Robust Dual-Signal Fusion: Hybrid Neuro-Symbolic Gating with Compressed Chain-of-Thought Refinement for Irony Detection in Social Media Texts Ankit Bhattacharjee∗1 & Krityapriya Bhaumik†1
arXiv:2606.16845v1 [cs.CL] 15 Jun 2026
1
Indian Institute of Technology Kharagpur
June 16, 2026
Abstract
TokenSkip [1] reduce inference overhead via token pruning but depend on SFT for domain adaptation and are primarily validated on structured datasets (MATH, GSM8K), exposing a critical gap for unstructured social media text. While fine-tuning remains the standard paradigm for adapting LLMs to narrow classification tasks, it carries significant practical drawbacks: catastrophic forgetting of generalist capabilities and the deployment overhead of maintaining distinct model weights for every target domain. The objective of our framework is not to replace supervised fine-tuning, but to bypass the necessity for it. By demonstrating that inference-time neuro-symbolic reasoning can recover fine-tuned performance levels, we establish a mechanism to achieve task-specific parity while strictly preserving the frozen base model’s generalist capabilities. We identify two concurrent failure modes in naive CoT compression applied to irony: (1) a literal bias in zeroshot LLMs that causes over-estimation of literal intent, and (2) a precision collapse when explicit heuristic priors are added to maximize recall, as aggressive symbolic overfiring floods the system with false positives. Neither failure can be resolved in isolation. We propose the Robust Dual-Signal (RDS) Fusion framework to resolve this tension. RDS is a hybrid neuro-symbolic architecture with three pillars: a GradientSensitized Local Guardian for surgical CoT compression, a linguistically-motivated symbolic prior for explicit irony detection, and an Entropy-Gated adaptive fusion engine. Crucially, we empirically demonstrate that the symbolic prior is load-bearing for recall while the CoT pipeline functions strictly as a precision-recovery engine, suppressing false positives generated by heuristic over-firing without SFT.
Large Language Models (LLMs) natively default to literal semantic interpretations, making zero-shot irony detection a persistent challenge. We introduce the Robust Dual-Signal (RDS) Fusion framework, a hybrid neuro-symbolic architecture that compresses Chain-of-Thought (CoT) reasoning [2] trajectories without Supervised Fine-Tuning (SFT). Evaluated on a strictly held-out TweetEval test set [3] (N = 734), RDS achieves 78.1% accuracy and a Macro F1 of 0.777, matching the absolute performance ceiling of the fine-tuned BERTweet [4]. On the heavily imbalanced iSarcasm dataset, the frozen CoT pipeline filters 22.5% of out-of-distribution hallucinations, yielding a zero-shot Macro F1 of 0.6726 and Ironic F1 of 0.4821, outperforming multiple heavily supervised SemEval transformer ensembles. A statistical ablation confirms this structural synergy: adding the symbolic prior to the neural baseline yields no significant gain (p = 0.242), and the marginal benefit of adding the CoT pipeline to that prior is heavily compressed (p = 0.149). Only the complete, concurrent fusion of all three signals achieves a statistically validated improvement over the baseline (p = 0.005).
1
Introduction
Detecting irony in unstructured social media text remains a persistent NLP challenge. LLMs natively default to literal interpretations, frequently missing the contextual inversions that define irony. In fact, standalone few-shot evaluation of Qwen2.5-3B-Instruct on unstructured irony yields sub-random performance (Macro F1 ¡ 0.500), definitively proving that LLMs cannot function as primary detectors in this domain without structural grounding. While Chain-of-Thought (CoT) prompting enables models to map complex pragmatics [2], unconstrained CoT introduces severe computational latency, making it impractical Primary Contributions for high-throughput classification [1]. Frameworks like • A parameter-free, few-shot CoT compression architecture: driven by a Gradient-Sensitized Local ∗ [email protected] Guardian, dynamically whitelisting structural linguistic † [email protected] 1
2
Bhattacharjee & Bhaumik (2026)
anchors without SFT. • Empirical reframing of few-shot reasoning in irony detection: the CoT pipeline acts strictly as a precision-restoration filter, not a primary detector. • TweetEval validation (N = 734): 78.1% accuracy, Macro F1 of 0.777, matching fine-tuned BERTweet. The only configuration achieving statistically significant improvement over fine-tuned RoBERTa (p = 0.005). • Robust few-shot cross-domain generalization on iSarcasm without parameter modification: 22.5% false positive suppression, Ironic F1 of 0.4821, outperforming multiple supervised SemEval systems.
3.1
Baseline Pruning Formulation
Given LLM M and CoT trajectory c = {ci }m i=1 , vanilla TokenSkip assigns importance I(ci ) to each token and prunes below the γ-quantile threshold Iγ = Qγ (I(c1 ), . . . , I(cm )), retaining c̃ = {ci | I(ci ) ≥ Iγ }. A static γ and probabilitybased I(ci ) are insufficient for unstructured irony.
3.2
Continuous Confidence and Two-Pass Fallback
RDS prompts the generator LLM to produce a structured CoT concluding with a continuous confidence score PCoT ∈ [0, 1]. Extraction uses a hierarchical parser: (1) explicit score matching; (2) negation-aware semantic scan 2 Related Work of the final three sentences, bounded to [0.15, 0.85]; and (3) 2.1 CoT Efficiency and Token Compres- full-text fallback. If primary extraction yields PCoT = 0.5 (unresolvable ambiguity), a two-pass fallback reprompts sion the LLM with a 10-token forced-verdict template to proCoT prompting decomposes complex reasoning into in- duce a definitive score. Full heuristic vocabulary and regex termediate steps [2] but introduces quadratic attention patterns are provided in Appendix A. overhead [5] and linear KV-cache growth. Compression approaches range from perplexity-based token filtering [7] to 3.3 Few-Shot Chain-of-Thought Prompt Design bidirectional token importance classification (LLMLingua2 [8]). TokenSkip [1] applies a quantile-thresholded prun- The generative pipeline relies on a 4-shot Chain-ofing over importance scores but suffers four key limita- Thought (CoT) prompt to evaluate irony. To prevent tions in noisy social media domains: (1) no calibration selection bias, the four in-context examples were manually guarantee—static γ ignores input complexity; (2) over- authored to represent canonical boundary cases (pure litsensitivity to surface tokens whose downstream classifica- eral, pure sarcasm, ambiguous, and hashtag-driven) rather tion impact differs from generation probability; (3) static than being sampled from the TweetEval or iSarcasm evalsnapshot constraints that ignore contextual threads; and uation sets. (4) inability to capture explicit social markers (hashtags, contradictory emojis) without SFT. RDS addresses all four You are an expert linguist specializing in detecting irony and sarcasm in social media text. via entropy-adaptive budgeting and gradient-sensitized local retention. DEFINITION: A tweet is IRONIC if there is a
2.2
Neuro-Symbolic and Training-Free Inference
Inference-time compute scaling (e.g., OpenAI o1 [6]) demonstrates that dynamic reasoning budgets improve performance without parameter updates. Hybrid neurosymbolic architectures [16, 17] combine neural continuous processing with symbolic discrete overrides, providing the structural grounding needed for explicit irony markers that neural pipelines smooth away [13]. RDS instantiates this paradigm for compressed CoT in social media irony.
3
Methodology
RDS optimizes CoT compression through two complementary modules—a macro-level dynamic token budgeter and a micro-level gradient-sensitized retention filter—followed by a rule-based symbolic prior and an adaptive fusion engine. Figure 1 illustrates the full pipeline.
contrast between its literal meaning and its intended meaning, or if the author says the opposite of what they actually mean (often to mock, criticize, or be humorous). A tweet is NON-IRONIC if it is a sincere, literal statement. KEY SIGNALS TO CHECK: - Does the literal meaning contradict the real-world situation? - Is there an exaggerated, over-the-top positive/negative tone? - Are there hashtags like #not, #sarcasm, #irony, #obviously that signal ironic intent? - Does the tweet mock or criticize something by pretending to praise it? - Would a reasonable reader take this at face value, or detect a hidden meaning? - Are there elongated words like ”Loooove” or ”Soooo” used sarcastically? CONFIDENCE SCALE: 0.0 = Absolutely certain NON-IRONIC (sincere, literal, no ambiguity at all) 0.1 = Very likely non-ironic, tiny residual doubt 0.3 = Probably non-ironic, some mixed signals present 0.5 = Completely uncertain — could genuinely be either 0.7 = Probably ironic, some mixed signals present 0.9 = Very likely ironic, tiny residual doubt 1.0 =
Robust Dual-Signal Fusion
3
Figure 1: The RDS pipeline architecture.
Absolutely certain IRONIC (clear sarcasm/irony, no ambiguity at all). EXAMPLES: Tweet: “Oh great, another Monday. Just what I needed.” Reasoning: “Just what I needed” is exaggeratedly positive about something universally disliked. Classic sarcasm with no ambiguity. Score: 0.95 Tweet: “Happy birthday to my best friend! Hope your day is amazing.” Reasoning: Sincere, literal birthday wish. No hidden meaning, no contrast, tone matches content perfectly. Score: 0.05 Tweet: “Wow, love how my flight got cancelled on the day of my interview. Truly blessed.” Reasoning: “Truly blessed” after describing a disaster is a clear ironic inversion. Very high confidence. Score: 0.92 Tweet: “This weather is something else today.” Reasoning: Ambiguous — could be genuine admiration or sarcastic complaint depending on context not available in the tweet alone. Score: 0.50 Now analyze the following tweet using the same reasoning process. Tweet: {tweet} Think step by step through the KEY SIGNALS above. End your response with EXACTLY: Score: X.XX (a number between 0.00 and 1.00, two decimal places)
3.4
Entropy-Based Dynamic Budgeting
We replace static γ with a dynamic ratio γdyn calibrated to input complexity via normalized token-level predictive entropy: m
H(c) = −
1 X P (ci |{cj }ij=1 ) log P (ci |{cj }ij=1 ). m i=1
(1)
High entropy indicates structural ambiguity requiring more context. We map H(c) to γdyn via a scaled sigmoid: γmax − γmin γdyn = γmin + , (2) 1 + exp[−k(H(c) − µ)] where µ is the empirical mean entropy and k controls sensitivity. Bounds (γmin , γmax ) = (0.1, 0.9) prevent degenerate compression. A non-linear sigmoid transformation is mathematically superior to a naive linear mapping because it stabilizes the compression variance for median-entropy inputs. By centering the inflection point at µ, the function prevents minor generation fluctuations from causing erratic budget shifts, actively reserving the extreme retention bounds strictly for high-uncertainty outliers and highly confident literal trajectories.
3.5
Gradient-Sensitized Retention (Local Guardian)
We replace probability-based I(ci ) with a task-aware piecewise-constant function. Let Gk denote the k = 5
4
Bhattacharjee & Bhaumik (2026)
tokens with highest gradient norms via backward pass Phase 2: Base Fusion. The trust weight α (confidence through the RoBERTa classification head [11, 12]: in the RoBERTa baseline) is linearly interpolated over [Hmin , Hmax ] with bounds [αmin , αmax ], then overridden if S(ci ) = ∥∇eci f (c)∥2 , (3) |P RoBERTa − P̃CoT | > ∆conf : where f (c) is the token distribution function and eci is the vector embedding of the i-th token. Let C denote contrastive connectors and W the English stopword list. The robust importance score is: τcrit if ci ∈ Gk τ conn if ci ∈ C Irobust (ci ) = (4) τ if ci ∈ / W, |ci | ≥ 3 cont τlow otherwise,
Pbase = αPRoBERTa + (1 − α)P̃CoT .
(9)
Calibrated values: Hmin = 0.2, Hmax = 0.5, αmin = 0.50, αmax = 0.88, Hsk = 0.25, λsk = 0.5, ∆conf = 0.85, αconf = 0.85.
Phase 3: Prior Injection. Strong signals cap α ← min(α, αcap = 0.30) and apply a heavy blend (ωstrong = with (τcrit , τconn , τcont , τlow ) = (1000, 500, 10, 1). The 0.60); weak signals use a moderate blend (ωweak = 0.75); multi-order-of-magnitude gaps guarantee lexicographic no signal defaults to Pbase . The final decision is ŷ = 1 if priority: gradient-critical tokens and contrastive connec- P (ŷ = 1|x) > 0.5. tors are guaranteed to survive any γdyn < 1, regardless of compression budget. The compressed trajectory is: 4 Experimental Setup c̃RDS = top
Irobust (ci ), S⊆c |S|=⌊(1−γdyn )m⌋
order preserved.
4.1
Datasets
(5) TweetEval Irony [3] (SemEval-2018 Task 3) : 784 annotated tweets serving as primary evaluation bench3.6 Linguistically-Motivated Prior Mod- mark. We designate the first 50 tweets as a validation ule split for parameter calibration; all reported results are on To capture explicit authorial markers that neural pipelines the strictly held-out 734-tweet test set. smooth away, we define a symbolic prior Sprior (x) iSarcasm (SemEval-2022 Task 6, English Task A) : over tiered signal sets. Let Rstrong denote definitive 1400 self-reported sarcasm tweets (N ironic = 200, Nliteral = author-labeled irony markers, Rweak soft hashtag signals, 1200). The severe class imbalance and implicit-intent Remoji contradictory emoji patterns, and Relong character- distribution provide an out-of-distribution stress test. The elongation tokens. The full lexicon definitions appear in full RDS framework is deployed without any parameter Appendix A. The prior is: modification on this dataset. ( ρstrong if x ∩ Rstrong ̸= ∅ Sprior (x) = (6) 4.2 Models and Evaluation Pipeline Scomp (x) otherwise, Three components constitute the pipeline: (1) Qwen2.53B-Instruct [9] generates uncompressed CoT trajectories under 4-bit NF4 quantization, greedy decoding, max new tokens=120; (2) LLMLingua-2 [8] executes token pruning at threshold Iγdyn ; (3) cardiffnlp/twitterroberta-base-irony provides the discriminative baseline Calibrated values: ρstrong = 0.88, ρweak = 0.15, ρemoji = and gradient signals. Hardware and library details are in 0.25, ρelong = 0.12, ρcap = 0.70, η = 3. Appendix B.
where the composite weak score is capped to prevent overfiring: Scomp (x) = min ρcap , δw ρweak + δe ρemoji + δℓ ρelong . (7)
3.7
Entropy-Gated Adaptive Inference 4.3 Evaluated Configurations Fusion Three configurations isolate each component’s contribu-
The multi-stage fusion engine balances trust among PCoT , tion: (i) RoBERTa-only baseline: task-specific finetuned encoder; (ii) Ablation (RoBERTa + Hashtag PRoBERTa , and Sprior . Prior): symbolic prior injected without CoT; (iii) Full Phase 1: CoT Skepticism Gate. To prevent overconfiRDS: complete neuro-symbolic pipeline. dent LLM hallucinations, extreme PCoT predictions are A standalone few-shot LLM configuration was evaluated contracted toward 0.5 when H(c) > Hsk : If H(c)>Hsk but excluded from the primary tables due to sub-random and |PCoT −0.5| ≥ 0.5−psk : performance (Macro F1 < 0.500), validating its restricted P̃CoT = 0.5 + λsk PCoT − 0.5 ; else P̃CoT = PCoT . (8) role as a supplementary filter.
Robust Dual-Signal Fusion
5
Results and Discussion
5.1
System-Level Performance
The full RDS framework resolves this via the compressed CoT pipeline. Rather than acting as a primary detector, the LLM functions as a precision-restoration filter— rescuing 68 false positives (FP: 178 → 110) while maintaining ironic recall of 0.824. This calibration yields 78.1% accuracy and Macro F1 of 0.777. The structural necessity of the Chain-of-Thought pipeline is visually confirmed across system metrics. As illustrated in Figure 2 and the corresponding confusion matrices (Figure 4), the symbolic ablation artificially maximizes ironic recall at the cost of a severe precision collapse, generating 178 false positives. The full RDS framework dynamically suppresses these heuristics to recover precision, systematically reducing false positives to 110. Furthermore, this precision recovery is tightly calibrated to reasoning complexity; Figure 3 demonstrates that RDS accuracy degrades logically as structural entropy increases, whereas the ablation’s error distribution remains entirely uncorrelated with predictive ambiguity. Table 1 presents the three-way comparison on the held-out TweetEval test set. The standalone RoBERTa achieves 73.2% accuracy with a low false positive rate but suffers from ironic recall of only 0.557. Injecting the symbolic prior (ablation) raises accuracy to 75.7% but at a critical cost: the fusion formula forces ironic recall to 1.000 (0 false negatives) by triggering on every hashtag-containing tweet, generating 178 false positives and collapsing precision to 0.619. The Macro F1 of 0.757 is therefore misleadingly inflated.
5.2
Statistical Significance
5
trophically fails on literal tweets with irony-associated hashtags (Non-Ironic F1 = 0.131). RDS fusion recovers Non-Ironic F1 from 0.131 to 0.508 via CoT-anchored suppression of over-fired priors. On the 317 non-triggered tweets, both configurations achieve Ironic F1 = 0.000, confirming that the few-shot CoT pipeline cannot detect implicit irony in a vacuum—the symbolic prior is strictly load-bearing for recall. Table 2 demonstrates the full subpopulation breakdown.
5.4
Calibration Dynamics
The Entropy Budgeter exhibits a statistically significant negative correlation between entropy and accuracy for RDS (r = −0.104, p = 0.005), versus no correlation for the ablation (r = −0.028, p = 0.456). The 10.9 percentagepoint accuracy spread across entropy quartiles (vs. 3.9 for the ablation) proves that the budgeter actively quantifies structural ambiguity rather than merely measuring text difficulty. The α distribution clusters heavily at the [0.30, 0.40] floor (55.4% of inferences), confirming the system’s conservative failsafe design. However, α > 0.5 on 41.3% of inferences demonstrates the parameter’s active contribution. The dynamic budget coefficient γdyn exhibits near-zero variance (mean=0.466, σ = 0.013), revealing that macrolevel budgeting operates near-constantly due to the tight CoT length distribution (mean=83.7, σ = 5.39, range 60– 104). This confirms that the precision recovery is driven primarily by the micro-level Local Guardian rather than the macro budgeter.
5.5
Comparison with Supervised State-ofthe-Art
Table 3 shows RDS achieving functional parity with BERTweet (78.1% vs. 78.2% accuracy; Ironic F1 0.747 vs. 0.746) and surpassing CCR-Net on Macro F1—entirely without downstream fine-tuning. The 50-tweet calibration holdout prevents a strictly identical numerical comparison, but proximity strongly suggests that well-calibrated neuro-symbolic fusion can approach the output of heavily domain-trained encoders through inference-time reasoning alone.
Paired McNemar’s tests across configurations confirm the synergistic premise. Adding the hashtag prior alone over the RoBERTa baseline does not yield significance (χ2 = 1.37, p = 0.242)—the recall gains are neutralized by the 178 false positives. The direct comparison between full RDS and ablation also falls short of significance (χ2 = 2.08, p = 0.149), a mathematical consequence of the ablation’s recall-saturated operating point compressing the disagreement space. Only the full RDS framework 5.6 Few-Shot Cross-Domain Robustness achieves a statistically validated improvement over the on iSarcasm standalone RoBERTa baseline (χ2 = 7.95, p = 0.005). Neither component alone is sufficient; only concurrent Deployed without modification on iSarcasm, the RoBERTa dual-signal synergy produces a validated classification baseline generates 315 false positives due to distribution shift. The full RDS framework suppresses 22.5% of shift. these false positives (315 → 244; McNemar’s χ2 = 21.44, 5.3 Subpopulation Analysis p < 0.0001), elevating Macro F1 from 0.6562 to 0.6726. Partitioning the test set by hashtag detector activation iso- Crucially, this precision recovery does not degrade relates the fusion mechanism. On the 417 triggered tweets, call; RDS strictly preserves the baseline’s Ironic F1 the ablation achieves a high Ironic F1 of 0.829 but catas- (0.4790 → 0.4821).
6
Bhattacharjee & Bhaumik (2026)
Figure 2: Ironic class metrics across configurations. Ablation’s perfect recall is a structural artifact forcing precision collapse; RDS CoT restores balance.
Figure 3: Accuracy by entropy bin (N = 734). RDS degrades monotonically with ambiguity; the ablation is uncorrelated with entropy.
Figure 4: Confusion matrices for all three TweetEval configurations. The RoBERTa baseline is conservative (FP=69, FN=127); the ablation enforces zero false negatives at the cost of FP=178; full RDS fusion reduces FP to 110 while keeping FN=60.
As shown in Table 4, few-shot RDS outperforms three supervised SemEval submissions. However, iSarcasm also exposes architectural limits: the hashtag prior fires on only 1.5% of samples (n = 21), confirming its inertia on implicit intent. On the 21 triggered instances, frozen TweetEval weights allow the LLM’s literal confidence to overpower valid priors (Ironic F1: 0.889 → 0.571), defining the cross-domain calibration boundary.
creating the isolated substring #not. The explicit hashtag detector immediately identifies this as a strong irony signal (SHTag = 0.88).
5.7
Case Studies
5.7.1
Case Study 1: Precision Recovery on a Literal Legal Headline (Tweet 86, TweetEval Dataset)
RDS correction: The CoT pipeline parses the semantic continuity of the legal phrase “plead not guilty” rather than treating “#not” as an isolated pragmatic inversion. During compression, the Local Guardian explicitly whitelists literal anchoring tokens (British, indicates) against the macro-budgeter. The LLM outputs PCoT = 0.15. Routed at α = 0.300, this contradicts the prior and pulls the fused probability to Pfused = 0.453, yielding a correct Non-Ironic prediction. This interaction
Tweet: #Myanmar #men #plead #not #guilty to #murder of #British #tourists... This is a strictly literal news headline. The author has tokenized the text using individual hashtags, inadvertently
Ablation failure: Despite RoBERTa’s low baseline probability (prob = 0.209), the strong symbolic trigger pushes the fused probability above 0.500, yielding a false positive. This is the canonical structural artifact of the overaggressive prior.
Robust Dual-Signal Fusion
Method
7
Accuracy
Macro F1
Ironic F1
Ironic Precision
Ironic Recall
Non-Ironic F1
73.2 75.7 78.1
0.706 0.757 0.777
0.620 0.765 0.747
0.700 0.619 0.684
0.557 1.000† 0.824
0.792 0.750 0.806
RoBERTa-only Ablation (Rob + Prior) RDS Fusion
Table 1: Performance on TweetEval Irony test set (N = 734). † Ablation’s ironic recall of 1.000 is a structural artifact forcing precision collapse. Subset
Model Ablation (Rob + Prior) RDS Fusion Ablation (Rob + Prior) RDS Fusion
Triggered (n = 417) Non-Triggered (n = 317)
Accuracy (%) 71.5 71.7 81.4 86.4
Macro F1 0.480 0.655 0.449 0.464
Ironic F1 0.829 0.801 0.000 0.000
Non-Ironic F1 0.131 0.508 0.897 0.927
Table 2: Full subpopulation analysis based on hashtag detector activation. CoT precision recovery elevates Non-Ironic F1 from 0.131 to 0.508 on triggered tweets. Few-shot irony detection is impossible on non-triggered tweets (Ironic F1 = 0.000 for both), confirming that the symbolic prior is strictly load-bearing for recall.
Model
Paradigm
Acc.
MF1
Ir.F1
CCR-Net [18] BERTweet [4] RDS Fusion
Supervised Supervised Hybrid
70.9 78.2 78.1
0.708 — 0.777
0.685 0.746 0.747
Table 3: Supervised comparison on TweetEval Irony. BERTweet and CCR-Net evaluated on full 784-tweet split; RDS on held-out N = 734.
Model stce [19] RDS Fusion Ablation (Rob + Prior) RoBERTa-only Baseline SarcasmDet [20] UTNLP [21] FII UAIC [22]
Eval.
Ir.F1
Supervised SOTA Hybrid Hybrid Supervised† Supervised Supervised Supervised
0.6052 0.4821 0.4793 0.4786 0.4300 0.3800 0.3700
Table 4: SemEval-2022 Task 6 leaderboard (selected subset) vs. RoBERTa, the Ablation Model & RDS Fusion. † RoBERTa is trained on the train set of TweetEval [3] training set.
applied to an objectively unpleasant event (a 3AM fire alarm). Both signals are strongly aligned. CoT pipeline: The LLM reasoning trajectory outputs PCoT = 0.85, independently parsing the pragmatic incongruence between the enthusiastic exclamation and the disruptive context. With entropy H(c) = 0.342 mapping to γdyn = 0.469, the Local Guardian whitelists critical tokens including indicates, disruptive, excitement, and Tone. The strong signal override recomputes Pbase = 0.30 × 0.980 + 0.70 × 0.85 = 0.889, and the final blend yields Pfused = 0.60 × 0.889 + 0.40 × 0.88 = 0.885. Correct Ironic prediction. This instance demonstrates the framework’s most coherent operating mode: the symbolic prior, the RoBERTa encoder, and the CoT pipeline are unanimously aligned, producing a high-confidence, correctly calibrated ironic prediction. 5.7.3
Case Study 3: Correct Rejection of Implicit Sarcasm (Tweet 143, TweetEval Dataset)
Tweet: how is it possible that somebody so interesting is also so, so, so boring? #irony #what This is a ground-truth ironic tweet containing a definidemonstrates exactly how the CoT engine functions as a tive author-labeled marker (#irony). The symbolic semantic safeguard against symbolic over-firing. prior correctly activates on this strong signal, assigning 5.7.2 Case Study 2: Correct Irony Classification S prior (x) = 0.88 and engaging the strong-signal override with Composite Signals (Tweet 172, Tweet- to cap the baseline trust weight at α = 0.300. Eval Dataset) RoBERTa baseline: The fine-tuned encoder assigns a Tweet: Yay for Fire Alarms at 3AM #not severely low irony probability (PRoBERTa = 0.087), enThis tweet contains a strong explicit irony marker tirely failing to resolve the self-contradictory predication (#not), which the symbolic prior immediately identifies, as- (“interesting” and “boring” applied to the same subject). signing Sprior (x) = ρstrong = 0.88. This caps the baseline CoT pipeline: The LLM outputs PCoT = 0.15 with a trust weight at α ← min(α, 0.30). predictive entropy of H(c) = 0.249. The reasoning trajecRoBERTa baseline: The fine-tuned encoder indepen- tory misinterprets the oxymoronic construction as a literal dently assigns a very high irony probability (PRoBERTa = complaint, lacking the contextual agility to recognize the 0.980), consistent with the overtly positive framing (“Yay”)
8
Bhattacharjee & Bhaumik (2026)
pragmatic inversion. Adaptive Fusion: Because both neural components fail with extreme confidence, their combined base probability collapses to Pbase = 0.30×0.087+0.70×0.15 = 0.131. Even after applying the heavy strong-signal blend (ωstrong = 0.60), the prior cannot rescue the prediction: Pfused = 0.60 × 0.131 + 0.40 × 0.88 = 0.431. The result falls below the 0.5 decision threshold, yielding a false negative. This case exposes a critical architectural boundary: the limits of the symbolic prior’s rescue capacity. When both the fine-tuned encoder and the zero-shot LLM confidently fail in unison, their combined neural literal bias mathematically overpowers even a perfectly fired, high-confidence explicit marker.
6
Limitations
Inference Latency. The framework incurs a multi-orderof-magnitude latency penalty: the RoBERTa baseline processes 1400 samples in <1 minute; full RDS requires ≈5 hours (≈14.1 sec/sample) due to sequential autoregressive generation and per-sample backward passes. RDS is appropriate for offline precision auditing, not real-time deployment. The Implicit Irony Bottleneck. On datasets where irony is strictly psychological or semantic rather than syntactic (e.g., iSarcasm), the framework suffers a dual systemic collapse. First, the symbolic prior remains largely dormant, activating on only 1.5% of samples and dropping RDS to a purely neural configuration. Second, within this neural fallback, the few-shot CoT pipeline exhibits severe literal inertia—yielding an isolated Ironic F1 score of 0.000 on ungrounded text. The signal from the CoT is thus incapable of recovering implicit irony when explicit linguistic anchors are absent. Static Fusion Weight Fragility. Calibration thresholds optimized on a 50-tweet TweetEval split cannot universally transfer across shifting linguistic distributions. On iSarcasm’s triggered subset, frozen blending weights allowed the LLM’s literal bias to overpower the valid symbolic prior, degrading Ironic F1 from 0.889 to 0.571. Narrow Dynamic Compression Range. The computed γdyn rarely utilizes theoretical extremes (γmin = 0.1, γmax = 0.9), hovering near median due to the uniformly moderate entropy of social media CoT. Compression efficiency gains are therefore partially muted.
7
Conclusion
We presented the Robust Dual-Signal Fusion (RDS) framework, a parameter-free hybrid neuro-symbolic architecture achieving few-shot irony detection at supervised parity. Our central empirical finding reframes CoT’s functional
role: within a dual-signal architecture, few-shot reasoning does not detect irony—it recovers precision by anchoring literal semantics to suppress heuristic over-firing. Crucially, isolated few-shot CoT evaluation yields sub-random performance (Macro F1 ¡ 0.500), definitively proving that the LLM cannot function as a primary irony detector in this domain; its utility is strictly unlocked when dynamically constrained as a precision-recovery gate. Validated on the held-out TweetEval test (N = 734), RDS matches BERTweet’s fine-tuned performance ceiling (78.1%, Macro F1 = 0.777) with statistical significance (p = 0.005). Deployed in a few-shot setting on the cross-domain iSarcasm benchmark without modification, it surpasses multiple supervised SemEval systems. These results demonstrate that calibrated neuro-symbolic fusion at inference time can bridge a substantial portion of the gap to domain-specific fine-tuning—while cleanly delineating the latency, calibration, and domain-dependency boundaries that define the frontier of few-shot hybrid systems. Data and Code Availability. The full RDS implementation, dataset sources, and baseline model weights are detailed in Appendix B.7. Acknowledgements. We thank Professor Sourangshu Bhattacharya (Department of Computer Science & Engineering, IIT Kharagpur) and mentor Vaishnovi Arun for guidance throughout this work. We also sincerely thank Yuvraj Veer and Maitreyee Chakraborty for their assistance with the preliminary presentation of this work.
References [1] H. Xia, C. T. Leong, W. Wang, Y. Li, and W. Li. TokenSkip: Controllable Chain-of-Thought Compression for Efficient LLM Inference. arXiv:2502.12067, 2024. [2] J. Wei, X. Wang, D. Schuurmans, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824–24837, 2022. [3] F. Barbieri, J. Camacho-Collados, et al. TweetEval: Unified Benchmark and Comparative Evaluation for Tweet Classification. Findings of EMNLP, pages 1644–1650, 2020. [4] D. Q. Nguyen, T. Vu, and A. T. Nguyen. BERTweet: A Pre-trained Language Model for English Tweets. EMNLP Demos, pages 9–14, 2020. [5] A. Vaswani, N. Shazeer, N. Parmar, et al. Attention is All You Need. NeurIPS, 30, 2017. [6] OpenAI. Learning to reason with LLMs. OpenAI Blog, 2024. [7] Y. Li, B. Dong, F. Guerin, and C. Lin. Compressing context to enhance inference efficiency of large language models. EMNLP, pages 6342–6353, 2023.
Robust Dual-Signal Fusion
[8] Z. Pan, Q. Wu, H. Jiang, et al. LLMLingua-2: Data distillation for efficient and faithful task-agnostic prompt compression. Findings of ACL, pages 963–981, 2024.
9
Appendices A
Extended Neuro-Symbolic Heuristics & Lexicons
A.1
Contrastive Connector Vocabulary
[9] A. Yang, B. Yang, B. Hui, et al. Qwen2 Technical Report. arXiv:2407.10671, 2024.
[10] Y. Liu, M. Ott, N. Goyal, et al. RoBERTa: (C) A Robustly Optimized BERT Pretraining Approach. The contrastive connector set used in the GradientarXiv:1907.11692, 2019.
Sensitized Local Guardian is:
[11] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image C = {but, yet, though, although, however, despite, while, classification models and saliency maps. arXiv preprint whereas, still, even, never, not, no, without, barely}. arXiv:1312.6034, 2013.
These tokens are disproportionately responsible for carry[12] Jiwei Li, Xinlei Chen, Eduard Hovy, and Dan Jurafsky. ing ironic inversion in natural language [14, 15] and are Visualizing and understanding neural models in NLP. guaranteed to survive any compression budget by virtue Proceedings of the 2016 Conference of the North American of τconn = 500 ≫ τcont = 10. Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pages 681– A.2 Symbolic Prior Lexicons 691, 2016.
Strong signal set Rstrong : definitive author-labeled irony
[13] E. Riloff, A. Qadir, P. Surve, et al. Sarcasm as Contrast hashtags and their close orthographic variants: between a Positive Sentiment and Negative Situation. #sarcasm, #sarcastic, #irony, #ironic, EMNLP, pages 704–714, 2013.
#not, #justkidding, #jk [14] A. Joshi, V. Sharma, and P. Bhattacharyya. Harnessing Context Incongruity for Sarcasm Detection. ACL- Presence of any element assigns Sprior (x) = ρstrong = 0.88. IJCNLP, pages 757–762, 2015. [15] A. Reyes, P. Rosso, and T. Veale. A Multidimensional Approach for Detecting Irony in Twitter. Language Resources and Evaluation, 47(1):239–268, 2013. [16] A. d’Avila Garcez and L. C. Lamb. Neuro-Symbolic AI: The 3rd Wave. Artificial Intelligence Review, 56:12387– 12406, 2023.
Weak signal set Rweak : soft hashtag signals often cooccurring with ironic intent but insufficient alone: #lol, #sure, #totally, #obviously, #right, #yeah, #great, #awesome, #fantastic, #wonderful, #perfect, #love, #amazing Each contributes ρweak = 0.15 to the composite score.
[17] H. Kautz. The Third Wave of Artificial Intelligence. AI Magazine, 43(2):93–98, 2022.
Emoji contrast rules Remoji : structural patterns where a positive-valence emoji is immediately followed by a [18] Z. Rahman, F. Din, and S. Khalid. CCR-Net: Surpassing negative-valence emoji or vice versa within the same token Baselines in Irony Detection via Supportive-Conflictive window. Positive cluster includes smiling, laughing, heart, Attention Fusion. Expert Systems with Applications, thumbs-up, and celebration symbols. Negative cluster 304:130814, 2026. includes crying, angry, broken-heart, and eye-roll symbols. A match contributes ρemoji = 0.25. [19] M. Yuan, M. Zhou, L. Jiang, Y. Mo, and X. Shi. stce at SemEval-2022 Task 6. SemEval-2022, 2022.
Elongation tokens Relong : tokens where any single character is consecutively repeated ≥ η = 3 times (e.g., sooooo, [20] M. Abdullah, D. Alnore, S. Swedat, et al. SarcasmDet at yesssss, reallyyyy). Detected via the regex (.)\1{2,}. SemEval-2022 Task 6. SemEval-2022, 2022. Contributes ρelong = 0.12. [21] A. Abaskohi, A. Rasouli, T. Zeraati, and B. Bahrak. UTNLP at SemEval-2022 Task 6. SemEval-2022, 2022. [22] T. Manoleasa and I. Sandu. FII UAIC at SemEval-2022 Task 6. SemEval-2022, 2022.
A.3
Negation-Aware Semantic Scan Vocabulary
Ironic indicator tokens : sarcasm, sarcastic, ironic, irony, clearly, [23] C. Van Hee, E. Lefever, and V. Hoste. SemEval-2018 obviously, totally, definitely, absolutely, Task 3: Irony Detection in English Tweets. SemEval- sure, right, great, wonderful, perfect, love, 2018, pages 39–50, 2018. exactly, certainly, naturally.
10
Bhattacharjee & Bhaumik (2026)
Figure 5: Distribution of uncompressed CoT token lengths. Tight variance (mean=83.7, σ = 5.39, range 60–104) demonstrates prompt-enforced deterministic structure; generation concludes naturally below the 120-token cap.
Figure 7: Alpha (α) distribution. Heavy clustering at [0.30, 0.40] floor (55.4% of inferences) confirms conservative failsafe design; α > 0.5 on 41.3% of inferences shows active contribution.
Figure 6: Distribution of γdyn . Near-zero variance (σ = 0.013) confirms that precision recovery is driven by the Local Guardian rather than the macro budgeter.
Figure 8: Hashtag activation analysis. Strong-signal firing rate on TweetEval vs. near-total inertia on iSarcasm (1.5%) confirms the prior’s dependence on explicit syntactic conventions.
(2×16 GB VRAM, CUDA 12.2). The full iSarcasm evaluation (N = 1400) required approximately 5 hours of wall-clock time on this configuration, averaging ≈12.86 seconds per sample. The TweetEval evaluation (N = 734) Local negation tokens (invert the score of the subse- required approximately 2.8 hours. quent indicator): B.2 Quantization Details not, no, never, without, hardly, barely, Qwen2.5-3B-Instruct was loaded using 4-bit NF4 (Norcannot, can’t, won’t, don’t. mal Float 4) quantization via bitsandbytes to optimize Score mapping: ironic token count ri and literal token memory footprint: ri count rl map to PCoT = 0.15 + 0.70 · ri +r where ϵ = l +ϵ • Quantization type: nf4 (NormalFloat 4-bit, opti−6 1 × 10 , clipped to [0.15, 0.85]. mized for normally distributed weights [9]) Literal indicator tokens : literal, sincere, genuine, actually, truly, honestly, factual, real, straightforward, serious, earnest, authentic.
B
• Double quantization: enabled (quantizes the quanti-
Hardware, Environment, and zation constants to further reduce memory) • Compute dtype: bfloat16 Reproducibility Details
• Full model memory footprint (postquantization): ≈ 2.1 GB VRAM greedy decoding (temperature=0, All experiments were executed on Kaggle and Google Co- • Generation: do sample=False), max new tokens=120 lab environments with NVIDIA dual T4 GPU acceleration
B.1
Hardware Configuration
Robust Dual-Signal Fusion
11
Figure 9: Cross-domain false positive reduction on iSarcasm. RDS suppresses 71 baseline hallucinations (315 → 244). Figure 10: Precision-recall trade-off on iSarcasm ironic class. Precision recovery drives system-level Macro F1 gains.
• Two-pass fallback: same model, max new tokens=10
B.3
Library Versions Library transformers torch bitsandbytes llmlingua numpy scikit-learn statsmodels
B.4
Version 4.40.0 2.2.0+cu121 0.43.1 0.2.2 1.26.4 1.4.2 0.14.1
Parameter Hmin , Hmax γmin , γmax τcrit , τconn τcont , τlow ρstrong , ρcap ρweak , ρemoji , ρelong αmin , αmax Hsk , λsk ∆conf , αconf ωstrong , ωweak
Value 0.2, 0.5 0.1, 0.9 1000, 500 10, 1 0.88, 0.70 0.15, 0.25, 0.12 0.50, 0.88 0.25, 0.5 0.85, 0.85 0.60, 0.75
Table 6: Consolidated pipeline hyperparameters.
Table 5: Library versions for reproducibility.
B.6
Generation Parsing Guardrails
To prevent data leakage, the first 50 tweets of the TweetEval test set were designated as a validation split before any parameter search. All calibration of Hmin , Hmax , αmin , αmax , ρstrong , and blending weights was strictly confined to these 50 samples (achieving an over-fitted validation accuracy of 88.0% on this calibration split). The remaining 734 tweets were treated as a completely unseen held-out test set. The iSarcasm evaluation used frozen TweetEval calibrations with zero additional tuning, guaranteeing a strictly few-shot cross-domain evaluation.
The following secondary guardrails were applied to CoT outputs before confidence extraction: 1. Encoding normalization: Outputs are decoded with skip special tokens=True and unicode-normalized to NFC. 2. Score boundary clamping: Any extracted score outside [0.0, 1.0] is hard-clamped to the nearest boundary. 3. Degenerate output detection: If the output length < 5 tokens (model failure), the extraction defaults to PCoT = 0.5 and triggers the two-pass fallback. 4. Repetition detection: Outputs exhibiting tokenlevel repetition loops (any token repeated > 10 consecutive times) are flagged and routed to the two-pass fallback without confidence extraction. 5. Entropy computation guard: Tokens with generation probability < 10−8 are floored before log computation to prevent numerical underflow in H(c).
B.5
Module Compression Budget Compression Bounds Retention Overrides Retention Overrides Prior Weights Weak Prior Weights Fusion Bounds Skepticism Gate Override Margins Blend Weights
Consolidated Calibration Thresholds
Table 6 consolidates all static thresholds and fusion parameters derived from the 50-tweet validation split.
B.7
Validation Protocol Leakage Guarantee
and
Anti-
Data & Code Availability
The complete RDS implementation and generation logs are available on GitHub. The raw evaluation datasets were sourced directly from the official TweetEval and iSarcasm repositories. The fine-tuned discriminative baseline (twitter-roberta-base-irony) is accessible via Hugging Face.