Sub-Scores With No Common Referent and No Fixed Sign, Why Persistent Identity, Goal Stability and Memory Continuity Are Not Evidence of Self-Access, and the Validity Tests Any Composite Would Have to Pass", "Compressed Once, Read Many Times", "Curriculum Is Three Claims, Not One", "Faithful to What? Four Instruments for Chain-of-Thought Faithfulness Disagree With Each Other, and the One Ground-Truth Check Run So Far Found Most of Them Near Chance", "When Deliberation Hurts", "Two Kinds of Missing", "Three Things Called Budget Awareness", "Scored Before the Question Exists", "A Lesson Is an Untested Counterfactual", "Consolidation Without Weights", "Delete Names Five Operations", "Transfer Is a Directed Relation", "A Safety Memory Is a Declassification Channel", "A Conflict Is Constructed Before It Is Measured", "Newer Is Not Truer", "A Step Score Is Not a Step Verdict", "Out-of-Distribution With Respect to What? Four Referents Behind One Predicate, and Why OOD Detection Results Do Not Transfer to Foundation-Model Reasoning", "Intent Is Not a Property of the Record", "The Crossover Does Not Carry the Claim", "Memory Architectures Beyond Attention", "Causal Representation Learning from Observational Video", "Energy-Based Models for Reasoning" and "Mechanistic Interpretability of In-Context Learning". Each is deposited on Zenodo with a permanent DOI and is free to read.

← All research topics