[ RESEARCH ]

Research

Five papers, each with a small demo of the idea. Most look at how language models fail and what post-training can do about it.

Google Scholar ↗

P.012026SemEval-2026 · ACL

ModusPonens: Breaking the Plausibility Trap in LLMs via Conflict-Aware Ensembling ↗

Two LoRA-adapted Qwen-2.5-14B models, one standard and one hardened on logical-conflict data, fused with logit-level soft voting. The ensemble shows less content bias at the same accuracy.

LLM reasoningLoRAEnsembling

↳INTERACTIVE DEMO

[ DEMO · BELIEF BIAS ]

Can a model tell valid from believable?

Pick an argument and flip whether its conclusion sounds true. The logic doesn't change, so watch which models get fooled.

ARGUMENT

CONCLUSION SOUNDS

P1All mammals breathe air.

P2Whales are mammals.

∴Whales breathe air.

P(VALID) BY MODELTRUTH: VALID

BelieverQwen-2.5-14B · standard LoRA

✓ CORRECT

0.92

SkepticQwen-2.5-14B · bias-hardened

✗ FOOLED

0.45

EnsembleLogit-level soft vote

✓ CORRECT

0.69

ILLUSTRATIVE. NUMBERS MIMIC THE FAILURE PATTERN FROM MODUSPONENS ↗, NOT ACTUAL MODEL OUTPUTS.

P.022026SemEval-2026 · ACL

Ambirig: Distributional Ordinal Modelling for Ambiguous Word Senses in Narrative Contexts ↗

A DeBERTa-v3 cross-encoder trained with an Earth Mover's Distance loss, so disagreement between annotators is modelled as a full distribution instead of being averaged away.

Word senseOrdinal regressionEMD loss

↳INTERACTIVE DEMO

[ DEMO · WORD-SENSE PLAUSIBILITY ]

How plausible is this meaning of “bank”?

Five people rate the sense on a 1–5 scale. Change the ending and compare how each loss captures the spread of human answers.

The old fisherman sat by the bank as the light faded.

RATE THIS SENSE: “THE LAND ALONGSIDE A RIVER”

STORY ENDING

TRAINING LOSS

RATING DISTRIBUTION · 1–5EMD TO HUMANS: 0.12

ANNOTATORSMODEL · EMD

ILLUSTRATIVE. DISTRIBUTIONS SHOW THE BEHAVIOUR, NOT THE PAPER’S OUTPUTS.

P.032025BLP-2025 · AACL-IJCNLP

CodeAnubad: Efficient Bangla-to-Python Code Generation via Iterative LoRA Fine-Tuning of Gemma-2 ↗

QLoRA fine-tuning of Gemma-2-9B with a self-improvement loop that grows a very small training set with verified-correct predictions. It outperforms translation and RAG baselines.

Low-resourceCode generationQLoRA

↳INTERACTIVE DEMO

[ DEMO · SELF-IMPROVEMENT LOOP ]

Fewer than 100 examples, then a loop

Step through the rounds and ranks. Each round adds only generations that pass their tests.

SELF-IMPROVEMENT ROUND

LORA RANK

· Fine-tune on the seed set
· Generate on dev and run the tests
· Keep only verified-correct pairs, retrain

GEMMA-2-9B · QLORA RANK 32ROUND SEED

একটি ফাংশন লিখুন যা একটি তালিকার সংখ্যাগুলোর যোগফল ফেরত দেয়।

“WRITE A FUNCTION THAT RETURNS THE SUM OF A LIST OF NUMBERS.”

def total(nums):
    s = 0
    for n in nums:
        s += n
    print(s)   # returns None

TESTS0 / 3 PASS
TRAIN POOL74 PAIRS
DEV PASS@140%

REPORTED BEST: 47% DEV, 37% HIDDEN TEST (RANK 32). INTERMEDIATE VALUES ARE ILLUSTRATIVE.

P.042025ACROSET

PlantDiseaseNet-RT50: A Fine-tuned ResNet50 Architecture for High-Accuracy Plant Disease Detection Beyond Standard CNNs ↗

A ResNet50 fine-tuned for plant disease classification from leaf images, outperforming standard CNN baselines.

Computer visionTransfer learning

↳INTERACTIVE DEMO

[ DEMO · LEAF CLASSIFIER ]

Standard CNN vs. fine-tuned ResNet50

Pick a leaf and a model. Look-alike diseases are where fine-tuning pays off.

LEAF SAMPLE

MODEL

TOP-3 PREDICTIONSPLANTDISEASENET-RT50

Tomato · early blight ✓93%

Tomato · late blight 4%

Tomato · septoria 2%

TRUE LABEL · TOMATO · EARLY BLIGHT✓ CORRECT

ILLUSTRATIVE CONFIDENCES. NOT THE PAPER’S REPORTED NUMBERS.

P.052024ICCCNT

Underwater colour: Unveiling Natural Based Underwater Image Colour Enhancement ↗

Colour enhancement for underwater images that restores colour lost to depth-dependent light absorption.

Image processingColour correction

↳INTERACTIVE DEMO

[ DEMO · COLOUR RECOVERY ]

What the camera sees at depth

Drag the depth. The middle row is what the camera captures; the bottom row is after enhancement.

DEPTH8 m

Water absorbs red first, then green, so deeper scenes turn blue-green. Enhancement estimates the lost channels and rebalances them.

SCENE COLOURSDEPTH 8 M

SURFACE

CAPTURED

ENHANCED

SIMPLIFIED ABSORPTION MODEL WITH GREY-WORLD REBALANCING, FOR INTUITION ONLY.